Compare commits
3 Commits
| Author | SHA1 | Date | |
|---|---|---|---|
 Praxis CI
|
5290d4d05d | ||
|
Praxis CI
|
ba928cf3b4 | ||
|
Praxis CI
|
f2a12f9fed |
+166
-1
@@ -746,4 +746,169 @@ Top risks for PLAN: R-AUTH-01 (Secure cookie + no-TLS → config-driven flag, gr
|
||||
|
||||
**Pip (pyproject.toml):** `asyncpg>=0.29` (Postgres driver), `argon2-cffi>=23.1` (password hashing), `slowapi>=0.1` (rate limiting). `pynacl`, `canonicaljson`, `base58` already present (v0.3).
|
||||
|
||||
**Npm (client/package.json):** `react-router-dom@^7` (React routing for /operator/*). No chart library — inline SVG sparklines (zero deps).
|
||||
**Npm (client/package.json):** `react-router-dom@^7` (React routing for /operator/*). No chart library — inline SVG sparklines (zero deps).
|
||||
|
||||
---
|
||||
|
||||
## v0.5 Live Assist Mode (On-the-Job Voice Companion)
|
||||
|
||||
> **Status:** Research-refined (v0.5 RESEARCH stage). Informed by `.ciagent/RESEARCH-v0.5-live-assist.md`.
|
||||
> **Decisions:** D-058 (wake-word invocation, REFINED by D-064), D-059 (context-binding), D-060 (3-layer guardrail, REFINED by D-068), D-061 (latency budget, AT RISK — see R-ASSIST-02), D-062 (shift-bounded sessions), D-063 (assist ≠ mastery), D-064 (Porcupine built-in WW + Vosk fallback), D-065 (Piper TTS for assist), D-066 (≤150-token assist prompt), D-067 (warm WebRTC per shift), D-068 (regex output filter + retry + canned fallback), D-069 (8h auto-end shift), D-070 (consent disclosure).
|
||||
> **Open flags for orchestrator:** (1) Picovoice MAU pricing has no recurring free tier — R-ASSIST-01; (2) C-8 <600ms latency at risk for assist (~655-770ms estimated) — R-ASSIST-02; (3) v0.5 may require a client upgrade from React-Web to React-Native for background wake-word — RESEARCH §7 Q1; (4) Canada consent law for ambient recording — R-ASSIST-08.
|
||||
|
||||
### v0.5 Component Map (additions to v0.4)
|
||||
|
||||
```
|
||||
Pipecat server (Python)
|
||||
├─ ... (v0.2 voice loop + v0.3 mastery/VC/IRT + v0.4 operator/auth/cohort unchanged) ...
|
||||
├─ Assist pipeline NEW (server/assist/) (v0.5 — D-061, D-065, D-066, D-067)
|
||||
│ ├─ build_assist_pipeline() (reuses _build_transport/stt/llm/tts; swaps context)
|
||||
│ ├─ AssistContextBinder (loads path week + scenario tag + learner theta from SQLite →
|
||||
│ │ ≤150-token context string — D-059, D-066)
|
||||
│ ├─ In-loop guardrail processor NEW (post-LLM frame processor, pre-TTS — D-060, D-068)
|
||||
│ │ └─ LiveAssistGuardrail.check(text) → GuardrailVerdict
|
||||
│ └─ Warm WebRTC connection manager NEW (shift-bounded, heartbeat every 30s — D-067)
|
||||
├─ LiveAssistGuardrail NEW (server/guardrails/live_assist.py) (v0.5 — D-060, D-068, REQ-ASSIST-03)
|
||||
│ ├─ Layer 1: coaching-mode system prompt (ask guiding questions, never give the answer,
|
||||
│ │ never speak on behalf of the learner, never claim false authority)
|
||||
│ ├─ Layer 2: regex output filter
|
||||
│ │ ├─ DIRECT_SCRIPT_RE ("you should say X" / "tell the customer Y" / "the answer is Z")
|
||||
│ │ ├─ IMPERATIVE_RE ("escalate to" / "offer a refund of" / "apologize by")
|
||||
│ │ ├─ FALSE_AUTHORITY_RE ("I am your manager" / "on behalf of the company")
|
||||
│ │ ├─ IMPERSONATION_RE (carry-forward from CustomerServiceGuardrail)
|
||||
│ │ ├─ COACHING_QUESTION_RE (ALLOW — "what do you think" / "how could you")
|
||||
│ │ └─ on block: one retry ("Rephrase as a coaching question") → canned fallback
|
||||
│ └─ Layer 3: audit log
|
||||
│ ├─ turns table gains guardrail_verdict JSON column (additive SQLite migration)
|
||||
│ └─ guardrail_block_count surfaces to cohort aggregation (operator safety signal)
|
||||
├─ Assist session API NEW (server/assist/routes.py) (v0.5)
|
||||
│ ├─ POST /api/assist/shift/start (declare context: path week + scenario tag → warm WebRTC)
|
||||
│ ├─ POST /api/assist/shift/end (close warm WebRTC, fire aggregation hook, auto-end after 8h — D-069)
|
||||
│ └─ (assist turns flow over the warm WebRTC connection, not separate HTTP endpoints)
|
||||
└─ Cohort aggregation extension (server/cohort/aggregator.py) (v0.5 — D-062, no schema change)
|
||||
├─ session_outcome gains session_type: 'practice' | 'assist'
|
||||
├─ _aggregate_assist() branch: assist_shifts_count, assist_turns_count,
|
||||
│ assist_avg_turns_per_shift, assist_active_learners_count, assist_guardrail_block_rate
|
||||
└─ k-anonymity ≥ 10 suppression identical to practice (D-034 carry-forward)
|
||||
|
||||
Client (Android — likely React Native upgrade, RESEARCH §7 Q1)
|
||||
├─ ... (v0.1 React web practice UI at / unchanged) ...
|
||||
├─ Praxis Assist foreground service NEW (v0.5 — D-058, D-064, D-067, D-070)
|
||||
│ ├─ Porcupine wake-word listener (built-in wake word for v0.5 pilot; custom post-pilot — D-064)
|
||||
│ ├─ Foreground service type: microphone (Android 14+ requirement)
|
||||
│ ├─ Persistent notification: "Praxis Assist is listening" (consent disclosure — D-070)
|
||||
│ ├─ Warm WebRTC connection to praxis server (opened at shift start, keepalive every 30s)
|
||||
│ └─ Tap-to-talk fallback (battery-saving mode / wake-word failure / noisy environment)
|
||||
└─ Assist control surface (minimal React: Start/End Shift toggle + context declaration)
|
||||
└─ ~100-150 LOC — below frontend-engineer reactivation threshold (PERSONAS §7.2)
|
||||
```
|
||||
|
||||
### Assist Voice Loop (distinct from the practice scenario loop)
|
||||
|
||||
```
|
||||
Shift start (learner: "Hey Praxis, starting my shift" or tap "Start Shift")
|
||||
├─ Foreground service starts (Porcupine on, warm WebRTC opens)
|
||||
├─ Learner declares context (path week + scenario tag) → AssistContextBinder
|
||||
│ └─ server reads progress.current_week from SQLite (D-007) + theta from learner_ability
|
||||
├─ Assist session row created (SQLite sessions, session_type='assist', started_at=now())
|
||||
|
||||
Assist turn (learner: "Hey Praxis" + situation/question)
|
||||
├─ Porcupine detects wake word (~200-500ms detection latency)
|
||||
├─ Foreground service routes audio to warm WebRTC → praxis server
|
||||
├─ Pipeline (reuses v0.1 services, assist-mode prompt):
|
||||
│ transport.input → stt (Deepgram) → AssistContextBinder (inject context) →
|
||||
│ llm (gemma4:cloud, ≤150-token assist prompt — D-066) →
|
||||
│ LiveAssistGuardrail (regex output filter — D-068) →
|
||||
│ tts (Piper ~80ms — D-065) → transport.output
|
||||
├─ Coaching plays in-ear. Turn logged (turns table + guardrail_verdict).
|
||||
└─ WebRTC stays warm for the next turn.
|
||||
|
||||
Shift end (learner: "Hey Praxis, ending shift" or tap "End Shift" or 8h auto-end — D-069)
|
||||
├─ Foreground service stops (Porcupine off, mic released, notification dismissed)
|
||||
├─ Warm WebRTC closed
|
||||
├─ Assist session row updated (ended_at, outcome, turn_count, guardrail_block_count)
|
||||
└─ on-session-end hook fires → cohort aggregation (session_type='assist') → Postgres
|
||||
(NOT the mastery flow — schedule_mastery=False per D-063)
|
||||
```
|
||||
|
||||
### Context-Binding (D-059, D-066)
|
||||
|
||||
The assist system prompt is ≤150 input tokens (D-066) to keep LLM prefill latency under 50ms:
|
||||
|
||||
```
|
||||
[Layer 1 coaching instruction — ~80 tokens, fixed]
|
||||
You are a live coaching AI in the learner's ear during a real customer interaction.
|
||||
Coach, do not do the learner's job. Ask guiding questions; never give the answer.
|
||||
Never speak on behalf of the learner. Never claim authority you don't have.
|
||||
Keep responses to 1-3 sentences for voice.
|
||||
|
||||
[Context-binding — ~50 tokens, per shift]
|
||||
Week {current_week}: {week_focus}. Scenario: {scenario_tag}.
|
||||
Learner theta: {theta:.1f}. Coaching focus: {top_rubric_criterion}.
|
||||
|
||||
[Voice-conciseness — ~20 tokens, fixed]
|
||||
Be brief. The customer is waiting.
|
||||
```
|
||||
|
||||
### Guardrail Extension (D-060, D-068, REQ-ASSIST-03)
|
||||
|
||||
The `Guardrail` interface (server/services/base.py) is extended with `LiveAssistGuardrail` (server/guardrails/live_assist.py). The 3 layers:
|
||||
|
||||
| Layer | Mechanism | On-voice-path? | Latency |
|
||||
|-------|-----------|-----------------|---------|
|
||||
| 1. Prompt rules | Coaching-mode system prompt (ask, don't tell) | Yes (system prompt) | 0ms (prefill only) |
|
||||
| 2. Output filter | Regex: DIRECT_SCRIPT_RE + IMPERATIVE_RE + FALSE_AUTHORITY_RE + IMPERSONATION_RE; COACHING_QUESTION_RE (allow) | Yes (post-LLM, pre-TTS) | <5ms (regex) |
|
||||
| 3. Audit log | turns table guardrail_verdict JSON + cohort aggregation guardrail_block_rate | No (async, off-voice-path) | 0ms on path |
|
||||
|
||||
Output filter logic: on direct-answer/false-authority/impersonation hit → block + log + one retry ("Rephrase as a coaching question"). If retry also blocks → canned fallback: "Think about what the customer needs right now. What's your next step?"
|
||||
|
||||
### Latency Budget for Assist Turns (D-061, R-ASSIST-02 — AT RISK)
|
||||
|
||||
| Segment | Budget | Note |
|
||||
|---------|--------|------|
|
||||
| Client capture + WebRTC uplink | ~50ms | warm connection (D-067) |
|
||||
| ASR (Deepgram Nova-3) | ~250ms | R1: measure |
|
||||
| LLM first token (gemma4:cloud, ≤150-token prompt — D-066) | ~225ms | +25ms prefill over v0.1 lean prompt |
|
||||
| TTS first audio (**Piper** — D-065) | ~80ms | R4 mitigation as assist default |
|
||||
| WebRTC downlink + playback | ~50ms | |
|
||||
| **Total (Piper + lean prompt, target)** | **~655ms** | ⚠️ ~55ms over C-8's <600ms |
|
||||
|
||||
**Wake-word → first-audio (distinct budget):** ~850-1150ms (warm WebRTC) — from Porcupine detection (~200-500ms) + the in-conversation turn budget above. This is the expected "time from saying 'Hey Praxis' to hearing coaching." Acceptable for live assist (not the in-conversation <600ms target).
|
||||
|
||||
**Mitigations to reach <600ms:** (a) measure R1/R3 — if Deepgram is ~200ms or Ollama Cloud is ~150ms, the total drops under 600ms; (b) accept ~650ms for the pilot, target <600ms in v0.6 with optimization. **Flag: C-8 is the binding constraint; the orchestrator may relax it for assist mode or push hardening to v0.6.**
|
||||
|
||||
### Aggregation Integration (D-062, no schema change)
|
||||
|
||||
The `cohort_aggregates` table (generic on `metric TEXT`) gains assist metrics as new metric strings — no DDL. The `session_outcome` dict gains `session_type: 'practice' | 'assist'`. The aggregator branches:
|
||||
|
||||
```python
|
||||
# server/cohort/aggregator.py extension (shape only)
|
||||
async def aggregate_session(pg_store, session_outcome):
|
||||
if session_outcome.get("session_type") == "assist":
|
||||
await _aggregate_assist(pg_store, session_outcome) # assist metrics
|
||||
else:
|
||||
await _aggregate_practice(pg_store, session_outcome) # existing v0.4 logic
|
||||
```
|
||||
|
||||
**Assist metrics:** `assist_shifts_count`, `assist_turns_count`, `assist_avg_turns_per_shift`, `assist_active_learners_count`, `assist_guardrail_block_rate`. All k-anonymized (≥10 distinct learners, else suppressed — D-034 carry-forward).
|
||||
|
||||
**Dashboard views (D-053 extension):** Practice volume → adds assist volume; Mastery progression → unchanged (assist ≠ mastery, D-063); Failure patterns → adds `assist_guardrail_block_rate` as a safety signal.
|
||||
|
||||
### v0.5 Risks (from RESEARCH-v0.5-live-assist.md)
|
||||
|
||||
Top risks for PLAN: R-ASSIST-01 (Picovoice MAU pricing — no recurring free tier, engage sales or use built-in wake word), R-ASSIST-02 (C-8 <600ms at risk for assist, ~655-770ms estimated), R-ASSIST-03 (wake-word→first-audio ~850-1150ms warm), R-ASSIST-07 (output filter false negatives — defense-in-depth + audit), R-ASSIST-08 (privacy/consent for ambient recording — legal review). Full table (14 risks) in RESEARCH-v0.5-live-assist.md.
|
||||
|
||||
### v0.5 New Dependencies
|
||||
|
||||
**Pip (server-side):** none new. The v0.1 voice pipeline (Pipecat + Deepgram + Cartesia + Piper + Ollama) is reused unchanged. The guardrail is pure-Python regex (no new dep). The aggregation extension uses existing asyncpg.
|
||||
|
||||
**Gradle (client-side, Android):** `ai.picovoice:porcupine-android` (wake-word detection — D-058, D-064). **Note:** the v0.1 client is React + WebRTC (D-015), which can't run a background foreground service on Android. v0.5 likely requires a **React Native upgrade** or a **separate native Android assist app** — see RESEARCH §7 Q1 (flag for orchestrator).
|
||||
|
||||
### v0.5 Open Architecture Questions (for PLAN stage)
|
||||
|
||||
- R-ASSIST-02: C-8 <600ms — relax for assist or push hardening to v0.6?
|
||||
- Client architecture: React Native upgrade, separate native app, or defer wake-word to v0.6 (tap-to-talk only for v0.5)?
|
||||
- Picovoice sales engagement timing (before PLAN or after v0.5 ships with tap-to-talk)?
|
||||
- Output filter regex corpus: how to build the tuning corpus before v0.5 ships?
|
||||
- Guardrail verdict storage: JSON column on `turns` or separate `guardrail_verdicts` table?
|
||||
- Canada consent law review for ambient recording (R-ASSIST-08).
|
||||
+12
-14
@@ -1,19 +1,17 @@
|
||||
{
|
||||
"phase": 3,
|
||||
"stage": "in_progress",
|
||||
"milestone": "v0.4",
|
||||
"phase_role": "final_review",
|
||||
"phase": 0,
|
||||
"stage": "grill",
|
||||
"milestone": "v0.5",
|
||||
"phase_role": "pre_execution",
|
||||
"attempts": 0,
|
||||
"updated_at": "2026-08-04T12:00:00Z",
|
||||
"updated_at": "2026-08-04T12:35:00Z",
|
||||
"milestone_complete": false,
|
||||
"milestone_merged_to_main": false,
|
||||
"tag": "v0.1.8",
|
||||
"release_url": "https://git.cloudinit.dev/coreci/praxis/releases/tag/v0.1.8",
|
||||
"release_status": "created",
|
||||
"next_milestone": null,
|
||||
"requirements": {
|
||||
"covered": ["REQ-MT-01", "REQ-AUTH-01", "REQ-NFR-AUTH-01", "REQ-NFR-MT-01", "REQ-MT-02", "REQ-DASH-01", "REQ-NFR-DASH-01", "REQ-NFR-DASH-02"],
|
||||
"active": [],
|
||||
"deferred": []
|
||||
}
|
||||
"next_milestone": "v0.5",
|
||||
"active_requirements": ["REQ-ASSIST-01", "REQ-ASSIST-02", "REQ-ASSIST-03", "REQ-NFR-ASSIST-01", "REQ-NFR-ASSIST-02", "REQ-NFR-ASSIST-03", "REQ-NFR-ASSIST-04", "REQ-IDEATE-01", "REQ-IDEATE-02", "REQ-IDEATE-03", "REQ-IDEATE-04", "REQ-IDEATE-05", "REQ-IDEATE-06", "REQ-IDEATE-07", "REQ-IDEATE-08", "REQ-IDEATE-09"],
|
||||
"v0.6_backlog": ["REQ-IDEATE-10", "REQ-IDEATE-11", "REQ-IDEATE-12", "REQ-IDEATE-13"],
|
||||
"tag_base": "v0.1.x",
|
||||
"next_tag": "v0.1.10",
|
||||
"ideate": true,
|
||||
"ideate_result": {"total": 13, "accepted_v0.5": 9, "accepted_v0.6": 4, "skipped": 0}
|
||||
}
|
||||
@@ -0,0 +1,628 @@
|
||||
# CIAgent Grill Report — v0.5 Live Assist (On-the-Job Voice Companion)
|
||||
|
||||
## Run: 2026-08-04 (mode: mechanical, focus: all axes + 6 v0.5-specific probes)
|
||||
|
||||
> **Reviewer:** adversarial technology executive (red-team)
|
||||
> **Subject:** v0.5 execution plan (Live Assist — On-the-Job Voice Companion) — 2 execution phases, 12 slices, 33 tasks, 16 active REQs (3 ASSIST + 4 NFR + 9 IDEATE)
|
||||
> **Stance:** plan is unfeasible, over-scoped, and too costly until evidence forces otherwise
|
||||
> **Artifacts reviewed:** PROJECT.md (D-058..D-073), REQUIREMENTS.md (16 active REQs + 4 v0.6 backlog), ROADMAP.md, ARCHITECTURE.md (v0.5 Live Assist Mode §), RESEARCH-v0.5-live-assist.md (14 risks R-ASSIST-01..14, 7 domains), PLAN-v0.5-live-assist.md (2 phases, 12 slices, 33 tasks), PERSONAS.md (5 active, 2 deactivated), GRILL-v0.4.md (format reference + G-001..G-041), REVIEW.md (8 v0.4 P1+ carried forward), AUDIT.md (v0.4 HEALTHY), config.json (autonomy=full), server/pipeline.py, server/services/base.py, server/guardrails/customer_service.py, server/session_recorder.py, server/__main__.py
|
||||
> **Binding status:** This grill verdict must be cleared (MUSTs resolved, escalations answered) before EXECUTE is authorized.
|
||||
|
||||
---
|
||||
|
||||
### Verdict: Proceed-with-conditions (confidence: 0.70)
|
||||
|
||||
The v0.5 plan is the project's first **safety-critical** milestone — the AI is in a learner's ear during *real* customer interactions, not role-play. This is a categorical shift from v0.1–v0.4 (practice surface, no real customers, no real consequences). The plan's single most important decision — **D-071 (tap-to-talk only, wake-word deferred to v0.6)** — is the correct call: it strips the client-architecture risk (React-Web can't do foreground services), the battery risk, the Picovoice MAU-pricing risk, and 5 of 14 research risks (R-ASSIST-01/04/05/13/14 all become N/A). What remains is the *core* safety surface: the guardrail (REQ-ASSIST-03), the context-binding (REQ-ASSIST-02), and the shift-bounded session model (REQ-NFR-ASSIST-04). This is the right 80/20.
|
||||
|
||||
However, four material issues must be resolved before EXECUTE: (1) **R-ASSIST-07 (guardrail false-negative)** is the single project-killing risk — a direct answer slips past the regex, the learner parrots it to a real customer, trust erodes. The plan *accepts* this residual risk ("adversarial FN rate is reported but not threshold-gated" — PLAN:419) without a documented acceptance threshold or an escalation. For a safety-critical surface, "we'll measure it and trend it nightly" is necessary but not sufficient — the grill must set the bar. (2) **D-073 (PIPEDA consent-law review)** is deferred to "Phase 1 implementation" — but shipping a recording device into real customer interactions without legal sign-off is a regulatory risk the CI agent cannot resolve under full autonomy. This is an escalation, not a binding decision. (3) The IDEATE stage **expanded v0.5 scope from 7 REQs to 16** (+128%) — the first use of ideation in the project. The 9 added REQs are *defensive* (guardrail tuning, mode-conflict, PII policy, audit-log, reconnect, tech-debt, cost, NFR measurement), not feature creep — but the grill must verify the expansion is risk-reduction, not scope inflation. (4) The **in-loop guardrail processor** (post-LLM, pre-TTS) is a *structural pipeline change*, not the "minimal delta / prompt swap" the research frames it as — the v0.1 pipeline has no in-loop guardrail (the CS guardrail runs on the debrief, not in-loop per RESEARCH §5.2). This is the highest-novelty code in v0.5 and it is on the safety-critical path.
|
||||
|
||||
The plan is **not** over-scoped *after* the D-071 deferral (16 REQs, but 9 are defensive; 33 tasks vs v0.4's 52). It is **not** unfeasible (0 new pip/npm deps, v0.1 pipeline reused). It is **not** a zombie (Live Assist is the explicitly-deferred v0.1 surface, now delivered). The conditions are binding and surgical — but two of them (R-ASSIST-07 threshold, PIPEDA escalation) touch the safety-critical core and cannot be waived.
|
||||
|
||||
---
|
||||
|
||||
### Axis 1 — Business Case
|
||||
|
||||
- **Q1: What problem does Live Assist solve that the practice surface (v0.1-v0.4) doesn't? Is "on-the-job coaching" the top priority, or a feature looking for a user?**
|
||||
- Evidence: PROJECT.md:45-47 — "v0.1–v0.4 built and validated the practice surface… v0.5 adds the companion surface: a hands-free voice assistant a learner invokes *while actually working*"; RESEARCH-v0.5 §4.1 — "No direct competitor does live-in-ear coaching during real customer calls on a $100 phone" (verified: Dialpad/Gong post-hoc, RealWear AR+industrial); ROADMAP.md:9-11 — "the key distinction from the practice surface is real-customer interaction."
|
||||
- Answer: Live Assist solves a problem the practice surface structurally cannot: coaching *during* real work, not *after* a role-play. The practice surface (v0.1-v0.4) teaches via simulated scenarios; Live Assist coaches during live customer interactions. This is the *transfer* moment — where practice meets the job. RESEARCH §4.1 confirms Praxis is novel (no competitor does this on a cheap phone). The priority is correct: v0.1-v0.4 built the practice foundation + operator visibility; v0.5 builds the transfer surface. The alternative (v0.6 low-bandwidth) would expand reach before the on-the-job value is proven.
|
||||
- Confidence: 0.80
|
||||
- Decision: **G-042** — Live Assist is the correct next priority (delivers the transfer surface the practice foundation was built for). Novel per RESEARCH §4.1. (0.80)
|
||||
|
||||
- **Q2: Who is the named executive sponsor for Live Assist specifically? (D-001 says "User-directed" for Canada — is there a sponsor for Live Assist?)**
|
||||
- Evidence: config.json:13 — `"level": "full"`; PROJECT.md:5 — "Autonomy: full"; D-001 (PROJECT.md:171) — "Launch market = Canada… User-directed"; no named human sponsor for Live Assist in any `.ciagent/` file.
|
||||
- Answer: No human sponsor. The CI agent is the executive sponsor under full autonomy — the established model since v0.1 (G-002 in GRILL-v0.4). The "sponsor makes a decision under pressure" test is met by this grill — the R-ASSIST-07 + PIPEDA decisions are the pressure decisions. D-001's "User-directed" applied to the *market* choice (Canada), not to Live Assist's scope.
|
||||
- Confidence: 0.80
|
||||
- Decision: **G-043** — CI is the named sponsor under full autonomy (no change from v0.1-v0.4 governance, G-002 carry-forward). (0.80)
|
||||
|
||||
- **Q3: What happens to the business if v0.5 is cancelled? (Does the v0.1-v0.4 practice surface work without it?)**
|
||||
- Evidence: ROADMAP.md:149-157 — future milestones (v0.6 low-bandwidth, v0.7 multi-language) do not depend on Live Assist; PROJECT.md:64-69 — v0.4 operator tier + v0.3 mastery + v0.1 voice loop carry forward unchanged.
|
||||
- Answer: If v0.5 is cancelled, the practice surface (v0.1-v0.4) continues to function. Live Assist is a *new surface*, not a dependency of the existing product. However, cancelling v0.5 means the *transfer* value (coaching during real work) is never delivered — the practice surface teaches, but the on-the-job bridge is missing. This is not a zombie (cancelling has a cost: the product's value proposition — "turn every smartphone into a master craftsperson that talks to you" — is unfulfilled without the live-coaching surface). But the practice surface is independently valuable.
|
||||
- Confidence: 0.78
|
||||
- Decision: **G-044** — v0.5 is not a zombie (delivers the transfer surface). The practice surface works without it, but the product's core promise (on-the-job coaching) is unfulfilled. Accept the non-zombie status. (0.78)
|
||||
|
||||
- **Q4: Is there an ROI calculation vs a counterfactual (skip to v0.6 low-bandwidth)?**
|
||||
- Evidence: MISSING — no ROI calculation in any `.ciagent/` file. D-012 (PROJECT.md:182) — "v0.1 cost ceiling = no enforced ceiling (pilot)"; REQ-IDEATE-07 (REQUIREMENTS.md:70) — assist cost tracking added by ideation.
|
||||
- Answer: No financial ROI. The counterfactual is "ship v0.5 vs skip to v0.6 (low-bandwidth)." Shipping v0.5 costs ~33 tasks of tokens + 0 new deps + the safety-critical guardrail work. Skipping to v0.6 would leave Live Assist permanently deferred (broken v0.1 out-of-scope promise: "Live Assist mode") and v0.6's low-bandwidth surfaces would build on a practice-only product with no on-the-job transfer. The ROI is *product-completeness* (delivering the v0.1-promised surface) + *safety-surface validation* (the guardrail work is the foundation for all future safety-critical domains per D-019). REQ-IDEATE-07 adds cost tracking — the *measurement* of ROI, not the calculation.
|
||||
- Confidence: 0.68
|
||||
- Decision: **G-045** — no financial ROI; the ROI is product-completeness (v0.1-promised surface) + safety-surface foundation (guardrail work extends D-019 for future domains). REQ-IDEATE-07 measures cost, doesn't justify it. Accept the non-financial ROI under full autonomy. (0.68)
|
||||
|
||||
---
|
||||
|
||||
### Axis 2 — Scope and Requirements
|
||||
|
||||
- **Q1: Is the scope stable? 16 active REQs + 4 v0.6 backlog — is this expanding?**
|
||||
- Evidence: REQUIREMENTS.md:8-81 — 16 active REQs (3 ASSIST + 4 NFR + 9 IDEATE); PROJECT.md:49 — "3 REQs + NFRs TBD after RESEARCH/IDEATE"; PLAN-v0.5:1011 — "16/16 REQ-IDs covered"; git log `b8c7de8` — "ideation results — 9 accepted into v0.5, 4 accepted into v0.6."
|
||||
- Answer: The scope **expanded** from 7 REQs (3 ASSIST + 4 NFR, post-CLARIFY) to 16 REQs (+9 IDEATE) — a +128% increase. This is the project's first use of the IDEATE stage. The 9 added REQs are: REQ-IDEATE-01 (guardrail tuning corpus), -02 (in-loop processor test), -03 (mode-conflict), -04 (measurable NFRs), -05 (PII policy), -06 (v0.4 tech-debt), -07 (cost tracking), -08 (WebRTC reconnect), -09 (incremental audit-log). **All 9 are defensive/risk-reduction, not features.** They address: guardrail false-positive/negative (the safety risk), mutual exclusivity (a correctness gap), PII (a privacy gap), NFR measurability (a verifiability gap), tech-debt (carried from v0.4), cost (C-3), resilience (WebRTC drop), audit completeness (abrupt termination). This is scope *hardening*, not scope *creep* — but it is still expansion, and the grill must verify each addition is risk-reduction, not gold-plating.
|
||||
- Confidence: 0.78
|
||||
- Challenge: The +128% expansion is the largest scope growth in the project's history (v0.4 was a clean handoff: 8 REQs, 0 added). The IDEATE stage is a new vector — without discipline, ideation becomes scope creep with a defensive veneer. The 9 REQs are individually justified, but the *aggregate* added 9 tasks of P1 surface + 4 P2 tasks. The grill accepts the expansion *because* each REQ maps to a named risk (R-ASSIST-06/07/08/09/11 + v0.4 P1+ findings), not because ideation is inherently good.
|
||||
- Decision: **G-046** — scope expanded +128% via IDEATE (7→16 REQs). Accepted because all 9 additions are risk-reduction (guardrail, PII, mode-conflict, resilience, audit, tech-debt, cost, NFR measurability), not feature creep. Each maps to a named risk. Future ideation must maintain this risk-reduction discipline. (0.78)
|
||||
|
||||
- **Q2: Are requirements frozen? (The 4 NFRs were `pending-research` → `research-grounded` — are they stable now?)**
|
||||
- Evidence: REQUIREMENTS.md:22-25 — 4 NFRs marked `research-grounded (R-ASSIST-XX)`; REQUIREMENTS.md:27 — "NFRs refined from `pending-research` to `research-grounded` after the v0.5 RESEARCH stage… Phase-1 measurement may further refine R-ASSIST-02 (latency) and R-ASSIST-14 (battery)."
|
||||
- Answer: The 4 NFRs are *research-grounded*, not *frozen*. REQ-NFR-ASSIST-01 (latency) is explicitly "AT RISK" — estimated ~655ms, target <600ms, pilot tolerance ≤650ms (D-072). REQ-NFR-ASSIST-02 (hands-free) was refined by D-071 (tap-to-talk only, wake-word deferred). REQ-NFR-ASSIST-03 (guardrail) is refined by D-068 (regex + retry + fallback). REQ-NFR-ASSIST-04 (session model) is stable (D-062). The NFRs are *stable enough* for PLAN, but REQ-NFR-ASSIST-01's target is a *pilot tolerance* (≤650ms), not the binding constraint (<600ms) — this is a deferred hardening, not a freeze. REQ-IDEATE-04 adds measurable targets (p95 ≤650ms, FP<5%) — this *is* the freeze for measurement purposes.
|
||||
- Confidence: 0.75
|
||||
- Decision: **G-047** — NFRs are research-grounded, not frozen. REQ-NFR-ASSIST-01 (latency) is at-risk with a pilot tolerance (D-072); REQ-IDEATE-04 provides the measurable freeze (p95 ≤650ms pilot, FP<5%). Accept as pilot-scale with v0.6 hardening for <600ms. (0.75)
|
||||
|
||||
- **Q3: What is explicitly out of scope? (Is the v0.5 out-of-scope list as explicit as v0.4's?)**
|
||||
- Evidence: PROJECT.md:54-62 — explicit out-of-scope list (9 items); REQUIREMENTS.md:83-92 — matching list.
|
||||
- Answer: Explicitly out of scope: full multi-path launch, low-bandwidth surfaces (WhatsApp/USSD/offline), multi-language, persona switching, full operator-suite dashboard, learner auth/multi-learner-per-device, session recording/replay, proactive intervention, multi-modal. The list is as explicit as v0.4's. The key deferral is **wake-word (D-071)** — the original D-058 scope (wake-word + tap-to-talk) is reduced to tap-to-talk only, with wake-word deferred to v0.6. This is the largest scope *reduction* in v0.5 and it is explicit (D-071 binding, PLAN:25).
|
||||
- Confidence: 0.85
|
||||
- Decision: **G-048** — out-of-scope is explicit and comprehensive. D-071 (wake-word deferred) is the key scope reduction, documented as binding. (0.85)
|
||||
|
||||
- **Q4: Hidden requirements? (PIPEDA legal review D-073 — is this a hidden regulatory requirement?)**
|
||||
- Evidence: D-073 (PROJECT.md:243) — "PIPEDA consent-law review = defer to v0.5 Phase 1 implementation"; R-ASSIST-08 (RESEARCH-v0.5 §2.6) — "Privacy/consent failure: the real customer didn't consent to being recorded/analyzed by an AI"; D-070 (PROJECT.md:240) — consent disclosure implemented regardless.
|
||||
- Answer: **Yes — PIPEDA is a hidden regulatory requirement.** The ambient mic captures the real customer (a third party); ASR transcribes their speech; the turns table stores it (REQ-IDEATE-05 acknowledges this as "STRIDE information-disclosure"). Canada's PIPEDA + provincial one-party/two-party consent laws govern recording. D-073 defers the legal review to "Phase 1 implementation" and frames it as "not a Phase 0 blocker." The disclosure (D-070) is the *engineering* mitigation, but it is NOT a *legal* determination — a disclosure does not make recording legal if the law requires two-party consent. The CI agent under full autonomy cannot resolve a legal question. This is an **escalation**, not a binding decision — the grill cannot determine with confidence ≥0.60 whether the disclosure is sufficient or whether legal review must block ship.
|
||||
- Confidence: 0.55
|
||||
- Challenge: PIPEDA is a regulatory requirement that the plan defers. For a safety-critical surface with real customers, deferring legal review is a risk the CI cannot own. This must be escalated.
|
||||
- Decision: **ESCALATION-01** — PIPEDA consent-law review (D-073) is a hidden regulatory requirement that cannot be resolved under full autonomy. The disclosure (D-070) is the engineering mitigation but not a legal determination. **Escalate to human attention:** determine whether Canada PIPEDA + provincial consent law requires explicit legal sign-off before shipping a recording device into real customer interactions. If the disclosure is legally sufficient, proceed; if two-party consent is required, the assist surface may need customer-facing consent (out of scope for v0.5) or geographic restriction. (0.55 — below threshold)
|
||||
|
||||
---
|
||||
|
||||
### Axis 3 — Architecture and Technical Feasibility
|
||||
|
||||
- **Q1: Has the assist pipeline architecture been validated? (D-061 says shares v0.1 pipeline — is build_assist_pipeline() validated or assumed?)**
|
||||
- Evidence: server/pipeline.py:44-185 — `build_pipeline()` with `_build_transport` (line 63), `_build_stt` (line 76), `_build_llm` (line 89), `_build_tts` (line 109), `LatencyObserver` (line 183); RESEARCH-v0.5 §5.2 — "v0.5 adds a `build_assist_pipeline()`… Reuses `_build_transport`, `_build_stt`, `_build_llm`, `_build_tts` unchanged"; PLAN-v0.5 TASK-05-01 — `build_assist_pipeline()` assembles the pipeline.
|
||||
- Answer: The v0.1 service constructors (`_build_transport/stt/llm/tts`) are verified present and reusable (pipeline.py:63-109). `build_assist_pipeline()` is *assumed* to reuse them — this is sound for the service layer. **However**, the in-loop guardrail processor (TASK-05-02 — `LiveAssistGuardrailProcessor` as a post-LLM, pre-TTS `FrameProcessor`) is a *structural pipeline change*, not a prompt swap. The v0.1 pipeline has NO in-loop guardrail processor — the CS guardrail runs on the debrief (post-session), not in-loop (RESEARCH §5.2: "the existing v0.1 pipeline doesn't have a post-LLM guardrail processor inline"). Inserting a frame processor between `llm` and `tts` is novel for this codebase. The research frames this as "~1 new Pipecat frame processor" (§5.2) — but Pipecat frame-processor semantics (when does `LLMFullResponseEndFrame` fire? can you inject a retry mid-stream?) are unvalidated. PLAN Open Question #4 (line 1046) defers the retry mechanism to EXECUTE: "verify Pipecat's `LLMContextAggregator` supports injecting a message + re-running the LLM within a single `process_frame` call. If not, the retry may need to be a separate pipeline task." This is the highest-novelty code in v0.5 and it is on the safety-critical path.
|
||||
- Confidence: 0.70
|
||||
- Challenge: The in-loop guardrail processor is a structural change deferred to EXECUTE. The retry mechanism (inject `RETRY_INSTRUCTION` + re-run LLM) is unvalidated against Pipecat's frame semantics. If Pipecat can't do mid-stream retry, the guardrail's "one retry" (D-068) becomes "canned fallback only" — a weaker safety posture.
|
||||
- Decision: **G-049 (MUST)** — The in-loop guardrail processor's retry mechanism (TASK-05-02) must be validated against Pipecat's frame-processor semantics BEFORE Wave 3 (SLICE-05). Add a Wave-1 or Wave-2 spike task: "Verify `LLMFullResponseEndFrame` fires after the full LLM response + that `LLMContextAggregator` supports injecting a retry message + re-running the LLM within `process_frame`." If Pipecat cannot do mid-stream retry, document the fallback (canned fallback only, no retry) and update D-068's safety posture. This is a binding contract, not an open question. (0.70)
|
||||
|
||||
- **Q2: Integration surface — v0.4 cohort aggregation (D-062), v0.1 voice pipeline (D-061), v0.3 mastery (D-063). Each is an integration point. Risk of quiet cost doubling?**
|
||||
- Evidence: PLAN-v0.5 SLICE-10 (aggregation extension), SLICE-05 (pipeline reuse), SLICE-01 (D-063 schedule_mastery=False); RESEARCH-v0.5 §6.1 — "no schema change to cohort_aggregates (the `metric` column is free-form TEXT)"; §4.3 — "D-063 is unambiguous: assist turns never update θ… `run_mastery_flow()` is invoked only for practice sessions."
|
||||
- Answer: Three integration points, all *additive*:
|
||||
1. **v0.4 cohort aggregation** — new `session_type='assist'` + 5 new metric strings (no schema change, D-062). Risk: low — the aggregator is metric-agnostic (RESEARCH §6.1, 0.90 confidence). But the aggregation cache persistence (v0.4 P1+ #7, REQ-IDEATE-06) directly corrupts `assist_active_learners_count` after restart — the tech-debt wave (SLICE-12) fixes this. **Dependency: the tech-debt fix is on the v0.5 critical path for correct assist metrics.**
|
||||
2. **v0.1 voice pipeline** — `build_assist_pipeline()` reuses services but adds the in-loop guardrail processor (see Q1). Risk: medium — the structural change is the novelty.
|
||||
3. **v0.3 mastery separation** — `schedule_mastery=False` for assist (D-063). Risk: low — the `end()` signature already supports the flag (RESEARCH §4.3, 0.90 confidence). Verified in code: `session_recorder.py` `end()` has `schedule_mastery` param.
|
||||
- The cost-doubling risk is concentrated in the in-loop guardrail processor (Q1). The aggregation + mastery integrations are low-risk additive extensions.
|
||||
- Confidence: 0.75
|
||||
- Decision: **G-050** — 3 integration points, all additive. Cohort aggregation (low risk, metric-agnostic) + mastery separation (low risk, flag exists) + voice pipeline (medium risk, in-loop guardrail is structural). The aggregation cache tech-debt (P1+ #7) is on the critical path for correct assist metrics — SLICE-12 fixes it. Accept with G-049 (guardrail retry validation). (0.75)
|
||||
|
||||
- **Q3: Is there an existing system being replaced? (No — Live Assist is new. But does it inherit v0.1-v0.4 tech debt?)**
|
||||
- Evidence: REVIEW.md:182-203 — 8 v0.4 P1+ findings; REQ-IDEATE-06 (REQUIREMENTS.md:64) — "Carry-forward the 8 v0.4 P1+ findings into the v0.5 backlog as a 'tech-debt wave'"; PLAN-v0.5 SLICE-12 — tech-debt wave (4 tasks).
|
||||
- Answer: No existing system replaced — Live Assist is new. It inherits 8 v0.4 P1+ findings, budgeted in P2 SLICE-12 (REQ-IDEATE-06): (1) argon2id blocking, (2) rate-limit mock test, (3) cookie-secret length, (4) credential status enum, (5) revocation audit log, (6) nightly zoneinfo, (7) aggregation cache persistence, (8) f-string SQL. The most consequential for v0.5 is #7 (aggregation cache) — it directly corrupts `assist_active_learners_count` after restart. The tech-debt wave is in P2 (not P1) — this means the assist metrics are *incorrect* for all of P1 + early P2 until SLICE-12 ships. This is a *deferred fix on the critical path*.
|
||||
- Confidence: 0.72
|
||||
- Challenge: The aggregation cache fix (P1+ #7) is in P2 SLICE-12, but it corrupts v0.5's assist metrics during P1. The plan accepts this (P1 doesn't ship to operators — it's the assist voice loop). But if P1 ships as v0.1.11 (per-phase ship, config.json:110), the assist metrics are wrong in any P1 deployment. This is a *sequencing* issue, not a missing task.
|
||||
- Decision: **G-051** — 8 v0.4 P1+ findings inherited, budgeted in P2 SLICE-12. The aggregation cache fix (P1+ #7) corrupts assist metrics during P1 — accept this because P1 ships the assist *voice loop* (no operator dashboard dependency), and the fix lands in P2 before operator visibility matters. Document in P1 ship notes: assist metrics are incorrect until P2 SLICE-12. (0.72)
|
||||
|
||||
- **Q4: Technical debt being inherited — is it budgeted for?**
|
||||
- Evidence: PLAN-v0.5 SLICE-12 (4 tasks: cache persistence, cookie-secret, credential status, argon2id+rate-limit+audit+zoneinfo); REQ-IDEATE-06 (should priority, P1).
|
||||
- Answer: Yes — budgeted in P2 SLICE-12 (4 tasks covering all 8 findings). The tech-debt wave is `should` priority (not `must`) — this is correct (the findings are non-blocking per REVIEW.md). The budget is 4 tasks in P2 Wave 2 — proportional to the 8 findings (some are one-liners: cookie-secret warning, zoneinfo swap).
|
||||
- Confidence: 0.80
|
||||
- Decision: **G-052** — tech-debt budgeted (4 tasks in P2 SLICE-12, `should` priority). Proportional to the 8 findings. Accept. (0.80)
|
||||
|
||||
---
|
||||
|
||||
### Axis 4 — People, Skills, and Organization
|
||||
|
||||
- **Q1: Key-person dependency — voice-engineer is REACTIVATED for the first time. Is there a knowledge concentration risk?**
|
||||
- Evidence: PERSONAS.md:577-593 — voice-engineer REACTIVATED, owns 7 P1 tasks (largest territory: build_assist_pipeline, in-loop guardrail processor, warm WebRTC, reconnect, tap-to-talk client, latency tuning); PLAN-v0.5:102-108 — persona load distribution.
|
||||
- Answer: The voice-engineer owns the largest P1 territory (7 tasks) and is activated for the *first time* in the project (proposed since v0.2 PERSONAS line 458, never operated). The in-loop guardrail processor + warm WebRTC + reconnect logic are all *new capabilities* this project has never built. If the voice-engineer is absent, the assist voice loop (SLICE-05, SLICE-06) has no owner — these are the core of v0.5. The security-engineer (6 tasks) owns the guardrail regex + tuning corpus — the other safety-critical path. The backend-engineer (6 tasks) owns the session API + context-binding. **Three personas are critical-path: voice-engineer, security-engineer, backend-engineer.** The voice-engineer is the highest key-person risk because the capability is *new* (no prior project experience), not just the territory.
|
||||
- Confidence: 0.78
|
||||
- Decision: **G-053** — key-person dependency: voice-engineer (new capability, largest territory), security-engineer (safety-critical guardrail), backend-engineer (session API + integration). All 3 critical-path. The voice-engineer is the highest risk (first activation, new capability). Accept under parallelization (max 5 concurrent, 5 active personas — exactly at the limit). (0.78)
|
||||
|
||||
- **Q2: Are the 5 active personas actually allocated? (CI agents, not humans. Are the agent capabilities sufficient for the voice-engineer territory?)**
|
||||
- Evidence: config.json:22-27 — parallelization enabled, max 5 concurrent; PERSONAS.md:556-646 — 5 active personas; config.json:52-81 — only 4 personas in config.json array (voice-engineer + security-engineer are emergent, defined in PERSONAS.md).
|
||||
- Answer: 5 active personas, max 5 concurrent — **exactly at the limit, no slack.** If all 5 are active in a wave, there is zero idle capacity for rework. P1 Wave 1 has 2 parallel slices (SLICE-01, SLICE-02) — 2 personas active (backend, backend+voice). P1 Wave 3 has 2 slices (SLICE-05, SLICE-06) — 2 personas (voice, voice). Peak parallelism is 2-3 slices per wave — within the 5-agent limit. The voice-engineer + security-engineer are NOT in config.json `personas` (emergent) — territory enforcement is `warn` (config.json:51), so they are not blocked. The capability question: the voice-engineer's frameworks (porcupine-android, webrtc, pipecat, piper-tts) are listed in PERSONAS.md but the voice-engineer has *never operated* in this project. The capability is *claimed*, not *demonstrated*. The in-loop guardrail processor (Q1, Axis 3) is the test of this capability.
|
||||
- Confidence: 0.72
|
||||
- Decision: **G-054** — 5 active personas, max 5 concurrent (at the limit, no slack). Peak parallelism 2-3 slices — within limit. Voice-engineer capability is claimed but undemonstrated (first activation). Accept with G-049 (guardrail retry validation) as the capability test. (0.72)
|
||||
|
||||
- **Q3: Is there a product owner with authority? (autonomy=full — the CI is the owner. Is that sound for a safety-critical surface?)**
|
||||
- Evidence: config.json:13 — `"level": "full"`; PROJECT.md:5; config.json:34-38 — security auto_accept_low_severity, auto_mitigate_medium, escalate_high_severity.
|
||||
- Answer: CI is the product owner under full autonomy — the established model since v0.1 (G-002, G-015 carry-forward). **For a safety-critical surface, this is the grill's hardest governance question.** The CI can auto-accept low-severity security issues + auto-mitigate medium — but R-ASSIST-07 (guardrail false-negative) is high-severity, and config.json:37 says `escalate_high_severity: true`. The plan *accepts* the residual risk (adversarial FN not threshold-gated) without escalating. This is a tension: the config says escalate high-severity, but the plan says accept. The grill must resolve this — either the residual risk is *not* high-severity (because defense-in-depth + audit + v0.6 LLM-as-judge mitigate it to medium), or the plan must escalate. See Probe 1.
|
||||
- Confidence: 0.68
|
||||
- Challenge: The CI-as-owner model is sound for practice surfaces (v0.1-v0.4) where the worst case is a bad role-play. For Live Assist, the worst case is a guardrail bypass during a real customer call. The config's `escalate_high_severity: true` is the safety valve — the plan must use it or justify why the risk is not high-severity.
|
||||
- Decision: **G-055** — CI is the product owner (full autonomy, carry-forward). For the safety-critical surface, the `escalate_high_severity: true` config (config.json:37) is the governing constraint. R-ASSIST-07 (guardrail false-negative) is high-severity per RESEARCH — the plan must either (a) escalate it (Probe 1) or (b) document why defense-in-depth + audit + v0.6 LLM-as-judge reduce it to medium (auto-mitigatable). This is resolved in Probe 1. (0.68)
|
||||
|
||||
- **Q4: Is the team building capability it doesn't have? (voice-engineer is new — has the guardrail/latency/pipeline work been done before in this project?)**
|
||||
- Evidence: RESEARCH-v0.5 §5.2 — "v0.5 adds an in-loop guardrail processor… the existing v0.1 pipeline doesn't have a post-LLM guardrail processor inline"; §3.3 — "prefill latency for gemma4:cloud is not yet measured (R3 from v0.1)"; PERSONAS.md:577-593 — voice-engineer frameworks include porcupine-android (not used in v0.5 per D-071), webrtc, pipecat.
|
||||
- Answer: Yes — three new capabilities:
|
||||
1. **In-loop Pipecat frame processor** — never built in this project. The v0.1 guardrail runs on the debrief (post-session), not in-loop. The frame-processor semantics (LLMFullResponseEndFrame, mid-stream retry) are unvalidated (G-049).
|
||||
2. **Warm WebRTC connection lifecycle** — v0.1 opens per-session cold connections; v0.5 keeps a warm connection for an 8h shift with heartbeat + reconnect. New state machine (REQ-IDEATE-08).
|
||||
3. **Regex guardrail tuning** — the CS guardrail (customer_service.py, 128 lines) is a fixed ruleset; v0.5 adds a tuning corpus + adversarial test + FP/FN measurement (REQ-IDEATE-01/04). New testing methodology.
|
||||
- All three are on the safety-critical or critical path. This is *acceptable for a pilot* (learning-as-you-go is the project's model since v0.1) but the grill must flag that the highest-novelty code (in-loop processor) is also the highest-safety-impact code.
|
||||
- Confidence: 0.72
|
||||
- Decision: **G-056** — team is building 3 new capabilities (in-loop frame processor, warm WebRTC lifecycle, regex guardrail tuning). All on the safety-critical/critical path. Acceptable for pilot with G-049 (guardrail retry validation) as the de-risking spike. The voice-engineer's first activation is the capability test. (0.72)
|
||||
|
||||
---
|
||||
|
||||
### Axis 5 — Timeline and Estimates
|
||||
|
||||
- **Q1: Was the deadline set before or after the scope was understood? (No deadline — CI pipeline. Is the 2-phase split evidence-based or arbitrary?)**
|
||||
- Evidence: ROADMAP.md:13-31 — v0.5 phases defined in ROADMAP (P0 pre-execution, P1 assist core, P2 integration, P3 review); PLAN-v0.5:17-25 — phase split rationale.
|
||||
- Answer: No calendar deadline (CI pipeline). The 2-phase split is *evidence-based*: P1 = the assist voice loop + guardrail (the safety-critical, on-voice-path surface — 12 REQs, 24 tasks); P2 = integration + measurement + tech-debt (the operator-facing + hardening surface — 4 REQs, 9 tasks). The split mirrors v0.4 (P1 infra / P2 feature) but inverts it (P1 feature / P2 hardening). P1 is independently shippable (a learner can start a shift, tap-to-talk, get coaching with guardrails, end the shift). This is the correct split — the safety-critical surface ships first, the measurement + tech-debt follows.
|
||||
- Confidence: 0.82
|
||||
- Decision: **G-057** — 2-phase split is evidence-based (P1 safety-critical voice loop, P2 hardening + measurement). P1 independently shippable. Not arbitrary. (0.82)
|
||||
|
||||
- **Q2: Critical path — what single thing would push v0.5 by a phase? (Likely the guardrail — REQ-ASSIST-03 is safety-critical. Is the guardrail on the critical path?)**
|
||||
- Evidence: PLAN-v0.5 wave dependency graph (P1:79-98); SLICE-03 (guardrail) → SLICE-04 (tuning corpus) → SLICE-05 (pipeline + in-loop processor) → SLICE-08 (e2e guardrail test); REQ-IDEATE-01 (tuning corpus + adversarial test).
|
||||
- Answer: The guardrail is on the critical path (SLICE-03 → 04 → 05 → 08). The single thing that would push v0.5 by a wave:
|
||||
- **Most likely: the guardrail tuning corpus fails FP<5% or direct-FN<5% (REQ-IDEATE-01).** TASK-04-02 asserts FP<5% on coaching responses + FN<5% on direct answers. If the regex over-matches (FP>5%) or under-matches (FN>5%), the regex needs retuning → pushes Wave 2 → Wave 3 → Wave 4. This is a *test-driven* gate — the tuning corpus is the proof.
|
||||
- **Less likely: the in-loop guardrail processor retry mechanism is infeasible in Pipecat (G-049).** If Pipecat can't do mid-stream retry, the guardrail weakens to "canned fallback only" — still safe, but D-068's "one retry" is unmet. This would push Wave 3 (SLICE-05) by a spike.
|
||||
- **Least likely: the warm WebRTC reconnect state machine (REQ-IDEATE-08).** The reconnect logic is specified (TASK-06-02) but the chaos test (TASK-06-03) is the proof. If the state machine has edge cases, it pushes Wave 3 (SLICE-06).
|
||||
- Confidence: 0.75
|
||||
- Decision: **G-058** — critical-path risk: guardrail tuning corpus (FP/FN rates, REQ-IDEATE-01). Mitigation: TASK-04-02 (test-driven gate). If FP>5% or direct-FN>5%, retune the regex → pushes by a wave. Accept with the test as the gate. G-049 (retry validation) de-risks the secondary path. (0.75)
|
||||
|
||||
- **Q3: Are the estimates evidence-based? (33 tasks across 2 phases — is this analogous to v0.4's 52 tasks/2 phases?)**
|
||||
- Evidence: PLAN-v0.5:1064 — 33 tasks (24 P1 + 9 P2); GRILL-v0.4:166 — v0.4 had 52 tasks (29 P1 + 23 P2); GRILL-v0.4:19 — v0.3 shipped ~40 tasks.
|
||||
- Answer: 33 tasks vs v0.4's 52 (-37%) and v0.3's 40 (-18%). The reduction is explained by D-071 (tap-to-talk only — wake-word deferral removed ~8-10 tasks: Porcupine integration, foreground service, battery management, OEM kill-switch handling) + 0 new deps (no dep-integration tasks). The scope is *smaller* than v0.4 despite +8 REQs (16 vs 8) because the IDEATE additions are mostly test/measurement tasks (low LOC) + the wake-word deferral stripped the client-architecture work. The tasks are bottom-up sized (each slice has 3-7 tasks with acceptance criteria). Evidence-based.
|
||||
- Confidence: 0.80
|
||||
- Decision: **G-059** — 33 tasks is evidence-based (smaller than v0.4's 52 due to D-071 wake-word deferral + 0 new deps; IDEATE additions are test/measurement tasks). Bottom-up sized. Accept. (0.80)
|
||||
|
||||
- **Q4: Definition of done — is "done" the grill's verdict or the verify stage's?**
|
||||
- Evidence: PLAN-v0.5 — per-slice acceptance criteria; ROADMAP.md:19-21 — per-phase ship + verify; config.json:28-33 — verification automated.
|
||||
- Answer: Definition of done = per-slice acceptance criteria + per-phase ship (v0.1.11, v0.1.12, v0.1.13) + verify stage. The grill is the P0 definition of done (this document). Established pattern since v0.2 (G-020 carry-forward). For the safety-critical surface, the *additional* done criterion is REQ-IDEATE-04's measurable NFRs (p95 ≤650ms, FP<5%) — these are the *quantitative* done bar for the guardrail.
|
||||
- Confidence: 0.82
|
||||
- Decision: **G-060** — definition of done = per-slice acceptance + per-phase ship + verify + REQ-IDEATE-04 measurable NFRs (p95 ≤650ms, FP<5%) as the quantitative guardrail bar. Established pattern + safety-critical addition. Accept. (0.82)
|
||||
|
||||
---
|
||||
|
||||
### Axis 6 — Budget and Financial Realism
|
||||
|
||||
- **Q1: Cost drivers — assist mode adds LLM calls (IDEATE-07 — 400 extra calls/month/learner). Is this in the budget?**
|
||||
- Evidence: REQ-IDEATE-07 (REQUIREMENTS.md:70) — "20 turns/shift × 20 shifts/month = 400 extra LLM calls"; PLAN-v0.5 SLICE-11 — per-turn cost tracking + C-3 check; TASK-11-02 — `check_c3_budget()`.
|
||||
- Answer: The cost driver is *budgeted* (SLICE-11, REQ-IDEATE-07). The estimate: 400 extra gemma4:cloud calls/month/learner at ~$0.0005/turn = ~$0.20/month — well under C-3's $3 (RESEARCH-v0.5, TASK-11-02). The cost is *diagnostic* (not enforced — D-012 says no enforced ceiling for pilot). The C-3 check (TASK-11-02) flags if practice + assist exceeds $3. This is the correct posture — measure, don't enforce, for the pilot.
|
||||
- Confidence: 0.80
|
||||
- Decision: **G-061** — assist cost driver budgeted (SLICE-11, ~$0.20/month, well under C-3). Diagnostic, not enforced (D-012 pilot relaxation). Accept. (0.80)
|
||||
|
||||
- **Q2: C-3 (≤$3/active learner/month) — does assist break it? (D-012 relaxed C-3 for the pilot, but is the relaxation still valid for v0.5?)**
|
||||
- Evidence: D-012 (PROJECT.md:182) — "v0.1 cost ceiling = no enforced ceiling (pilot)"; GRILL-v0.4 G-012 — "no TLS → accepted as pilot-scale constraint"; REQ-IDEATE-07 — C-3 check.
|
||||
- Answer: The C-3 relaxation (D-012) was set for v0.1 and carried through v0.4 (G-012). v0.5 adds ~$0.20/month/learner for assist — the total (practice + assist) is still well under $3 at pilot scale. The relaxation remains valid *for the pilot*. The architecture must not preclude meeting $3 post-pilot (D-012) — the assist cost is LLM calls, which the post-pilot path (self-hosted gemma4:e4b, D-020) reduces. The relaxation is valid for v0.5.
|
||||
- Confidence: 0.78
|
||||
- Decision: **G-062** — C-3 relaxation (D-012) remains valid for v0.5 pilot. Assist adds ~$0.20/month, total well under $3. Post-pilot path (self-hosted model) preserves the $3 target. Accept. (0.78)
|
||||
|
||||
- **Q3: Burn rate — token cost of 33 tasks + 2 phases + grill + review + audit. Is this proportional to v0.4?**
|
||||
- Evidence: git log — v0.4 shipped in ~1.3 days (GRILL-v0.4 G-023); v0.5 has 33 tasks vs v0.4's 52 (-37%).
|
||||
- Answer: v0.5 is ~37% smaller than v0.4 by task count. Expected burn: ~0.8-1.0 days of CI agent time (proportional reduction). The token cost is the CI agent's operational cost — not tracked, but the pace is established (4 milestones in ~4 days). Proportional.
|
||||
- Confidence: 0.78
|
||||
- Decision: **G-063** — burn rate: ~0.8-1.0 days estimated (proportional to v0.4, -37% tasks). Accept. (0.78)
|
||||
|
||||
- **Q4: Is the budget contingent on anything? (Porcupine pricing D-064 — MAU-priced, no recurring free tier. Is the pilot contingent on Picovoice sales engagement?)**
|
||||
- Evidence: D-064 (PROJECT.md:234) — Porcupine MAU pricing; D-071 (PROJECT.md:241) — tap-to-talk only in v0.5, wake-word deferred to v0.6; R-ASSIST-01 (RESEARCH-v0.5 §1.2) — "no recurring free tier."
|
||||
- Answer: **No — D-071 removed the Picovoice contingency.** The wake-word (Porcupine) is deferred to v0.6. v0.5 ships tap-to-talk only — no Porcupine dependency, no MAU pricing, no sales engagement needed. This is the single biggest budget de-risking of v0.5: the entire Picovoice commercial question is v0.6's problem, not v0.5's. The v0.5 budget is contingent on *nothing* external (0 new deps, no vendor engagement, full autonomy).
|
||||
- Confidence: 0.85
|
||||
- Decision: **G-064** — no budget contingency. D-071 (tap-to-talk only) removed the Picovoice MAU-pricing dependency. v0.5 has 0 external commercial dependencies. Accept. (0.85)
|
||||
|
||||
---
|
||||
|
||||
### Axis 7 — Risks, Assumptions, and Dependencies
|
||||
|
||||
- **Q1: Top 3 assumptions — evidence for each?**
|
||||
- Evidence: RESEARCH-v0.5 risks (R-ASSIST-01..14); D-071, D-068, D-072.
|
||||
- Answer:
|
||||
1. **Tap-to-talk is sufficient UX (D-071).** Evidence: none — this is an *unvalidated* assumption. No user testing, no pilot data. The practice surface (v0.1-v0.4) uses a WebRTC connection per session; tap-to-talk is a button-hold pattern. Whether a learner on a real shift will tap a button on their phone (which may be in their pocket) is *untested*. The alternative (wake-word) is deferred to v0.6. **Confidence: 0.60** — the assumption is reasonable (tap-to-talk is a proven pattern for walkie-talkie apps) but unvalidated for this use case.
|
||||
2. **Regex guardrail is adequate (D-068).** Evidence: RESEARCH §2.3 (0.78 confidence) — the regex patterns target direct-answer + false-authority + impersonation. The tuning corpus (REQ-IDEATE-01) + adversarial test will measure FP/FN. The adversarial FN rate is "reported but not threshold-gated" (PLAN:419) — this is a *residual risk acceptance*, not a proof of adequacy. **Confidence: 0.65** — the regex is the fast on-voice-path filter; the LLM-as-judge (v0.6) is the accurate off-voice-path backstop. Defense-in-depth is the mitigation, not regex alone.
|
||||
3. **≤650ms latency is achievable (D-072).** Evidence: RESEARCH §3.3 — estimated ~655ms (Piper + lean prompt), unmeasured. The estimate is a *budget math* calculation, not a measurement. R1/R3/R4 (Deepgram/Ollama/Piper latencies) are unmeasured since v0.1. **Confidence: 0.65** — the budget math is sound but the actual latencies are unmeasured. D-072 accepts ≤650ms as pilot tolerance; <600ms is v0.6 hardening.
|
||||
- Confidence: 0.63
|
||||
- Decision: **G-065** — 3 core assumptions: tap-to-talk UX (0.60, unvalidated), regex guardrail adequacy (0.65, residual risk accepted), ≤650ms latency (0.65, unmeasured). All accepted as pilot-scale constraints with v0.6 hardening paths. The tap-to-talk assumption is the lowest-confidence — flag for v0.6 user testing. (0.63)
|
||||
|
||||
- **Q2: Dependencies — Picovoice (D-064, deferred to v0.6), PIPEDA (D-073), v0.4 cohort pipeline (D-062), v0.1 voice pipeline (D-061).**
|
||||
- Evidence: D-071 (Picovoice deferred), D-073 (PIPEDA deferred), D-062 (cohort aggregation), D-061 (voice pipeline reuse).
|
||||
- Answer:
|
||||
- **Picovoice**: NOT a v0.5 dependency (D-071 — tap-to-talk only). Deferred to v0.6. ✅
|
||||
- **PIPEDA**: Deferred to "Phase 1 implementation" (D-073). This is the escalation (ESCALATION-01, Axis 2). The disclosure (D-070) is the engineering mitigation. ⚠️
|
||||
- **v0.4 cohort pipeline**: D-062 — additive extension (session_type=assist, new metric strings, no schema change). Verified: aggregator.py is metric-agnostic (RESEARCH §6.1, 0.90). ✅
|
||||
- **v0.1 voice pipeline**: D-061 — service reuse (transport/stt/llm/tts) + in-loop guardrail processor (structural change, G-049). ⚠️
|
||||
- The PIPEDA dependency is the only one that requires human attention. The others are internal + additive.
|
||||
- Confidence: 0.75
|
||||
- Decision: **G-066** — 4 dependencies: Picovoice (deferred, ✅), PIPEDA (escalation, ⚠️ — ESCALATION-01), cohort pipeline (additive, ✅), voice pipeline (structural change, ⚠️ — G-049). Accept the internal dependencies; escalate PIPEDA. (0.75)
|
||||
|
||||
- **Q3: Single risk that kills v0.5? (R-ASSIST-07 — guardrail false-negative reaches learner's ear during real customer call. Is there a mitigation beyond "defense-in-depth + post-v0.5 LLM-as-judge"?)**
|
||||
- Evidence: R-ASSIST-07 (RESEARCH-v0.5 §2.6) — "The 'parrot' failure: the AI gives a verbatim script, the learner repeats it word-for-word, the customer detects the robotic delivery → trust erosion"; PLAN-v0.5:1025 — "defense-in-depth (prompt + regex + audit) + adversarial test + nightly FN trending + post-v0.5 LLM-as-judge (REQ-IDEATE-10, v0.6)"; PLAN:419 — "adversarial FN rate is reported but not threshold-gated."
|
||||
- Answer: R-ASSIST-07 is the single project-killing risk. A direct answer that slips past the regex → learner parrots it → real customer hears robotic delivery → trust erosion + potential escalation. The mitigation is *defense-in-depth* (3 layers: prompt + regex + audit) + *measurement* (tuning corpus + adversarial test + nightly FN trending) + *future backstop* (v0.6 LLM-as-judge). **The gap: the adversarial FN rate is "reported but not threshold-gated" (PLAN:419).** This means the plan *accepts* an unknown residual risk without a ceiling. For a safety-critical surface, this is insufficient — the grill must set the bar. The bar cannot be "0% FN" (regex can't catch every paraphrase) — but it must be a *documented acceptance threshold* with an escalation if exceeded. config.json:37 says `escalate_high_severity: true` — R-ASSIST-07 is high-severity, so the plan must either escalate or document why the residual risk is acceptable.
|
||||
- Confidence: 0.68
|
||||
- Challenge: The plan accepts an unquantified residual risk on a safety-critical surface. "We'll measure it and trend it nightly" is necessary but not sufficient — what happens if the nightly trend shows 15% FN? The plan has no trigger. This is the grill's hardest call.
|
||||
- Decision: **G-067 (MUST)** — R-ASSIST-07 (guardrail false-negative) must have a *documented acceptance threshold* before EXECUTE. The adversarial FN rate (REQ-IDEATE-01) must be: (a) measured pre-ship (TASK-04-02), (b) compared against a threshold (e.g., "adversarial FN ≤ 20% acceptable for pilot because defense-in-depth + audit + v0.6 LLM-as-judge mitigate; >20% triggers a re-tuning wave or escalation"), (c) the threshold + the mitigation rationale documented in the ship notes. This is NOT a "0% FN" demand — it is a "know your residual risk + decide if it's acceptable" demand. The plan's current "reported but not threshold-gated" is insufficient for a safety-critical surface. config.json:37 `escalate_high_severity: true` is the governing constraint. (0.68)
|
||||
|
||||
- **Q4: Pre-mortem — "It's 12 months from now and v0.5 failed. Why?"**
|
||||
- Evidence: RESEARCH-v0.5 risks; PLAN-v0.5 risk matrix.
|
||||
- Answer: The most likely failure modes (in order):
|
||||
1. **A guardrail bypass incident during a real customer call (R-ASSIST-07).** A direct answer slipped past the regex, the learner parroted it, the customer escalated to a real manager who disavowed the "AI's advice." The nightly FN trend showed 18% but no one acted because there was no threshold (G-067 gap). This is the *highest-consequence* failure — it breaks trust in the product + the learner's job.
|
||||
2. **PIPEDA complaint (R-ASSIST-08 / D-073).** A real customer discovered they were recorded by the learner's mic without their consent. The disclosure (D-070) was shown to the *learner*, not the *customer*. Canada's two-party consent law (if applicable in the province) was not reviewed. This is the *highest-legal-consequence* failure.
|
||||
3. **The in-loop guardrail processor's retry mechanism was infeasible in Pipecat (G-049).** The "one retry" (D-068) became "canned fallback only" — safe but degraded. The assist coaching quality dropped (every block → canned fallback, no second chance). Learners stopped using assist because the coaching felt robotic.
|
||||
4. **The latency was >650ms in practice (R-ASSIST-02).** The ~655ms estimate was optimistic; actual p95 was ~720ms. Coaching arrived after the customer moment passed. Learners abandoned assist for being "too slow to be useful."
|
||||
- Confidence: 0.75
|
||||
- Decision: **G-068** — pre-mortem top-4: guardrail bypass (highest consequence, G-067 gap), PIPEDA complaint (ESCALATION-01), in-loop retry infeasible (G-049), latency >650ms (D-072 pilot tolerance). All four are addressed in binding decisions/escalations. (0.75)
|
||||
|
||||
---
|
||||
|
||||
### Axis 8 — Governance, Decision-Making, and Communication
|
||||
|
||||
- **Q1: Decision-maker — autonomy=full, the CI decides. Is there a human escalation path for safety-critical decisions? (config.json escalation_hooks: deploy, delete_data, merge_to_main — none for "ship safety-critical guardrail". Is this a gap?)**
|
||||
- Evidence: config.json:14 — `"escalation_hooks": ["deploy", "delete_data", "merge_to_main"]`; config.json:37 — `"escalate_high_severity": true`; PROJECT.md:5 — "Autonomy: full."
|
||||
- Answer: The escalation_hooks list does NOT include "ship safety-critical guardrail" or "legal review." The `escalate_high_severity: true` security config is the *only* safety valve — it says the CI *should* escalate high-severity security issues, but the *mechanism* (how? to whom?) is unspecified. For v0.1-v0.4 (practice surface), this was acceptable — the worst case was a bad role-play. For v0.5 (Live Assist, real customers), the worst case is a guardrail bypass during a real call + a PIPEDA complaint. The escalation path for these is *the grill itself* — this document is the escalation mechanism. The grill's ESCALATION-01 (PIPEDA) + G-067 (guardrail threshold) are the safety-critical escalations/binding decisions. **The gap: there is no *ongoing* human escalation path post-ship.** If the nightly FN trend spikes post-ship, the CI auto-mitigates (config.json:36) but does not escalate to a human (no hook for "safety signal spike"). This is a v0.6+ governance gap, not a v0.5 blocker — v0.5 ships the measurement (REQ-IDEATE-04 nightly trending); v0.6 adds the LLM-as-judge + the escalation on spike.
|
||||
- Confidence: 0.70
|
||||
- Decision: **G-069** — escalation path: the grill is the safety-critical escalation mechanism (ESCALATION-01 + G-067). config.json `escalate_high_severity: true` is the governing constraint. Post-ship ongoing escalation (safety signal spike → human) is a v0.6+ governance gap — v0.5 ships the measurement, v0.6 adds the response. Accept for pilot with documented gap. (0.70)
|
||||
|
||||
- **Q2: Governance cadence — the pipeline stages are the governance. Is the grill the right gate for a safety-critical surface?**
|
||||
- Evidence: ROADMAP.md:21 — "Pipeline stages: SPECIFY → CLARIFY → RESEARCH → IDEATE → PLAN → GRILL → SHIP"; ROADMAP.md:30 — "GRILL-v0.5.md (adversarial review — real-customer interaction warrants grill)."
|
||||
- Answer: The grill is the right gate — ROADMAP.md:30 explicitly flags "real-customer interaction warrants grill." The pipeline stages (SPECIFY→…→GRILL→SHIP) are the governance cadence; the grill is the crisis-cadence (this document). For a safety-critical surface, the grill is the *only* human-in-the-loop checkpoint (the CI runs the rest autonomously). This is the correct model — the grill surfaces the safety-critical decisions (G-067, ESCALATION-01) for human attention before SHIP.
|
||||
- Confidence: 0.82
|
||||
- Decision: **G-070** — grill is the right gate for a safety-critical surface (ROADMAP:30 explicit). The grill is the human-in-the-loop checkpoint. Accept. (0.82)
|
||||
|
||||
- **Q3: What's omitted from status reports? (The LSP errors in server/__main__.py, test_scenario_library.py — are these reported or hidden?)**
|
||||
- Evidence: Task context mentions "LSP errors in server/__main__.py, test_scenario_library.py"; verification: `python3 -m py_compile server/__main__.py` → exit 0 (clean); `python3 -m py_compile tests/test_scenario_library.py` → exit 0 (clean).
|
||||
- Answer: The "LSP errors" claim in the task context is **unverified** — both files compile cleanly (`py_compile` exit 0). This may refer to type-checking (pyright/mypy) warnings, not syntax errors, or it may be stale. The grill does not flag this as a material omission — the files compile, the v0.4 tests pass (317 pass, 0 fail per REVIEW.md). If there are type-checking warnings, they are non-blocking (the codebase doesn't enforce strict typing in CI). **No omission found.**
|
||||
- Confidence: 0.80
|
||||
- Decision: **G-071** — no status-report omission found. The "LSP errors" claim is unverified (files compile clean). Type-checking warnings, if any, are non-blocking. Accept. (0.80)
|
||||
|
||||
- **Q4: Stop-the-project trigger — is there one? (If the grill returns RETHINK, does the pipeline stop?)**
|
||||
- Evidence: config.json:13 — full autonomy; GRILL-v0.4 G-032 — "no human stop trigger (full autonomy). The grill is the stop mechanism."
|
||||
- Answer: No human stop trigger (full autonomy, G-032 carry-forward). The grill is the stop mechanism — if the verdict were "Rethink" or "Escalate" on a material axis, the pipeline would stop. This grill's verdict is "Proceed-with-conditions" — the project proceeds after the MUSTs (G-049, G-067) + the escalation (ESCALATION-01) are resolved. The escalation (PIPEDA) is the *de facto* stop trigger — if the human legal review determines the disclosure is insufficient, v0.5 cannot ship the assist surface as designed.
|
||||
- Confidence: 0.78
|
||||
- Decision: **G-072** — no human stop trigger (full autonomy). The grill is the stop mechanism. ESCALATION-01 (PIPEDA) is the de facto stop trigger for the assist surface. This grill = proceed with conditions. (0.78)
|
||||
|
||||
---
|
||||
|
||||
### Axis 9 — Change, Adoption, and Operational Readiness
|
||||
|
||||
- **Q1: Who uses Live Assist? (The learner — during a real shift. How does their work change? They now have an AI in their ear.)**
|
||||
- Evidence: PROJECT.md:45-47 — "a hands-free voice assistant a learner invokes *while actually working*"; PERSONAS.md — no learner persona (learners are external to the CI agent); D-071 — tap-to-talk invocation.
|
||||
- Answer: The learner uses Live Assist during a real shift. Their work changes: they now have an AI coach in their ear (via earbuds) that they invoke by tapping a button (D-071 — tap-to-talk, not wake-word). "What's in it for them" = real-time coaching during real customer interactions — the transfer moment from practice to job. **This is unvalidated** — no user testing, no pilot data on whether learners will actually tap a button on their phone during a real customer call (the phone may be in their pocket, the tap may be socially awkward). The tap-to-talk UX (D-071) is the lowest-confidence assumption (G-065, 0.60). The alternative (wake-word, hands-free) is deferred to v0.6. For v0.5 pilot, tap-to-talk is the *validation* — does a learner use it? The measurement is the assist usage metrics (REQ-NFR-ASSIST-04, cohort aggregation).
|
||||
- Confidence: 0.65
|
||||
- Challenge: The adoption risk is *real* — tap-to-talk during a real customer call is socially + ergonomically awkward (phone in pocket, earbuds in, tap a button on the phone screen). The "we'll measure usage" answer is correct but the pilot may show low adoption. This is a v0.5 *validation* risk, not a v0.5 *blocker*.
|
||||
- Decision: **G-073** — Live Assist's first user is the learner during a real shift. Tap-to-talk (D-071) is the unvalidated UX assumption (G-065, 0.60). v0.5 pilot *validates* adoption (assist usage metrics); v0.6 adds wake-word if tap-to-talk adoption is low. Document in ship notes: v0.5 validates the coaching/guardrail/context-binding value, not the hands-free UX (that's v0.6). (0.65)
|
||||
|
||||
- **Q2: Is the ops team involved? (CI project — ops is the LXC deploy. Does v0.5 need deploy changes? D-071 says no — v0.4 LXC carries forward. Is that sound?)**
|
||||
- Evidence: PERSONAS.md:651-660 — devops-engineer DEACTIVATED for v0.5 ("No deploy changes — v0.4's LXC + Docker-in-LXC + Postgres + backup cron carries forward unchanged"); PLAN-v0.5:1071 — "New pip deps: 0… New npm deps: 0."
|
||||
- Answer: v0.5 needs NO deploy changes — 0 new pip deps, 0 new npm deps, no new Docker services, no CT bump. The assist surface is server-side code (server/assist/) + a React route (client/src/AssistControl.tsx) on the existing v0.4 LXC. devops-engineer deactivation is sound. The ops surface (LXC, Postgres, backup) is unchanged. This is the correct posture — v0.5 is a *feature* milestone, not an *infra* milestone.
|
||||
- Confidence: 0.85
|
||||
- Decision: **G-074** — v0.5 needs no deploy changes (0 new deps, no CT bump, v0.4 LXC carries forward). devops-engineer deactivation is sound. Accept. (0.85)
|
||||
|
||||
- **Q3: Rollback plan — if v0.5 ships and a guardrail incident occurs, what's the rollback? (Disable assist mode? Revert to v0.1.9?)**
|
||||
- Evidence: config.json:40 — `"branching_strategy": "phase"`; PLAN-v0.5 — per-phase ship (v0.1.11, v0.1.12, v0.1.13); git revert pattern (GRILL-v0.4 G-035).
|
||||
- Answer: Rollback is per-phase git revert (G-035 carry-forward). But for a *guardrail incident* (R-ASSIST-07), the rollback is *operational*, not just git:
|
||||
- **Preventive rollback**: disable assist mode (revert to v0.1.9 = v0.4). The assist routes (`/api/assist/*`) + the assist WebRTC endpoint are removed. The practice surface (v0.1-v0.4) continues unchanged. This is a clean revert — the assist surface is additive (new routes, new server/assist/ package, new SQLite migration 0004). Reverting removes the routes + the package; the migration is additive (session_type defaults to 'practice', guardrail_verdict_json is nullable) so existing practice sessions are unaffected.
|
||||
- **Corrective rollback**: impossible. Once a guardrail bypass reaches a learner's ear during a real call, the turn has played. The audit log (REQ-IDEATE-09 incremental write) records it for investigation, but the *incident* cannot be rolled back. This is the nature of a live surface — rollback is preventive (disable), not corrective.
|
||||
- The preventive rollback (disable assist) is clean + tested (the assist surface is additive). The corrective impossibility is accepted (the audit log is the post-incident tool, not a rollback).
|
||||
- Confidence: 0.75
|
||||
- Decision: **G-075** — rollback is preventive (disable assist mode → revert to v0.1.9). The assist surface is additive (clean revert). Corrective rollback is impossible (a live turn cannot be un-played) — the audit log (REQ-IDEATE-09) is the post-incident tool. Accept the preventive-only rollback. (0.75)
|
||||
|
||||
- **Q4: Has anyone validated the success criteria with the people who will judge v0.5 successful? (NFRs are research-grounded, not measurement-validated.)**
|
||||
- Evidence: REQUIREMENTS.md:22-25 — NFRs `research-grounded`; REQ-IDEATE-04 — measurable targets (p95 ≤650ms, FP<5%); config.json:13 — full autonomy (CI is the judge).
|
||||
- Answer: No human judge (full autonomy, G-036 carry-forward). The CI is the judge. The success criteria = 16/16 REQ coverage + per-slice acceptance + REQ-IDEATE-04 measurable NFRs. The NFRs are *research-grounded* (estimated, not measured) — REQ-IDEATE-04 + SLICE-09 (P2) add the *measurement*. The validation path: P2 SLICE-09 measures p95 latency + FP/FN rates. If p95 >650ms or FP>5%, the P2 verify stage flags it. This is the *measurement-validated* path — but it happens in P2, not pre-ship. **Gap: the success criteria are validated *during* P2, not *before* P1 ship (v0.1.11).** If P1 ships with a guardrail that has FP>5%, the P1 ship is premature. The mitigation: TASK-04-02 (guardrail tuning test) is in P1 Wave 2 — it runs *before* P1 ship. If it fails, P1 doesn't ship. This is the correct gate.
|
||||
- Confidence: 0.72
|
||||
- Decision: **G-076** — success criteria are research-grounded, measurement-validated in P2 (SLICE-09). The P1 gate is TASK-04-02 (guardrail tuning test, FP<5% / direct-FN<5%) — runs before P1 ship. If it fails, P1 doesn't ship. Accept with TASK-04-02 as the P1 gate + SLICE-09 as the P2 measurement. (0.72)
|
||||
|
||||
---
|
||||
|
||||
### Meta — Closing Review
|
||||
|
||||
- **Q1: If you were the auditor, what would you flag?**
|
||||
- Evidence: all axes above.
|
||||
- Answer: Four flags:
|
||||
1. **R-ASSIST-07 residual risk acceptance without a threshold (G-067).** The plan accepts an unquantified adversarial FN rate on a safety-critical surface. This is the grill's hardest call — the bar must be set.
|
||||
2. **PIPEDA legal review deferred (ESCALATION-01).** Shipping a recording device into real customer interactions without legal sign-off is a regulatory risk the CI cannot own.
|
||||
3. **IDEATE scope expansion +128% (G-046).** The first use of ideation expanded v0.5 from 7 to 16 REQs. The additions are defensive, but the expansion is the largest in project history — future ideation must maintain risk-reduction discipline.
|
||||
4. **In-loop guardrail processor is a structural pipeline change (G-049).** The research frames it as "~1 new frame processor" but the retry mechanism is unvalidated against Pipecat semantics. This is the highest-novelty code on the safety-critical path.
|
||||
- Confidence: 0.78
|
||||
- Decision: **G-077** — auditor flags: R-ASSIST-07 threshold gap, PIPEDA escalation, IDEATE scope expansion, in-loop processor novelty. All addressed in binding decisions/escalations. (0.78)
|
||||
|
||||
- **Q2: What is v0.5 NOT doing that it should? (PIPEDA legal review is deferred D-073 — should it block ship?)**
|
||||
- Evidence: D-073 (PROJECT.md:243); ESCALATION-01 (Axis 2).
|
||||
- Answer:
|
||||
1. **PIPEDA legal review** — deferred, escalated (ESCALATION-01). The grill cannot determine if it blocks ship — that's a legal question. The disclosure (D-070) is the engineering mitigation; the legal review is the *regulatory* mitigation.
|
||||
2. **Post-ship safety signal escalation** — the nightly FN trend (REQ-IDEATE-04) measures but does not escalate on spike (G-069). v0.6 adds the LLM-as-judge + the escalation response.
|
||||
3. **Guardrail red-team prompt set** — REQ-IDEATE-01 builds a *synthetic* tuning corpus (LLM-generated coaching vs direct-answer responses). This is NOT a *human red-team* prompt set — a determined adversary (or a clever learner) may find paraphrases the synthetic corpus doesn't cover. The adversarial test (TASK-04-02) is the best available, but it's synthetic, not human. This is an accepted limitation (pilot).
|
||||
- Confidence: 0.75
|
||||
- Decision: **G-078** — v0.5 is NOT doing: PIPEDA legal review (escalated), post-ship safety escalation (v0.6), human red-team prompt set (synthetic corpus accepted for pilot). All documented. Accept with ESCALATION-01 as the human-action item. (0.75)
|
||||
|
||||
- **Q3: Simplest possible version — is tap-to-talk-only (D-071) the 80/20, or is even that over-scoped?**
|
||||
- Evidence: D-071 (PROJECT.md:241); PLAN-v0.5:25 — "Why tap-to-talk only in v0.5."
|
||||
- Answer: Tap-to-talk-only (D-071) IS the 80/20. The wake-word deferral stripped the client-architecture risk (React Native), the battery risk, the Picovoice commercial risk, and 5 of 14 research risks. What remains is the *core* value: the guardrail (safety), the context-binding (relevance), the shift-bounded session (aggregation), the in-loop processor (pipeline integration). Cutting further (e.g., deferring the guardrail tuning corpus REQ-IDEATE-01) would ship a guardrail "blind on its two most safety-critical metrics" (FP/FN) — unacceptable for a safety-critical surface. The 16 REQs are the *minimum* for a safe assist pilot. **Tap-to-talk is the 80/20; the guardrail work is the non-negotiable 20%.**
|
||||
- Confidence: 0.80
|
||||
- Decision: **G-079** — tap-to-talk-only (D-071) is the 80/20. The guardrail work (REQ-IDEATE-01/04, the tuning corpus + measurable NFRs) is the non-negotiable 20%. Cutting further would ship an unvalidated safety-critical surface. Accept the scope. (0.80)
|
||||
|
||||
- **Q4: What would have to be true for v0.5 to succeed in the next 90 days, and is it true today?**
|
||||
- Evidence: all axes.
|
||||
- Answer: For v0.5 to succeed:
|
||||
1. **The guardrail must be robust (FP<5%, direct-FN<5%, adversarial FN known + thresholded).** Is it true today? No — the tuning corpus is built in P1 Wave 2 (TASK-04-01/02). Will be true after P1 if the test passes. G-067 sets the threshold.
|
||||
2. **The in-loop guardrail processor must work in Pipecat (retry mechanism).** Is it true today? No — unvalidated (G-049). Will be true after the Wave-1/2 spike.
|
||||
3. **PIPEDA must be addressed (legal review or disclosure-sufficient determination).** Is it true today? No — deferred (ESCALATION-01). Will be true only after human legal review.
|
||||
4. **The latency must be ≤650ms.** Is it true today? No — unmeasured (D-072). Will be true after P2 SLICE-09 measurement.
|
||||
5. **The tap-to-talk UX must be usable during a real shift.** Is it true today? No — unvalidated (G-065). Will be true only after pilot deployment (v0.5's validation purpose).
|
||||
- 2 of 5 are addressable in P1/P2 (guardrail robustness, in-loop processor). 1 requires human action (PIPEDA). 2 are post-ship validation (latency measurement, UX adoption). This is the expected state for a pilot — the *plan* is ready; the *proof* is in execution.
|
||||
- Confidence: 0.72
|
||||
- Decision: **G-080** — 5 success conditions: guardrail robustness (P1 gate, G-067), in-loop processor (P1 spike, G-049), PIPEDA (human escalation, ESCALATION-01), latency (P2 measurement), UX adoption (post-ship validation). 2 addressable in P1/P2, 1 requires human, 2 post-ship. Accept — the plan is ready, the proof is in execution. (0.72)
|
||||
|
||||
---
|
||||
|
||||
### v0.5-Specific Probes (Signature Questions)
|
||||
|
||||
#### Probe 1 — R-ASSIST-07 (Guardrail false-negative): Is "defense-in-depth + audit + v0.6 LLM-as-judge" enough for a safety-critical surface?
|
||||
|
||||
**Question:** The AI is in a learner's ear during a *real* customer call. The regex output filter (D-068) is the on-voice-path guardrail. The adversarial FN rate is "reported but not threshold-gated" (PLAN:419). If a direct answer slips past the regex, the learner may parrot it. Is the 3-layer defense (prompt + regex + audit) + nightly trending + v0.6 LLM-as-judge sufficient, or does the grill need to set a binding threshold?
|
||||
|
||||
**Evidence:**
|
||||
- R-ASSIST-07 (RESEARCH-v0.5 §2.6) — "The 'parrot' failure: the AI gives a verbatim script, the learner repeats it word-for-word, the customer detects the robotic delivery → trust erosion."
|
||||
- D-068 (PROJECT.md:238) — "regex-based direct-answer + false-authority + impersonation patterns, with one retry on block + canned coaching redirect fallback."
|
||||
- PLAN-v0.5:419 — "The adversarial FN rate is reported but not threshold-gated (it's the residual risk, mitigated by defense-in-depth)."
|
||||
- config.json:37 — `"escalate_high_severity": true`.
|
||||
- REQ-IDEATE-10 (v0.6 backlog) — "LLM-as-judge guardrail evaluation (nightly, off-voice-path) — measure the true false-negative rate the regex filter cannot."
|
||||
|
||||
**Analysis:**
|
||||
The plan's posture is: regex is the fast on-voice-path filter (D-068); the LLM-as-judge is the accurate off-voice-path backstop (v0.6, REQ-IDEATE-10). The *gap* is v0.5: the regex is the only on-voice-path guardrail, and its adversarial FN rate is *unthresholded*. For a safety-critical surface where the worst case is a guardrail bypass during a real customer call, "we'll measure it and trend it nightly" is necessary but not sufficient — the plan needs a *decision*: what FN rate is acceptable for the pilot, and what happens if it's exceeded?
|
||||
|
||||
The config says `escalate_high_severity: true` — R-ASSIST-07 is high-severity. The plan *accepts* the residual risk without escalating. This is the tension G-055 identified. The resolution: the grill sets the threshold (G-067) — the adversarial FN rate must be measured pre-ship (TASK-04-02), compared against a documented threshold, and the threshold + mitigation rationale documented in the ship notes. This is NOT a "0% FN" demand (impossible for regex) — it is a "know your residual risk + decide if it's acceptable" demand.
|
||||
|
||||
The defense-in-depth (prompt + regex + audit) is the *correct* architecture — the grill does not dispute the 3-layer pattern (RESEARCH §2.1, 0.85 confidence). The issue is the *threshold*, not the architecture. The v0.6 LLM-as-judge is the *future* backstop, not the *current* mitigation — v0.5 ships with regex + audit only.
|
||||
|
||||
**Verdict:** Defense-in-depth is the correct architecture; the missing piece is a *documented acceptance threshold* for the adversarial FN rate. G-067 (MUST) sets this. The plan's "reported but not threshold-gated" is insufficient for a safety-critical surface — the grill requires a threshold + an escalation if exceeded. **Confidence: 0.68.**
|
||||
|
||||
---
|
||||
|
||||
#### Probe 2 — D-073 (PIPEDA consent-law review): Should legal review block ship?
|
||||
|
||||
**Question:** The ambient mic captures the real customer (a third party). ASR transcribes their speech. The turns table stores it (REQ-IDEATE-05). Canada's PIPEDA + provincial consent laws govern recording. D-073 defers the legal review to "Phase 1 implementation." The disclosure (D-070) is shown to the *learner*, not the *customer*. Is the disclosure sufficient, or does the legal review need to block ship?
|
||||
|
||||
**Evidence:**
|
||||
- D-073 (PROJECT.md:243) — "PIPEDA consent-law review = defer to v0.5 Phase 1 implementation; document as R-ASSIST-08 in the grill."
|
||||
- D-070 (PROJECT.md:240) — consent disclosure: "Praxis Assist is on — those around you may be recorded by your mic."
|
||||
- R-ASSIST-08 (RESEARCH-v0.5 §2.6) — "the real customer didn't consent to being recorded/analyzed by an AI."
|
||||
- REQ-IDEATE-05 (REQUIREMENTS.md:52) — "The ambient mic captures BOTH the learner and the real customer; ASR transcribes both; the turns table stores transcribed text. The customer is a third party."
|
||||
- config.json:13 — full autonomy (CI cannot resolve legal questions).
|
||||
|
||||
**Analysis:**
|
||||
This is a *legal* question, not a technical one. The CI agent under full autonomy cannot determine whether Canada's PIPEDA + provincial consent law requires:
|
||||
- (a) One-party consent (the learner's consent is sufficient — the disclosure D-070 covers this).
|
||||
- (b) Two-party consent (the *customer* must consent — Praxis cannot notify the customer, so the assist surface may be illegal in two-party provinces).
|
||||
- (c) A PIPEDA-compliant privacy policy + data handling agreement.
|
||||
|
||||
The disclosure (D-070) is the *engineering* mitigation — it makes the *learner* aware. It does NOT make the *customer* aware, and it does NOT determine the legal consent regime. The PII policy (REQ-IDEATE-05) retains customer speech with redaction + 30-day retention — this is a *data handling* mitigation, not a *consent* determination.
|
||||
|
||||
The grill's confidence that the disclosure is sufficient: **0.55** — below the 0.60 threshold. The grill cannot resolve this under full autonomy. This is an escalation.
|
||||
|
||||
**Verdict:** PIPEDA legal review is a hidden regulatory requirement that the CI cannot resolve. The disclosure (D-070) is the engineering mitigation but not a legal determination. **Escalate to human attention** (ESCALATION-01): determine whether the disclosure is legally sufficient or whether two-party consent / a PIPEDA privacy policy is required before ship. If the disclosure is sufficient, proceed; if not, the assist surface may need geographic restriction or customer-facing consent (out of scope for v0.5). **Confidence: 0.55 — below threshold, escalated.**
|
||||
|
||||
---
|
||||
|
||||
#### Probe 3 — IDEATE scope expansion (+128%): Risk-reduction or scope creep?
|
||||
|
||||
**Question:** v0.5 started with 7 REQs (3 ASSIST + 4 NFR, post-CLARIFY). IDEATE added 9 REQs (+128%) — the largest scope growth in project history. Are the 9 additions risk-reduction (guardrail, PII, mode-conflict, resilience, audit, tech-debt, cost, NFR measurability) or scope creep with a defensive veneer?
|
||||
|
||||
**Evidence:**
|
||||
- git log `b8c7de8` — "ideation results — 9 accepted into v0.5, 4 accepted into v0.6."
|
||||
- REQUIREMENTS.md:29-70 — 9 IDEATE REQs.
|
||||
- PLAN-v0.5:1011 — "16/16 REQ-IDs covered."
|
||||
|
||||
**Analysis:**
|
||||
The 9 IDEATE REQs map to named risks:
|
||||
- REQ-IDEATE-01 (guardrail tuning corpus) → R-ASSIST-06/07 (FP/FN).
|
||||
- REQ-IDEATE-02 (in-loop processor test) → REQ-IDEATE-02 interface gap (GuardrailContext.role).
|
||||
- REQ-IDEATE-03 (mode-conflict) → D-061 mutual exclusivity gap.
|
||||
- REQ-IDEATE-04 (measurable NFRs) → REQ-NFR-ASSIST-01/03 verifiability.
|
||||
- REQ-IDEATE-05 (PII policy) → R-ASSIST-08 (STRIDE information-disclosure).
|
||||
- REQ-IDEATE-06 (tech-debt) → 8 v0.4 P1+ findings.
|
||||
- REQ-IDEATE-07 (cost tracking) → C-3 budget.
|
||||
- REQ-IDEATE-08 (WebRTC reconnect) → R-ASSIST-09.
|
||||
- REQ-IDEATE-09 (incremental audit-log) → R-ASSIST-14 abrupt termination.
|
||||
|
||||
**Every addition maps to a named risk or a carried-forward finding.** None are features. The expansion is risk-reduction, not scope creep. The +128% is large but justified — v0.5 is the first *safety-critical* milestone, and the IDEATE stage surfaced the defensive requirements the practice surface (v0.1-v0.4) didn't need. The 4 deferred to v0.6 (REQ-IDEATE-10..13) are also risk-reduction (LLM-as-judge, assist-weaning, offline mode, voice-only context) — the ideation was disciplined.
|
||||
|
||||
**Verdict:** The IDEATE expansion is risk-reduction, not scope creep. Every REQ maps to a named risk. Accepted (G-046). Future ideation must maintain this discipline — the grill will flag any IDEATE addition that doesn't map to a named risk. **Confidence: 0.78.**
|
||||
|
||||
---
|
||||
|
||||
#### Probe 4 — In-loop guardrail processor (structural pipeline change): Is the "minimal delta" framing accurate?
|
||||
|
||||
**Question:** RESEARCH §5.2 frames the assist pipeline as "minimal delta: ~1 new pipeline builder, ~1 new guardrail processor." But the v0.1 pipeline has NO in-loop guardrail (the CS guardrail runs on the debrief). Is the in-loop processor a "minimal delta" or a structural change?
|
||||
|
||||
**Evidence:**
|
||||
- server/pipeline.py:143-185 — `build_pipeline()` has no in-loop guardrail processor (transport → stt → latency → user_agg → llm → latency → tts → latency → transport → assistant_agg).
|
||||
- RESEARCH-v0.5 §5.2 — "v0.5 adds an in-loop guardrail processor for assist mode. This is a pipeline-structure change but a small one (~1 new Pipecat frame processor)."
|
||||
- server/guardrails/customer_service.py — CS guardrail runs `check()` standalone, not as a frame processor.
|
||||
- PLAN-v0.5 TASK-05-02 — `LiveAssistGuardrailProcessor(FrameProcessor)` between llm and tts.
|
||||
- PLAN-v0.5 Open Question #4 (line 1046) — "verify Pipecat's `LLMContextAggregator` supports injecting a message + re-running the LLM within a single `process_frame` call. If not, the retry may need to be a separate pipeline task."
|
||||
|
||||
**Analysis:**
|
||||
The "minimal delta" framing is *partially accurate*. The service reuse (transport/stt/llm/tts) is genuinely minimal — the constructors are env-driven and reusable (verified: pipeline.py:63-109). **But the in-loop guardrail processor is a structural change**: the v0.1 pipeline has no post-LLM frame processor; v0.5 inserts one between `llm` and `tts`. This is novel for this codebase. The retry mechanism (inject `RETRY_INSTRUCTION` + re-run LLM mid-stream) is *unvalidated* against Pipecat's frame semantics — Open Question #4 defers this to EXECUTE, which is too late for a safety-critical path.
|
||||
|
||||
The risk: if Pipecat's `LLMFullResponseEndFrame` doesn't fire as expected, or if the `LLMContextAggregator` can't inject a retry mid-stream, the guardrail's "one retry" (D-068) becomes "canned fallback only" — safe but degraded. The coaching quality drops (every block → canned fallback, no second chance). This is a *quality* risk, not a *safety* risk (the canned fallback is safe) — but it affects the product's value.
|
||||
|
||||
**Verdict:** The in-loop guardrail processor is a structural change, not a minimal delta. The retry mechanism must be validated before Wave 3 (G-049 MUST). If Pipecat can't do mid-stream retry, document the fallback (canned-only) + update D-068's safety posture. The "minimal delta" framing should be corrected in the plan. **Confidence: 0.70.**
|
||||
|
||||
---
|
||||
|
||||
#### Probe 5 — Tap-to-talk UX (D-071): Is the unvalidated adoption risk acceptable for a pilot?
|
||||
|
||||
**Question:** D-071 ships tap-to-talk only (no wake-word). The learner taps a button on their phone during a real customer call. The phone may be in their pocket. The tap may be socially awkward. No user testing validates this UX. Is the pilot the validation, or is this a feature looking for a user?
|
||||
|
||||
**Evidence:**
|
||||
- D-071 (PROJECT.md:241) — "tap-to-talk ONLY (no wake-word in v0.5)… learner taps a button to invoke an assist turn during a real shift."
|
||||
- G-065 (Axis 7) — tap-to-talk UX assumption confidence 0.60 (lowest).
|
||||
- RESEARCH-v0.5 §4.1 — "No direct competitor does live-in-ear coaching during real customer calls on a $100 phone" (novel surface, no comparable UX to benchmark).
|
||||
|
||||
**Analysis:**
|
||||
Tap-to-talk is a *proven* pattern for walkie-talkie apps (Zello, Voxer) — users tap+hold to speak, release to send. This is a reasonable UX for hands-free-adjacent interaction. **But** those apps are *the* primary interface (the user opens the app to talk); Praxis assist is a *secondary* interface (the learner is in a real customer call, the phone is in their pocket, they tap a button on a screen they can't see). The social + ergonomic gap is real: the learner must (a) have earbuds in, (b) have the phone accessible, (c) tap a button without looking, (d) do this during a live customer interaction. This is a *high-friction* UX.
|
||||
|
||||
The pilot is the validation — v0.5 measures assist usage (REQ-NFR-ASSIST-04 cohort metrics). If adoption is low, v0.6 adds wake-word (the hands-free target). This is the correct pilot posture: ship the *value* (coaching/guardrail/context-binding), validate the *UX* (tap-to-talk adoption), iterate in v0.6. The risk is that low adoption makes the pilot a *failure* — but the pilot's purpose is to *find out*, not to *prove* adoption.
|
||||
|
||||
**Verdict:** Tap-to-talk is an unvalidated but reasonable UX for a pilot. The pilot is the validation. v0.6 adds wake-word if adoption is low. Accept with documented risk (G-073). **Confidence: 0.65.**
|
||||
|
||||
---
|
||||
|
||||
#### Probe 6 — 2-phase split: Is P1 (assist core + guardrail) independently shippable without P2 (measurement + tech-debt)?
|
||||
|
||||
**Question:** P1 ships v0.1.11 (assist core + guardrail, 12 REQs). P2 ships v0.1.12 (integration + tech-debt + NFR measurement, 4 REQs). Is P1 independently shippable — does a learner get a safe assist experience without P2?
|
||||
|
||||
**Evidence:**
|
||||
- PLAN-v0.5:17-23 — P1 = assist voice loop + guardrail (12 REQs, 24 tasks); P2 = integration + measurement + tech-debt (4 REQs, 9 tasks).
|
||||
- config.json:110 — `"per_phase": true` (per-phase ship).
|
||||
|
||||
**Analysis:**
|
||||
P1 delivers: the assist voice loop (build_assist_pipeline), the 3-layer guardrail (LiveAssistGuardrail + tuning corpus + adversarial test), the shift-bounded session model, the tap-to-talk client, the warm WebRTC + reconnect, the incremental audit-log, the mode-conflict guard, the PII policy. A learner can start a shift, tap-to-talk, get coaching with guardrails, end the shift. **This is a safe, usable assist experience.**
|
||||
|
||||
P2 adds: the cohort aggregation assist metrics (operator visibility), the cost tracking (C-3 check), the NFR measurement (p95 latency, FP/FN rates), the tech-debt wave (8 v0.4 P1+ findings). **P2 is hardening + visibility, not safety.** The guardrail's safety is in P1 (SLICE-03/04/08); P2 *measures* the guardrail's FP/FN rates (SLICE-09) but the guardrail itself ships in P1.
|
||||
|
||||
The one caveat: the aggregation cache tech-debt (P1+ #7) corrupts `assist_active_learners_count` during P1 (G-051). But P1 doesn't ship operator visibility (the cohort dashboard extension is P2 SLICE-10) — so the corrupted metric is not *visible* during P1. The fix lands in P2 before the dashboard extension. This is a *sequencing* dependency, not a P1 safety gap.
|
||||
|
||||
**Verdict:** P1 is independently shippable — a learner gets a safe assist experience. P2 is hardening + operator visibility + measurement. The split is clean (P1 = safety-critical voice loop, P2 = hardening). The aggregation cache corruption during P1 is not visible (no dashboard in P1) and fixed in P2 before visibility. **Confidence: 0.82.**
|
||||
|
||||
---
|
||||
|
||||
### v0.4 Grill Deferred Items — Coverage Check
|
||||
|
||||
The v0.4 grill (GRILL-v0.4.md) deferred no items to v0.5 (v0.4 was the operator tier, complete). The v0.4 grill's 8 P1+ findings are carried forward as REQ-IDEATE-06 (tech-debt wave, P2 SLICE-12). Let me verify:
|
||||
|
||||
| v0.4 Grill/Finding | v0.5 Coverage | Status |
|
||||
|---------------------|---------------|--------|
|
||||
| G-008 (backup drill) | v0.4 complete (REVIEW.md:240) | ✅ Resolved in v0.4 |
|
||||
| G-011 (two-store fallback) | v0.4 complete (REVIEW.md:241) | ✅ Resolved in v0.4 |
|
||||
| G-027 (first-boot no v0.3 key) | v0.4 complete (REVIEW.md:242) | ✅ Resolved in v0.4 |
|
||||
| G-031 (R-AUTH-01 reframe) | v0.4 complete (REVIEW.md:243) | ✅ Resolved in v0.4 |
|
||||
| G-038 (differencing-attack test) | v0.4 complete (REVIEW.md:244) | ✅ Resolved in v0.4 |
|
||||
| G-041 (SPA fallback subclass) | v0.4 complete (REVIEW.md:245) | ✅ Resolved in v0.4 |
|
||||
| P1+ #1 (argon2id blocking) | REQ-IDEATE-06, TASK-12-04 | ✅ Covered in v0.5 P2 |
|
||||
| P1+ #2 (rate-limit mock test) | REQ-IDEATE-06, TASK-12-04 | ✅ Covered in v0.5 P2 |
|
||||
| P1+ #3 (cookie-secret length) | REQ-IDEATE-06, TASK-12-02 | ✅ Covered in v0.5 P2 |
|
||||
| P1+ #4 (credential status enum) | REQ-IDEATE-06, TASK-12-03 | ✅ Covered in v0.5 P2 |
|
||||
| P1+ #5 (revocation audit log) | REQ-IDEATE-06, TASK-12-04 | ✅ Covered in v0.5 P2 |
|
||||
| P1+ #6 (nightly zoneinfo) | REQ-IDEATE-06, TASK-12-04 | ✅ Covered in v0.5 P2 |
|
||||
| P1+ #7 (aggregation cache) | REQ-IDEATE-06, TASK-12-01 | ✅ Covered in v0.5 P2 (critical path for assist metrics — G-051) |
|
||||
| P1+ #8 (f-string SQL) | REQ-IDEATE-06, TASK-12-03 | ✅ Covered in v0.5 P2 |
|
||||
|
||||
**Verdict:** 6/6 v0.4 grill MUSTs resolved in v0.4. 8/8 v0.4 P1+ findings covered in v0.5 P2 SLICE-12 (REQ-IDEATE-06). The aggregation cache fix (P1+ #7) is on the v0.5 critical path for correct assist metrics (G-051).
|
||||
|
||||
---
|
||||
|
||||
### Binding Decisions
|
||||
|
||||
| ID | Axis | Decision | Confidence | Type |
|
||||
|----|------|----------|-----------|------|
|
||||
| G-042 | 1 | Live Assist is the correct next priority (delivers the transfer surface). Novel per RESEARCH §4.1. | 0.80 | ACCEPT |
|
||||
| G-043 | 1 | CI is the named sponsor under full autonomy (G-002 carry-forward). | 0.80 | ACCEPT |
|
||||
| G-044 | 1 | v0.5 is not a zombie (delivers the transfer surface). Practice surface works without it. | 0.78 | ACCEPT |
|
||||
| G-045 | 1 | No financial ROI; ROI is product-completeness + safety-surface foundation. REQ-IDEATE-07 measures cost. | 0.68 | ACCEPT |
|
||||
| G-046 | 2 | IDEATE scope expanded +128% (7→16 REQs). Accepted — all 9 additions are risk-reduction, map to named risks. Future ideation must maintain discipline. | 0.78 | ACCEPT |
|
||||
| G-047 | 2 | NFRs are research-grounded, not frozen. REQ-NFR-ASSIST-01 at-risk (D-072 pilot tolerance). REQ-IDEATE-04 provides measurable freeze. | 0.75 | ACCEPT |
|
||||
| G-048 | 2 | Out-of-scope is explicit. D-071 (wake-word deferred) is the key scope reduction, binding. | 0.85 | ACCEPT |
|
||||
| **G-049** | **3** | **MUST: In-loop guardrail processor retry mechanism (TASK-05-02) must be validated against Pipecat frame semantics BEFORE Wave 3. Add a Wave-1/2 spike: verify LLMFullResponseEndFrame + LLMContextAggregator retry injection. If infeasible, document canned-fallback-only + update D-068. Binding contract, not open question.** | **0.70** | **MUST** |
|
||||
| G-050 | 3 | 3 integration points, all additive. Cohort aggregation (low) + mastery separation (low) + voice pipeline (medium, G-049). Aggregation cache tech-debt on critical path (G-051). | 0.75 | ACCEPT |
|
||||
| G-051 | 3 | 8 v0.4 P1+ findings inherited, budgeted in P2 SLICE-12. Aggregation cache fix corrupts assist metrics during P1 — accept (P1 ships voice loop, not operator dashboard). Document in P1 ship notes. | 0.72 | ACCEPT |
|
||||
| G-052 | 3 | Tech-debt budgeted (4 tasks in P2 SLICE-12, `should` priority). Proportional. | 0.80 | ACCEPT |
|
||||
| G-053 | 4 | Key-person: voice-engineer (new capability, largest territory), security-engineer (guardrail), backend-engineer (session API). Voice-engineer highest risk (first activation). | 0.78 | ACCEPT |
|
||||
| G-054 | 4 | 5 active personas, max 5 concurrent (at limit, no slack). Peak parallelism 2-3 slices. Voice-engineer capability claimed but undemonstrated — G-049 is the test. | 0.72 | ACCEPT |
|
||||
| G-055 | 4 | CI is product owner (full autonomy). For safety-critical surface, `escalate_high_severity: true` governs. R-ASSIST-07 must be escalated or documented as medium (Probe 1). | 0.68 | ACCEPT |
|
||||
| G-056 | 4 | Team building 3 new capabilities (in-loop processor, warm WebRTC, regex tuning). All on safety-critical/critical path. Acceptable for pilot with G-049 de-risking. | 0.72 | ACCEPT |
|
||||
| G-057 | 5 | 2-phase split evidence-based (P1 safety-critical voice loop, P2 hardening + measurement). P1 independently shippable. | 0.82 | ACCEPT |
|
||||
| G-058 | 5 | Critical-path: guardrail tuning corpus (FP/FN rates). TASK-04-02 is the gate. G-049 de-risks secondary path. | 0.75 | ACCEPT |
|
||||
| G-059 | 5 | 33 tasks evidence-based (smaller than v0.4's 52 due to D-071 + 0 new deps). Bottom-up sized. | 0.80 | ACCEPT |
|
||||
| G-060 | 5 | Definition of done = per-slice acceptance + per-phase ship + verify + REQ-IDEATE-04 measurable NFRs (p95 ≤650ms, FP<5%). | 0.82 | ACCEPT |
|
||||
| G-061 | 6 | Assist cost driver budgeted (SLICE-11, ~$0.20/month, well under C-3). Diagnostic, not enforced. | 0.80 | ACCEPT |
|
||||
| G-062 | 6 | C-3 relaxation (D-012) remains valid for v0.5 pilot. Assist adds ~$0.20/month. Post-pilot path preserves $3. | 0.78 | ACCEPT |
|
||||
| G-063 | 6 | Burn rate: ~0.8-1.0 days estimated (proportional to v0.4, -37% tasks). | 0.78 | ACCEPT |
|
||||
| G-064 | 6 | No budget contingency. D-071 removed Picovoice MAU-pricing dependency. 0 external commercial dependencies. | 0.85 | ACCEPT |
|
||||
| G-065 | 7 | 3 core assumptions: tap-to-talk UX (0.60, unvalidated), regex guardrail (0.65, residual risk), ≤650ms latency (0.65, unmeasured). All pilot-scale with v0.6 hardening. | 0.63 | ACCEPT |
|
||||
| G-066 | 7 | 4 dependencies: Picovoice (deferred ✅), PIPEDA (escalation ⚠️), cohort pipeline (additive ✅), voice pipeline (structural ⚠️ G-049). | 0.75 | ACCEPT |
|
||||
| **G-067** | **7** | **MUST: R-ASSIST-07 (guardrail false-negative) must have a documented acceptance threshold before EXECUTE. Adversarial FN rate (REQ-IDEATE-01) must be: (a) measured pre-ship (TASK-04-02), (b) compared against a threshold (e.g., "≤20% acceptable for pilot because defense-in-depth + audit + v0.6 LLM-as-judge mitigate; >20% triggers re-tuning or escalation"), (c) threshold + rationale documented in ship notes. Not a "0% FN" demand — a "know your residual risk + decide" demand. config.json:37 escalate_high_severity governs.** | **0.68** | **MUST** |
|
||||
| G-068 | 7 | Pre-mortem top-4: guardrail bypass (G-067 gap), PIPEDA (ESCALATION-01), in-loop retry (G-049), latency >650ms (D-072). All addressed. | 0.75 | ACCEPT |
|
||||
| G-069 | 8 | Escalation path: grill is the safety-critical mechanism (ESCALATION-01 + G-067). Post-ship ongoing escalation (safety spike → human) is v0.6+ gap. Accept for pilot. | 0.70 | ACCEPT |
|
||||
| G-070 | 8 | Grill is the right gate for safety-critical surface (ROADMAP:30 explicit). Human-in-the-loop checkpoint. | 0.82 | ACCEPT |
|
||||
| G-071 | 8 | No status-report omission. "LSP errors" claim unverified (files compile clean). Type-checking warnings non-blocking. | 0.80 | ACCEPT |
|
||||
| G-072 | 8 | No human stop trigger (full autonomy). Grill is the stop mechanism. ESCALATION-01 (PIPEDA) is the de facto stop trigger for the assist surface. | 0.78 | ACCEPT |
|
||||
| G-073 | 9 | Live Assist's first user is the learner during a real shift. Tap-to-talk (D-071) is unvalidated UX (0.60). v0.5 validates adoption; v0.6 adds wake-word if low. | 0.65 | ACCEPT |
|
||||
| G-074 | 9 | v0.5 needs no deploy changes (0 new deps, no CT bump, v0.4 LXC carries forward). devops-engineer deactivation sound. | 0.85 | ACCEPT |
|
||||
| G-075 | 9 | Rollback is preventive (disable assist → revert to v0.1.9). Assist surface is additive (clean revert). Corrective rollback impossible (live turn cannot be un-played) — audit log is post-incident tool. | 0.75 | ACCEPT |
|
||||
| G-076 | 9 | Success criteria research-grounded, measurement-validated in P2 (SLICE-09). P1 gate = TASK-04-02 (guardrail tuning test, FP<5%/FN<5%). P2 = SLICE-09 measurement. | 0.72 | ACCEPT |
|
||||
| G-077 | Meta | Auditor flags: R-ASSIST-07 threshold gap, PIPEDA escalation, IDEATE scope expansion, in-loop processor novelty. All addressed. | 0.78 | ACCEPT |
|
||||
| G-078 | Meta | v0.5 NOT doing: PIPEDA legal review (escalated), post-ship safety escalation (v0.6), human red-team prompt set (synthetic corpus accepted for pilot). | 0.75 | ACCEPT |
|
||||
| G-079 | Meta | Tap-to-talk-only (D-071) is the 80/20. Guardrail work (REQ-IDEATE-01/04) is the non-negotiable 20%. Cutting further ships an unvalidated safety-critical surface. | 0.80 | ACCEPT |
|
||||
| G-080 | Meta | 5 success conditions: guardrail robustness (P1 gate), in-loop processor (P1 spike), PIPEDA (human escalation), latency (P2 measurement), UX adoption (post-ship). Plan ready, proof in execution. | 0.72 | ACCEPT |
|
||||
|
||||
---
|
||||
|
||||
### Escalations
|
||||
|
||||
**ESCALATION-01 — PIPEDA consent-law review (D-073, R-ASSIST-08).** Confidence: 0.55 (below 0.60 threshold).
|
||||
|
||||
The ambient mic captures the real customer (a third party); ASR transcribes their speech; the turns table stores it (REQ-IDEATE-05). Canada's PIPEDA + provincial one-party/two-party consent laws govern recording. D-073 defers the legal review to "Phase 1 implementation." The disclosure (D-070) is shown to the *learner*, not the *customer* — it is the engineering mitigation, not a legal determination.
|
||||
|
||||
**The CI agent under full autonomy cannot resolve a legal question.** This must be escalated to human attention:
|
||||
|
||||
1. **Determine the consent regime:** Does Canada PIPEDA + the pilot province's consent law require one-party consent (learner's consent sufficient — D-070 covers) or two-party consent (customer must consent — Praxis cannot notify the customer)?
|
||||
2. **If one-party:** the disclosure (D-070) is sufficient. Proceed with v0.5.
|
||||
3. **If two-party:** the assist surface may need geographic restriction (one-party provinces only) or customer-facing consent (out of scope for v0.5 — would block the assist surface in two-party provinces).
|
||||
4. **If a PIPEDA privacy policy / data handling agreement is required:** the PII policy (REQ-IDEATE-05, 30-day retention + redaction) may need to be formalized into a PIPEDA-compliant policy before ship.
|
||||
|
||||
**Action required:** Human legal review of Canada PIPEDA + provincial consent law for ambient recording during coaching, before v0.5 SHIP. The grill cannot determine with confidence ≥0.60 whether the disclosure is sufficient. This is the de facto stop trigger for the assist surface (G-072).
|
||||
|
||||
---
|
||||
|
||||
### MUST Conditions Summary (blocking — must be resolved before Phase 1 EXECUTE)
|
||||
|
||||
1. **G-049 — In-loop guardrail processor retry validation.** Add a Wave-1/2 spike task: verify Pipecat's `LLMFullResponseEndFrame` fires after the full LLM response + that `LLMContextAggregator` supports injecting a retry message + re-running the LLM within `process_frame`. If infeasible, document the fallback (canned-fallback-only, no retry) + update D-068's safety posture. This is a binding contract, not an open question (PLAN Open Question #4 must be resolved pre-EXECUTE).
|
||||
|
||||
2. **G-067 — R-ASSIST-07 guardrail false-negative acceptance threshold.** The adversarial FN rate (REQ-IDEATE-01) must be: (a) measured pre-ship (TASK-04-02), (b) compared against a *documented threshold* (e.g., "≤20% acceptable for pilot because defense-in-depth + audit + v0.6 LLM-as-judge mitigate; >20% triggers a re-tuning wave or escalation"), (c) the threshold + mitigation rationale documented in the v0.5 ship notes. The plan's current "reported but not threshold-gated" (PLAN:419) is insufficient for a safety-critical surface. config.json:37 `escalate_high_severity: true` is the governing constraint.
|
||||
|
||||
---
|
||||
|
||||
### Escalations Requiring Human Attention (before SHIP)
|
||||
|
||||
**ESCALATION-01 — PIPEDA consent-law review.** Determine whether Canada PIPEDA + provincial consent law requires one-party or two-party consent for ambient recording during coaching. If the disclosure (D-070) is legally sufficient, proceed. If two-party consent is required, the assist surface may need geographic restriction or customer-facing consent (out of scope for v0.5). This is the de facto stop trigger for the assist surface.
|
||||
|
||||
---
|
||||
|
||||
### FIX Conditions (non-blocking — tracked in VERIFY-P1/P2)
|
||||
|
||||
- **G-046** — Document in v0.5 ship notes: IDEATE expanded scope +128% (7→16 REQs). All additions are risk-reduction. Future ideation must maintain risk-reduction discipline.
|
||||
- **G-051** — Document in P1 ship notes: assist metrics (assist_active_learners_count) are incorrect during P1 due to the aggregation cache tech-debt (v0.4 P1+ #7). Fix lands in P2 SLICE-12 before operator dashboard visibility.
|
||||
- **G-065** — Document in v0.5 ship notes: tap-to-talk UX (D-071) is the lowest-confidence assumption (0.60, unvalidated). v0.5 pilot validates adoption; v0.6 adds wake-word if low.
|
||||
- **G-069** — Document in v0.5 ship notes: post-ship safety signal escalation (nightly FN trend spike → human) is a v0.6+ governance gap. v0.5 ships the measurement (REQ-IDEATE-04); v0.6 adds the LLM-as-judge + the escalation response.
|
||||
- **G-073** — Document in v0.5 ship notes: v0.5 validates the coaching/guardrail/context-binding value, not the hands-free UX (tap-to-talk is the pilot validation; wake-word is v0.6).
|
||||
- **G-078** — Document in v0.5 ship notes: the guardrail tuning corpus (REQ-IDEATE-01) is synthetic (LLM-generated), not a human red-team prompt set. Accepted limitation for pilot.
|
||||
|
||||
---
|
||||
|
||||
### ACCEPT Items (proceed as-is)
|
||||
|
||||
- Live Assist is the correct next priority (G-042).
|
||||
- CI is the named sponsor under full autonomy (G-043).
|
||||
- v0.5 is not a zombie (G-044).
|
||||
- IDEATE scope expansion is risk-reduction, not scope creep (G-046, Probe 3).
|
||||
- Out-of-scope is explicit; D-071 wake-word deferral is the key scope reduction (G-048).
|
||||
- 3 integration points are additive (G-050).
|
||||
- Tech-debt is budgeted in P2 SLICE-12 (G-052).
|
||||
- Key-person dependency is manageable under parallelization (G-053).
|
||||
- 2-phase split is evidence-based; P1 independently shippable (G-057, Probe 6).
|
||||
- 33 tasks is evidence-based (G-059).
|
||||
- Assist cost is budgeted, well under C-3 (G-061, G-062).
|
||||
- No budget contingency — D-071 removed Picovoice dependency (G-064).
|
||||
- No deploy changes needed (G-074).
|
||||
- Rollback is preventive (disable assist → revert to v0.1.9) (G-075).
|
||||
- Tap-to-talk is the 80/20; guardrail work is the non-negotiable 20% (G-079).
|
||||
- v0.4 grill MUSTs (6/6) resolved in v0.4; v0.4 P1+ findings (8/8) covered in v0.5 P2.
|
||||
|
||||
---
|
||||
|
||||
### Bottom Line
|
||||
|
||||
The v0.5 plan is **not unfeasible** — the D-071 tap-to-talk deferral stripped the client-architecture risk, the battery risk, the Picovoice commercial risk, and 5 of 14 research risks. The remaining scope (guardrail + context-binding + shift-bounded session + in-loop processor) is the *core* safety surface, well-researched and cleanly phased. The plan is **not over-scoped** after the deferral (16 REQs, but 9 are defensive; 33 tasks vs v0.4's 52). The plan is **not a zombie** (Live Assist is the v0.1-promised surface, now delivered).
|
||||
|
||||
The 2 MUST conditions are surgical:
|
||||
- 1 is a *validation spike* (in-loop guardrail processor retry mechanism — G-049).
|
||||
- 1 is a *threshold* (R-ASSIST-07 adversarial FN rate acceptance — G-067).
|
||||
|
||||
The 1 escalation is a *legal question* the CI cannot resolve (PIPEDA consent-law review — ESCALATION-01). This is the de facto stop trigger for the assist surface.
|
||||
|
||||
**Resolve the 2 MUSTs, answer the 1 escalation, and v0.5 is a GO.**
|
||||
|
||||
The v0.5 milestone is the project's first **safety-critical** surface — the AI is in a learner's ear during *real* customer interactions. The grill's binding decisions (G-067 threshold, G-049 validation) + the escalation (ESCALATION-01 PIPEDA) are the safety-critical gates. The plan's architecture (3-layer guardrail, defense-in-depth, audit + nightly trending) is sound — the grill's conditions ensure the *residual risk* is *known + decided*, not *assumed + deferred*.
|
||||
+150
-1
@@ -541,4 +541,153 @@ Two personas are **phase-specific** for v0.4:
|
||||
- devops-engineer (create-operator.py) ↔ security-engineer (argon2id hashing) — D-052
|
||||
- The **security-engineer and devops-engineer are NOT in config.json `personas`** — emergent personas defined in PERSONAS.md (same pattern as v0.2/v0.3). Territory enforcement (warn mode) picks up globs from PERSONAS.md.
|
||||
- R-AUTH-01 (Secure cookie + no-TLS) is a security-engineer + lead-developer collaboration point for GRILL-v0.4 (config-driven flag resolution must be grill-approved).
|
||||
- R-VC-MIG-01 (VC key migration) is a security-engineer + data-engineer collaboration point (archive v0.3 public key before activating new key).
|
||||
- R-VC-MIG-01 (VC key migration) is a security-engineer + data-engineer collaboration point (archive v0.3 public key before activating new key).
|
||||
|
||||
---
|
||||
|
||||
# Praxis — Persona Assessment (v0.5 Live Assist)
|
||||
|
||||
> **Generated:** v0.5 RESEARCH stage
|
||||
> **Project:** Praxis (v0.5 — Live Assist: on-the-job voice companion, wake-word, guardrails, cohort aggregation extension)
|
||||
> **Source:** v0.5 RESEARCH-v0.5-live-assist.md + v0.5 REQUIREMENTS.md (REQ-ASSIST-01/02/03, REQ-NFR-ASSIST-01..04) + actual `server/` structure + `db/` structure
|
||||
|
||||
## v0.5 Persona Roster
|
||||
|
||||
### Active personas (5)
|
||||
|
||||
The v0.5 milestone is **voice-pipeline-heavy (wake-word + assist mode + latency tuning) + safety-critical guardrails + cohort aggregation extension**. The **voice-engineer reactivates** (proposed at line 458 for v0.5+ — now confirmed). The **devops-engineer deactivates** (no deploy changes — v0.4 LXC carries forward). The **frontend-engineer deactivates provisionally** (assist UI is minimal — ~100-150 LOC, below the reactivation threshold; reactivate if the assist control surface exceeds ~200 LOC). The security-engineer and data-engineer are retained (guardrails + aggregation).
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: lead-developer
|
||||
active: true
|
||||
phase_specific: false
|
||||
reason: Coordinates across assist pipeline (voice-engineer), guardrails (security-engineer), context-binding + session API (backend-engineer), and aggregation extension (data-engineer). Owns the build_assist_pipeline() design decision (mode param vs separate builder) and the warm-WebRTC-connection lifecycle (D-067). Owns the C-8 latency tension for assist (R-ASSIST-02 — the binding-constraint risk). Required for every milestone.
|
||||
domain: coordination
|
||||
frameworks: [pipecat, fastapi, sqlite, postgres, webrtc, docker]
|
||||
constraints: [pragmatic, latency-budget-aware, hybrid-storage-no-cross-db-joins, k-anonymity-floor-10, assist-does-not-affect-mastery, warm-webrtc-per-shift]
|
||||
territory:
|
||||
- "docker-compose.yml"
|
||||
- ".env.example"
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: voice-engineer
|
||||
active: true
|
||||
phase_specific: true
|
||||
reason: REACTIVATED for v0.5 (proposed at PERSONAS.md line 458 for v0.5+). Owns the wake-word client (Picovoice Porcupine Android foreground service — D-058, D-064), the assist audio pipeline (warm WebRTC connection per shift — D-067, wake-word → first-audio latency — R-ASSIST-03), latency tuning (the C-8 <600ms assist budget — R-ASSIST-02, Domain 3), the in-loop guardrail processor (post-LLM frame processor — D-060 layer 2), and the build_assist_pipeline() (reuses v0.1 services, swaps the system prompt + adds the guardrail processor). This is the largest new territory in v0.5: the assist voice loop is a new mode alongside the practice scenario loop. Will deactivate in v0.6 unless voice work continues (accent modeling, multi-voice personas, multi-learner concurrency).
|
||||
domain: voice
|
||||
frameworks: [porcupine-android, webrtc, silero-vad, pipecat, audio-codecs, piper-tts, cartesia-tts, deepgram-nova3, ollama-cloud]
|
||||
constraints: [sub-600ms-latency-assist, warm-webrtc-per-shift, foreground-service-background-mic, wake-word-detection-latency, piper-tts-for-assist, lean-assist-system-prompt-150-tokens, in-loop-guardrail-processor]
|
||||
territory:
|
||||
- "**/server/pipeline.py"
|
||||
- "**/server/assist/pipeline.py"
|
||||
- "**/server/asr/**"
|
||||
- "**/server/tts/**"
|
||||
- "**/server/latency.py"
|
||||
- "**/server/guardrails/live_assist.py"
|
||||
- "**/client/wake-word/**"
|
||||
- "**/client/assist-service/**"
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: backend-engineer
|
||||
active: true
|
||||
phase_specific: false
|
||||
reason: RETAINED from v0.4. Owns the assist context-binding endpoints (load path week + scenario tag + learner theta from SQLite into the assist prompt — D-059), the assist session API (POST /api/assist/shift/start + /end — D-062, D-069), the SessionRecorder extension (session_type field, assist turn logging, _build_session_outcome assist branch), and the cohort hook extension for session_type='assist' (D-062). Collaborates with security-engineer on the LiveAssistGuardrail ruleset (backend owns the in-loop processor integration; security owns the regex patterns + safety logic). The assist session API + context-binding is the largest backend territory in v0.5.
|
||||
domain: backend
|
||||
frameworks: [pipecat, pydantic, fastapi, uvicorn, aiosqlite, asyncpg]
|
||||
constraints: [api-first, type-safe, mastery-off-voice-path, aggregation-off-voice-path, latency-budget-aware, no-cross-db-joins, assist-does-not-update-mastery, schedule-mastery-false-for-assist]
|
||||
territory:
|
||||
- "**/server/**"
|
||||
- "**/server/assist/**"
|
||||
- "**/server/guardrails/**"
|
||||
- "**/server/cohort/**"
|
||||
- "**/server/session_recorder.py"
|
||||
- "**/db/migrations/**"
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: data-engineer
|
||||
active: true
|
||||
phase_specific: false
|
||||
reason: RETAINED from v0.4. Owns the assist aggregation integration into the v0.4 cohort pipeline (new assist metrics in cohort_aggregates — no schema change, new metric strings: assist_shifts_count, assist_turns_count, assist_avg_turns_per_shift, assist_active_learners_count, assist_guardrail_block_rate — D-062), the turns-table guardrail_verdict field migration (SQLite, additive — D-060 layer 3), and the assist session_type field in the sessions table. Also owns the k-anonymity suppression extension for assist metrics (assist_active_learners_count distinct-count, ≥10 threshold). Smaller v0.5 surface than v0.4 but on the critical path for operator visibility into assist usage + guardrail safety signals.
|
||||
domain: data
|
||||
frameworks: [sqlite, postgres16, aiosqlite, asyncpg]
|
||||
constraints: [schema-first, migration-driven, no-cross-db-joins, k-anonymity-floor-10, opaque-learner-ref, write-time-suppression, assist-metrics-no-schema-change]
|
||||
territory:
|
||||
- "**/db/**"
|
||||
- "**/db/migrations/**"
|
||||
- "**/server/cohort/aggregator.py"
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: security-engineer
|
||||
active: true
|
||||
phase_specific: true
|
||||
reason: RETAINED from v0.4. Owns the LiveAssistGuardrail enforcement (REQ-ASSIST-03 — the most safety-critical requirement in v0.5: the AI is in the learner's ear during real customer interactions). The 3-layer guardrail (D-060, REFINED by D-068) is the security-engineer's v0.5 surface: (1) prompt rules (coaching-mode system prompt — ask guiding questions, never give the answer, never claim false authority), (2) output filter patterns (direct-answer vs coaching-question regex — DIRECT_SCRIPT_RE, IMPERATIVE_RE, FALSE_AUTHORITY_RE, IMPERSONATION_RE, COACHING_QUESTION_RE + one retry + canned fallback), (3) audit logging (turns table guardrail_verdict + cohort aggregation guardrail_block_rate safety signal for operators). Also owns the privacy/consent disclosure surface (R-ASSIST-08 — foreground-service notification + learner-facing "Assist is on — those around you may be recorded by your mic" disclosure — D-070). REQ-ASSIST-03 blocks ship if the guardrail is not robust.
|
||||
domain: security
|
||||
frameworks: [pynacl, canonicaljson, base58, regex, llm-guardrail-patterns, argon2-cffi, starlette-sessionmiddleware]
|
||||
constraints: [coaches-not-does, no-direct-answer-patterns, no-false-authority, no-impersonation, audit-all-assist-turns, guardrail-block-rate-operator-visible, consent-disclosure-required, output-filter-false-negative-mitigation-defense-in-depth]
|
||||
territory:
|
||||
- "**/server/guardrails/live_assist.py"
|
||||
- "**/server/guardrails/**"
|
||||
- "**/server/vc/**" # retained from v0.4 (no v0.5 change expected)
|
||||
- "**/server/auth/**" # retained from v0.4 (no v0.5 change expected)
|
||||
---
|
||||
```
|
||||
|
||||
### Deactivated personas (2)
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: devops-engineer
|
||||
active: false
|
||||
phase_specific: true
|
||||
reason: DEACTIVATED for v0.5. No deploy changes — v0.4's LXC + Docker-in-LXC + Postgres + backup cron carries forward unchanged. The assist foreground service is a client-side concern (voice-engineer territory), not a deploy/infra change. No new Docker services, no CT resource bump, no new backup scripts, no new deploy scripts. Will reactivate in v0.6+ if deploy hardening (TLS, multi-instance for assist concurrency, autoscaling) or a CT bump is needed.
|
||||
domain: devops
|
||||
frameworks: [proxmox-lxc, docker, systemd, bash, pg_dump, cron]
|
||||
constraints: [idempotent-deploy, rollback-on-failure, secrets-never-committed]
|
||||
territory: []
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: frontend-engineer
|
||||
active: false
|
||||
phase_specific: true
|
||||
reason: DEACTIVATED for v0.5 (PROVISIONAL). v0.5 assist mode is invoked by wake-word (audio) — the UI surface is minimal: a "Start Shift" / "End Shift" toggle + a context-declaration screen (path week + scenario tag selector). Estimated ~100-150 LOC of React — below the reactivation threshold (~200 LOC). This is small enough that the voice-engineer (client/wake-word + client/assist-service) can own the minimal control surface alongside the audio pipeline, OR the backend-engineer can add a minimal React route. No full frontend surface (no new dashboard, no complex components, no chart library). Will reactivate in v0.6+ if a richer assist control surface (shift history, guardrail-block review, assist coaching-quality dashboard) is needed. NOTE FOR ORCHESTRATOR: if the assist control surface (start/stop shift + context declaration + shift history) is judged non-trivial (>200 LOC of React), reactivate frontend-engineer. Current estimate: ~100-150 LOC.
|
||||
domain: frontend
|
||||
frameworks: [react, react-router-dom, pipecat-client-sdk, webrtc, vite]
|
||||
constraints: [component-first, voice-first-ui, minimal-client-javascript, assist-control-surface-minimal]
|
||||
territory: []
|
||||
---
|
||||
```
|
||||
|
||||
## v0.5 Notes for PLAN/EXECUTE
|
||||
|
||||
- Territory enforcement mode: `warn` (per config.json `personas.territory_enforcement`)
|
||||
- The **voice-engineer owns the largest v0.5 task surface**: wake-word client (Porcupine Android foreground service), assist pipeline (build_assist_pipeline + in-loop guardrail processor), warm WebRTC lifecycle, latency tuning (the C-8 <600ms assist budget is the binding-constraint risk — R-ASSIST-02), and the minimal assist control surface. This is the first voice-engineer activation (proposed since v0.2 PERSONAS line 458).
|
||||
- The **security-engineer's v0.5 surface is the most safety-critical**: REQ-ASSIST-03 (coaches not does, never lies to real customers). The 3-layer guardrail (D-060, D-068) blocks ship if not robust. R-ASSIST-07 (output filter false negatives) is the residual risk — mitigated by defense-in-depth (prompt + regex + audit) + a post-v0.5 LLM-as-judge.
|
||||
- The **backend-engineer's v0.5 surface**: assist session API + context-binding + SessionRecorder extension + cohort hook extension. Solid mid-size surface.
|
||||
- The **data-engineer's v0.5 surface is the smallest** but on the operator-visibility critical path: assist metrics (no schema change, new metric strings) + guardrail_verdict migration.
|
||||
- Cross-persona collaboration points:
|
||||
- voice-engineer (in-loop guardrail processor) ↔ security-engineer (LiveAssistGuardrail regex + safety logic) — D-060/D-068
|
||||
- voice-engineer (assist pipeline) ↔ backend-engineer (assist session API + context-binding) — D-059/D-061
|
||||
- backend-engineer (session_outcome session_type) ↔ data-engineer (aggregator _aggregate_assist branch) — D-062
|
||||
- security-engineer (guardrail_verdict audit) ↔ data-engineer (guardrail_block_rate cohort metric) — D-060 layer 3 + D-062
|
||||
- lead-developer (C-8 latency tension) ↔ voice-engineer (latency tuning) — R-ASSIST-02
|
||||
- The **voice-engineer is NOT in config.json `personas`** — emergent persona defined in PERSONAS.md (same pattern as v0.2 devops-engineer, v0.3/v0.4 security-engineer). Territory enforcement (warn mode) picks up globs from PERSONAS.md.
|
||||
- R-ASSIST-01 (Picovoice MAU pricing) is a lead-developer + voice-engineer collaboration point (decide: built-in wake word for v0.5, custom post-pilot, or Vosk fallback).
|
||||
- R-ASSIST-02 (C-8 <600ms at risk) is a lead-developer + voice-engineer collaboration point for GRILL-v0.5 (relax C-8 for assist or push hardening to v0.6).
|
||||
- R-ASSIST-08 (privacy/consent) is a security-engineer + lead-developer collaboration point (legal review of Canada consent law for ambient recording — flag for orchestrator).
|
||||
- **Client architecture flag (RESEARCH §7 Q1):** v0.5 may require a client upgrade from React-Web (v0.1, D-015) to React-Native or a separate native Android assist app, because background wake-word needs an Android foreground service (which React-Web can't provide). Alternative: defer wake-word to v0.6 and ship v0.5 assist as tap-to-talk only. **This is a scope decision for the orchestrator.**
|
||||
File diff suppressed because it is too large
Load Diff
+66
-4
@@ -1,9 +1,9 @@
|
||||
# Praxis — Voice-first AI Apprenticeship Platform
|
||||
|
||||
**Milestone:** v0.4 (Operator tier — cohort dashboard, auth, Postgres)
|
||||
**Status:** phase 3 — final review (active milestone); P0-P2 complete (v0.1.6/v0.1.7/v0.1.8 tagged)
|
||||
**Milestone:** v0.5 (Live Assist — on-the-job voice companion)
|
||||
**Status:** phase 0 — pre-execution (active milestone)
|
||||
**Autonomy:** full
|
||||
**Previous milestone:** v0.3 (Mastery scoring + competency rubrics + verifiable credentials) — complete, tagged v0.1.5, release #380
|
||||
**Previous milestone:** v0.4 (Operator tier — cohort dashboard, auth, Postgres) — complete, tagged v0.1.9, release created, merged to main
|
||||
|
||||
## Vision
|
||||
|
||||
@@ -42,7 +42,47 @@ v0.3 activated the mastery/assessment layer deferred from v0.1/v0.2 (per D-021,
|
||||
- Voice loop (Deepgram Nova-3 + Cartesia + Pipecat + Ollama Cloud)
|
||||
- v0.1 scenario (`cs_refund_ca_v01.yaml`) + guardrails + debrief
|
||||
|
||||
## v0.4 Scope (Operator Tier — Cohort Dashboard + Auth + Postgres)
|
||||
## v0.5 Scope (Live Assist — On-the-Job Voice Companion)
|
||||
|
||||
v0.5 activates the Live Assist surface deferred from v0.1 (per the original out-of-scope list: "Live Assist mode"). v0.1–v0.4 built and validated the practice surface — learners practice scenarios with AI tutors, scored against rubrics, progress via mastery gates, with a v0.4 operator tier observing cohort patterns. v0.5 adds the **companion surface**: a hands-free voice assistant a learner invokes *while actually working* on the job, context-aware of their current scenario/skill path, coaching in real time without doing the job for them.
|
||||
|
||||
**v0.5 in scope (activated REQ groups — 3 REQs + NFRs TBD after RESEARCH/IDEATE):**
|
||||
- **Hands-free voice companion (REQ-ASSIST-01):** voice companion invocable while working — distinct from the practice voice loop (v0.1). Hands-free (earbuds/phone-in-pocket), always-listening or wake-word/hotkey-activated, short coaching turns interleaved with real work. Reuses the v0.1 voice pipeline (Pipecat + Deepgram + Cartesia + Ollama Cloud) but in a new "assist" mode, not the practice scenario loop.
|
||||
- **Context-aware (REQ-ASSIST-02):** knows the learner's current scenario/skill path — binds to the learner's active path week (D-037) + scenario context, so coaching is relevant to the job they're actually doing, not generic. Carries forward learner state from SQLite (D-007 preserved).
|
||||
- **Guardrails (REQ-ASSIST-03):** coaches, does not do the job; never lies to real customers — the safety-critical distinction from the practice surface. The AI is in the learner's ear during real customer interactions; it must never impersonate, never give answers the learner parrots, never claim authority it doesn't have. Extends D-019 guardrail layer with Live-Assist-specific ruleset. Safety-sensitive: real customers, real consequences.
|
||||
|
||||
**v0.5 out of scope (still deferred):**
|
||||
- REQ-PATH-01 (full multi-path launch) — still Customer Service path only; Live Assist binds to that path
|
||||
- REQ-LOWBW-01..03 (WhatsApp/USSD/offline) — v0.5 is voice; low-bandwidth surfaces later
|
||||
- REQ-VOICE-05/06 (multi-language, persona switching) — Canadian English only in v0.5
|
||||
- REQ-DASH-02 (full operator-suite dashboard) — v0.4's foundational cohort view is sufficient; Live Assist telemetry feeds the same aggregation pipeline
|
||||
- Learner auth / multi-learner-per-device — still single-learner-per-device (D-007)
|
||||
- Live Assist session recording/replay — v0.5 is live coaching, not recording; replay later
|
||||
- Proactive intervention (AI speaks unprompted) — v0.5 is learner-invoked; proactive later
|
||||
- Multi-modal (camera/screen context) — audio-only (C-4)
|
||||
|
||||
**Carries forward from v0.4 (already in production):**
|
||||
- Operator-tier Postgres + cohort aggregation + operator auth + cohort dashboard (v0.4)
|
||||
- Mastery scoring + competency rubrics + IRT + VC issuer (v0.3)
|
||||
- Scenario library + Customer Service 6-week path (v0.3)
|
||||
- Docker-in-LXC deployment (v0.2)
|
||||
- Voice loop: Deepgram Nova-3 + Cartesia + Pipecat + Ollama Cloud (v0.1)
|
||||
|
||||
**Open questions for CLARIFY/RESEARCH:**
|
||||
1. ✅ **RESOLVED (D-058, D-064):** Invocation model = wake-word (Picovoice Porcupine on-device) + tap-to-talk fallback. Refined: built-in wake word for v0.5 pilot (MAU pricing has no recurring free tier — R-ASSIST-01); custom "Hey Praxis" post-pilot; Vosk fallback. **NEW open: client architecture — React-Web (v0.1) can't do background wake-word; React-Native upgrade or defer wake-word to v0.6 (RESEARCH §7 Q1).**
|
||||
2. ✅ **RESOLVED (D-059):** Context-binding = learner declares context at session start (path week + scenario tag); server reads `progress.current_week` from SQLite. Auto-detection impossible (C-4).
|
||||
3. ✅ **RESOLVED (D-060, D-068):** "Coaches not does" enforced via 3-layer guardrail: (1) prompt rules (coaching-mode system prompt), (2) output filter (regex direct-answer + false-authority + impersonation patterns + one retry + canned fallback), (3) audit log (turns table guardrail_verdict + cohort guardrail_block_rate). **NEW open: privacy/consent for ambient recording (R-ASSIST-08) — legal review of Canada PIPEDA.**
|
||||
4. ⚠️ **AT RISK (D-061, R-ASSIST-02):** <600ms latency budget for assist turns estimated ~655-770ms (all-cloud) / ~655ms (Piper + lean prompt). Mitigations: D-065 (Piper TTS for assist), D-066 (≤150-token prompt). **Flag for orchestrator: relax C-8 for assist or push hardening to v0.6.** The same pipeline handles both modes (no second Pipecat instance) — confirmed. Wake-word → first-audio is a separate ~850-1150ms budget (warm WebRTC — D-067).
|
||||
5. ✅ **RESOLVED (D-058, D-064, D-067):** Hands-free UX = Porcupine on-device (offline, ~1MB RAM, <4% core — verified). Battery ~4-9% per 8h shift (estimated — R-ASSIST-14, needs Phase-1 measurement). Warm WebRTC per shift (D-067). Foreground service for background mic (Android 14+ requirement).
|
||||
6. ✅ **RESOLVED (D-062, D-069):** Session model = shift-bounded ("starting shift" / "ending shift"), with assist turns within. Auto-end after 8h (D-069). Aggregates as `session_type=assist` in v0.4 cohort pipeline (no schema change). Does NOT update mastery (D-063).
|
||||
|
||||
**NEW open questions from research (for orchestrator + PLAN):**
|
||||
7. **Client architecture for v0.5** (RESEARCH §7 Q1): React-Web (v0.1, D-015) can't run a background foreground service on Android. Options: (a) upgrade to React Native, (b) separate native Android assist app, (c) defer wake-word to v0.6 and ship v0.5 assist as tap-to-talk only. **Recommendation: (c) for v0.5 pilot.** Scope decision.
|
||||
8. **Picovoice sales engagement timing** (R-ASSIST-01): before PLAN or after v0.5 ships with tap-to-talk? If wake-word deferred to v0.6, sales engagement is v0.6.
|
||||
9. **Output filter regex corpus** (R-ASSIST-06): how to build the tuning corpus before v0.5 ships? Synthetic corpus via LLM (prompt gemma4:cloud to produce coaching + direct-answer responses, label, tune). Phase-1 task.
|
||||
10. **Canada consent law review** (R-ASSIST-08, D-070): PIPEDA + provincial one-party/two-party consent for ambient recording during coaching. Legal review recommended before v0.5 ship.
|
||||
|
||||
## v0.4 Scope (Operator Tier — Cohort Dashboard + Auth + Postgres — complete)
|
||||
|
||||
v0.4 activates the operator tier deferred from v0.3 per GRILL-v0.3.md Axis 2 (the operator tier was originally v0.8 on this ROADMAP; pulling it into v0.3 created a 2-milestone program disguised as one). The v0.3 mastery/VC/scenario work carries forward unchanged; v0.4 layers the operator surface on top of it.
|
||||
|
||||
@@ -185,6 +225,22 @@ v0.3 activated the mastery/assessment layer deferred from v0.1/v0.2 (per D-021).
|
||||
| D-055 | Postgres backup = **nightly `pg_dump` to a named Docker volume, 7-day retention** | CLARIFY auto-decide. Postgres data lives on a named Docker volume (`pgdata`) inside the LXC CT. Nightly cron job runs `pg_dump praxis | gzip > /backups/praxis-$(date).sql.gz` to a second named volume (`pgbackups`). 7-day retention (rotates oldest). Operator can `pct pull` backups to the PVE host. No streaming replication (single CT, no replica target). This is pilot-tier backup; a later milestone adds off-CT replication. | 0.70 | No backups (data loss risk), WAL streaming to a replica (no replica in v0.4), S3 push (no S3 in LXC pilot) |
|
||||
| D-056 | Auth session store = **signed stateless cookies (HMAC-SHA256), no server-side session table** | CLARIFY auto-decide. D-041 said "session-cookie" — clarifying: the cookie is a self-contained signed token (user_id, issued_at, expiry, HMAC). No `sessions` table in Postgres. Verification = recompute HMAC + check expiry. Logout = client clears cookie (stateless — no server revocation list in v0.4). Rate limit is in-memory (single-instance). This minimizes DB load + simplifies the auth surface. A later milestone adds a revocation list if multi-instance or forced-logout is needed. | 0.75 | Postgres sessions table (DB load + cleanup job), Redis sessions (extra service), JWT with claims (same idea, more complex tooling) |
|
||||
| D-057 | Auth enforcement = **server-side on every `/api/operator/*` request + React route guard for UX, never trust the client** | CLARIFY auto-decide. FastAPI middleware checks the signed cookie on every `/api/operator/*` request; 401 if missing/invalid/expired. React `/operator/*` routes check a `/api/operator/me` call on mount and redirect to `/operator/login` if 401 — this is UX only, the server is the authority. The cohort dashboard reads only k-anonymized aggregates (D-034) so even an auth bypass leaks no PII (defense in depth). VC issuance endpoints (`/api/operator/credentials/*`) are also auth-gated. | 0.85 | Server-only (poor UX — no redirect), React-only (insecure — bypassable), no auth on issuance (credential forgery risk) |
|
||||
| D-058 | Live Assist invocation model = **wake-word (Picovoice Porcupine on-device) + tap-to-talk fallback, NOT always-listening** | CLARIFY auto-decide (full autonomy). Always-listening drains battery on a $100 Android phone the learner is actively using for work + raises privacy concerns (listening to real customers). Wake-word is the hands-free UX without always-on microphone. Picovoice Porcupine is on-device, offline, low-power, free-tier supports custom wake words. Tap-to-talk fallback covers wake-word failure or noisy environments. Research phase to validate Porcupine on Android + battery impact. | 0.65 | Always-listening (battery + privacy), pure tap-to-talk (not hands-free), cloud wake-word (latency + connectivity dependency) |
|
||||
| D-059 | Live Assist context-binding source = **learner declares context at session start (path + scenario tag), server reads active path week from SQLite for rubric/coaching alignment** | CLARIFY auto-decide (full autonomy). Live Assist cannot auto-detect which real scenario the learner is in (no camera per C-4, no screen context). Learner taps their current path week / scenario tag when starting an assist session (or voice-declares it). Server reads the learner's `progress.current_week` from SQLite (D-007) for rubric alignment + coaching context. This keeps the learner in control + makes context explicit. Auto-detection from calendar/location is out of scope. | 0.70 | Full auto-detection (impossible without sensors), pure SQLite read without learner declaration (ambiguous which real scenario), no context (generic coaching — violates REQ-ASSIST-02) |
|
||||
| D-060 | Live Assist "coaches not does" guardrail enforcement = **(1) prompt-layer rules (system prompt forbids giving direct answers), (2) output filter (post-generation check for direct-answer patterns), (3) session audit log of all assist turns** | CLARIFY auto-decide (full autonomy). REQ-ASSIST-03 is safety-critical. Three layers: (1) system prompt explicitly instructs the LLM to ask guiding questions, never give the answer, never speak on behalf of the learner. (2) Output filter scans the LLM response for direct-answer patterns (e.g., "you should say X to the customer") and rewrites/blocks. (3) All assist turns logged to SQLite for audit + the operator cohort dashboard (v0.4). Research phase to validate filter patterns + false-positive rate. | 0.70 | Prompt-only (single layer — bypassable), output-filter-only (inconsistent with prompt), no logging (no audit trail — unsafe for safety-critical surface) |
|
||||
| D-061 | Live Assist latency budget = **shares the v0.1 voice pipeline (Pipecat + Deepgram + Cartesia + Ollama) but assist turns are short (≤30s), and the <600ms round-trip (C-8) must hold for assist turns** | CLARIFY auto-decide (full autonomy). Live Assist does NOT run concurrently with a practice session — it's a separate mode. The learner invokes assist, gets short coaching turns (≤30s each), dismisses. The same pipeline handles both modes (no second Pipecat instance). C-8's <600ms budget applies to assist turns too — coaching that arrives after the customer moment has passed is useless. Research phase to validate wake-word → first-audio latency + whether assist context adds LLM tokens that break the budget. | 0.75 | Separate pipeline (doubles infra cost + complexity), relaxed latency for assist (useless coaching), longer turns (loses the real-time moment) |
|
||||
| D-062 | Live Assist session model = **shift-bounded sessions (learner starts "I'm starting my shift", ends "ending shift"), with individual coaching turns within the shift; assist turns feed the v0.4 cohort aggregation as a new `session_type=assist`** | CLARIFY auto-decide (full autonomy). A shift-bounded session matches the real-world use case (a learner works a shift, invokes assist as needed). Within the shift, each assist turn is a discrete coaching exchange. Assist turns aggregate into the v0.4 cohort pipeline (D-045) as `session_type=assist` — operators see assist usage patterns alongside practice patterns. No double-counting with mastery: assist turns are coaching, not assessment, so they don't update θ (D-035) or count toward mastery gates (D-032). Continuous (no start/end) is ambiguous for aggregation. | 0.70 | Continuous (no aggregation boundary), per-turn sessions (too granular for cohort view), no aggregation (operators blind to assist usage) |
|
||||
| D-063 | Live Assist does NOT update mastery score (D-035) or count toward mastery gates (D-032) — assist is coaching, not assessment | CLARIFY auto-decide (full autonomy). Mastery gates require demonstrated performance across varied scenarios (D-032). Live Assist is the AI helping during real work — it's coaching, not a performance demonstration. Counting assist turns toward mastery would be gaming (the AI did the work). Assist turns are logged for audit + cohort aggregation (D-062) but never update θ or open gates. A later milestone may add "assist-weaning" (track reducing assist reliance as a mastery signal) but v0.5 keeps them separate. | 0.85 | Assist counts toward mastery (gaming risk), assist updates θ (contaminates the ability estimate), no logging (no audit) |
|
||||
| D-064 | Live Assist wake-word engine = **Picovoice Porcupine (built-in wake word for v0.5 pilot; custom "Hey Praxis" post-pilot)**, with **Vosk as the documented open-source fallback** | RESEARCH-derived (RESEARCH-v0.5 §1.2). R-ASSIST-01: Porcupine MAU pricing has no recurring free tier (verified via Picovoice general FAQ). v0.5 ships with a built-in Porcupine wake word (e.g., "Bumblebee") to avoid custom-training costs during the pilot. Post-pilot, engage Picovoice sales for a custom "Hey Praxis" under a pilot/educational tier. Vosk (Apache 2.0, offline) is the fallback if Porcupine pricing is unsustainable. Snowboy rejected (deprecated). | 0.70 | Vosk for v0.5 (free but heavier), TFLite DIY (engineering effort), Snowboy (deprecated) |
|
||||
| D-065 | Live Assist TTS = **Piper (self-hosted on pilot server) as the default for assist turns**, Cartesia as the quality fallback for practice mode | RESEARCH-derived (RESEARCH-v0.5 §3.3). R-ASSIST-02: assist turns are latency-critical (C-8). Piper ~80ms first audio vs Cartesia ~120ms. The v0.1 R4 mitigation pre-stages Piper; v0.5 assist mode defaults to Piper to claw back ~40ms toward the <600ms budget. Practice mode retains Cartesia (quality over latency for practice). | 0.75 | Cartesia for both (simpler, but +40ms on assist), Piper for both (lower quality for practice) |
|
||||
| D-066 | Live Assist system prompt = **≤150 input tokens** (coaching instruction ~80 tokens + context-binding ~50 tokens + voice-conciseness ~20 tokens) | RESEARCH-derived (RESEARCH-v0.5 §3.3). R-ASSIST-02: extra input tokens add prefill latency (~0.5ms/token). A lean prompt keeps the prefill delta under 50ms vs v0.1 practice. Avoid dumping the full rubric or scenario YAML into the prompt — context-binding is terse (path week, scenario tag, one-line coaching focus). | 0.78 | Verbose prompt (easier coaching quality, but +100-200ms latency) |
|
||||
| D-067 | Live Assist WebRTC connection = **warm for the entire shift** (foreground service keepalive; not per-turn cold connect) | RESEARCH-derived (RESEARCH-v0.5 §3.4). R-ASSIST-03: cold WebRTC connect (~500-1000ms) is unacceptable for live assist. The assist foreground service opens a warm connection at shift start, keeps it alive (heartbeat every 30s), and reuses it for every assist turn. Closed at shift-end. Between turns, only keepalive flows (no audio streaming) to save battery. | 0.78 | Per-turn cold connect (too slow), always-streaming (battery + privacy) |
|
||||
| D-068 | Live Assist guardrail output filter = **regex-based direct-answer + false-authority + impersonation patterns, with one retry on block + canned coaching redirect fallback** | RESEARCH-derived (RESEARCH-v0.5 §2.3). R-ASSIST-06/07: regex is the fast on-voice-path filter (matches the existing CustomerServiceGuardrail pattern). One retry gives the LLM a chance to self-correct; the canned fallback ensures a safe response if the retry also blocks. LLM-as-judge deferred to post-v0.5 (off-voice-path, more accurate, nightly). | 0.78 | LLM-as-judge on-voice-path (too slow for <600ms), no filter (unsafe) |
|
||||
| D-069 | Live Assist shift = **auto-end after 8 hours** (configurable via `PRAXIS_ASSIST_MAX_SHIFT_HOURS=8`) | RESEARCH-derived (RESEARCH-v0.5 §4.2). R-ASSIST-11: learners may forget "ending shift", leaving orphaned WebRTC connections + stale sessions. Auto-end after 8h (a typical shift length) closes the shift cleanly, fires the aggregation hook, and releases the foreground service. The learner can restart a new shift if needed. | 0.75 | No auto-end (orphan risk), shorter (4h — too short for some shifts), longer (12h — battery risk) |
|
||||
| D-070 | Live Assist consent disclosure = **foreground-service notification + learner-facing "Assist is on — those around you may be recorded by your mic" disclosure at shift start** | RESEARCH-derived (RESEARCH-v0.5 §2.6). R-ASSIST-08: the ambient mic may pick up the real customer. Ethical and legal (one-party/two-party consent law) requires disclosure. The foreground service notification (Android requirement) + an in-app disclosure at shift start covers the learner's awareness. The customer's consent is the learner's responsibility (Praxis can't notify the customer). **Flag for orchestrator: legal review of Canada consent law (PIPEDA) for ambient recording during coaching.** | 0.65 | No disclosure (legal/ethical risk), explicit customer consent prompt (impractical — the customer isn't a Praxis user) |
|
||||
| D-071 | Live Assist client architecture for v0.5 = **tap-to-talk ONLY (no wake-word in v0.5)** — React-Web (D-015) keeps the assist surface as a tap-to-talk web control; wake-word deferred to v0.6 with a React-Native or native Android app | RESEARCH-flagged decision (full autonomy). R-ASSIST-13: React-Web (v0.1, D-015) cannot run an Android background foreground service for on-device wake-word detection. Adding wake-word requires a React-Native upgrade or a separate native Android assist app — a client-architecture change too large for v0.5's scope. v0.5 ships assist as tap-to-talk (the existing fallback from D-058): learner taps a button to invoke an assist turn during a real shift. This preserves the "hands-free goal" as the v0.6 target while delivering the coaching/guardrail/context-binding value in v0.5 on the existing web client. D-058's wake-word is deferred, not abandoned. | 0.70 | Force React-Native in v0.5 (scope creep — client rewrite + assist feature together), defer all of v0.5 assist to v0.6 (no value delivered), ship wake-word on web (technically infeasible) |
|
||||
| D-072 | Live Assist C-8 latency budget for v0.5 pilot = **target <600ms (C-8) retained; accept ≤650ms as pilot tolerance with hardening in v0.6** — Piper TTS (D-065) + ≤150-token prompt (D-066) are the mitigations; if measurement shows >650ms, document as R-ASSIST-02 carried to v0.6 | RESEARCH-flagged decision (full autonomy). R-ASSIST-02: research estimates ~655-770ms all-cloud, ~655ms with Piper + lean prompt. C-8 is a binding constraint but v0.5 is a pilot — a 50ms tolerance (≤650ms) is acceptable if trending down, with <600ms as the v0.6 hardening target. The alternative (relax C-8 formally) weakens the constraint for all future milestones; the alternative (block v0.5 ship until <600ms) delays the safety-critical guardrail work. Accept pilot tolerance, measure in Phase 1, harden in v0.6. | 0.65 | Relax C-8 to 700ms (weakens constraint permanently), block v0.5 until <600ms (delays guardrail work), ignore the gap (unsafe) |
|
||||
| D-073 | Live Assist PIPEDA consent-law review = **defer to v0.5 Phase 1 implementation; document as R-ASSIST-08 in the grill** — the ambient-mic legal question is a grill-axis candidate, not a Phase 0 blocker | RESEARCH-flagged decision (full autonomy). R-ASSIST-08: Canada PIPEDA + provincial consent law for ambient recording during coaching needs legal review. This is not a Phase 0 research blocker — the disclosure (D-070) is the engineering mitigation. Legal review runs in parallel with Phase 1 implementation. The grill (next stage) should include an axis on consent/privacy. If the grill returns a MUST for legal review before ship, schedule it before Phase 1 SHIP. | 0.60 | Block Phase 0 on legal review (over-cautious — no implementation yet), ignore the legal risk (unsafe), no disclosure (D-070 already addresses) |
|
||||
|
||||
### Confidence updates from research
|
||||
|
||||
@@ -192,6 +248,12 @@ v0.3 activated the mastery/assessment layer deferred from v0.1/v0.2 (per D-021).
|
||||
|----|--------|-------|--------|
|
||||
| D-003 | 0.75 | **0.95** | Both Ollama model IDs verified in catalog as real, current, cloud-hosted tags |
|
||||
| D-007 | 0.80 | **0.90** | SQLite confirmed appropriate for v0.1 single-learner scale; no evidence favors alternatives |
|
||||
| D-058 | 0.65 | **0.70 (REFINED)** | Porcupine verified (on-device, offline, low-power, Android SDK, custom WW). MAU pricing / no recurring free tier contradicts the free-tier assumption — refined by D-064 (built-in WW for pilot, custom post-pilot, Vosk fallback). |
|
||||
| D-059 | 0.70 | **0.82** | `PraxisStore.get_progress()` confirmed returns `current_week`; auto-detection impossible (C-4); learner declaration is the right model. |
|
||||
| D-060 | 0.70 | **0.85** | 3-layer pattern confirmed as industry-standard; existing CustomerServiceGuardrail proves the regex output-filter approach. Refined by D-068 (regex + retry + canned fallback). |
|
||||
| D-061 | 0.75 | **0.70 (AT RISK)** | Estimated assist latency ~655-770ms (all-cloud) / ~655ms (Piper + lean prompt) — C-8 <600ms is at risk. Mitigations identified (D-065 Piper, D-066 lean prompt) but may not fully close the gap. Flag for orchestrator. |
|
||||
| D-062 | 0.70 | **0.85** | Shift-bounded model confirmed as matching real CS work; no schema change to cohort_aggregates (new metric strings); on-session-end hook extended cleanly. |
|
||||
| D-063 | 0.85 | **0.90** | `SessionRecorder.end(schedule_mastery=False)` for assist shifts confirmed — the mastery flow is practice-only by the existing flag. |
|
||||
|
||||
## Target Users (v0.3: Canada pilot — Customer Service path)
|
||||
|
||||
|
||||
+92
-10
@@ -1,11 +1,97 @@
|
||||
# Praxis — Requirements
|
||||
|
||||
**Milestone:** v0.4 (Operator tier — cohort dashboard, auth, Postgres)
|
||||
**Status:** phase 3 — final review (active milestone); P0-P2 complete — 8/8 v0.4 REQ covered (v0.1.6/v0.1.7/v0.1.8 tagged); v0.3 complete — released as v0.1.5 (13/13 v0.3 REQ covered)
|
||||
**Milestone:** v0.5 (Live Assist — on-the-job voice companion) — active, phase 0
|
||||
**Status:** phase 0 pre-execution — v0.4 complete (released as v0.1.9, merged to main, 8/8 v0.4 REQ covered); v0.3 complete (released as v0.1.5, 13/13 v0.3 REQ covered)
|
||||
|
||||
Formal requirements with REQ-IDs. Scoped to the active milestone unless noted. v0.1/v0.2/v0.3 requirements (complete) are retained for reference with their final status. Later-milestone requirements are marked `deferred`.
|
||||
Formal requirements with REQ-IDs. Scoped to the active milestone unless noted. v0.1/v0.2/v0.3/v0.4 requirements (complete) are retained for reference with their final status. Later-milestone requirements are marked `deferred`.
|
||||
|
||||
## v0.4 Active Requirements
|
||||
## v0.5 Active Requirements
|
||||
|
||||
### Live Assist (v0.5 core)
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-ASSIST-01 | Hands-free voice companion invocable while working — distinct from the practice voice loop (v0.1). Always-listening or wake-word/hotkey-activated, short coaching turns interleaved with real work. Reuses the v0.1 voice pipeline (Pipecat + Deepgram + Cartesia + Ollama Cloud) in a new "assist" mode. | must | P1 | active |
|
||||
| REQ-ASSIST-02 | Context-aware — knows the learner's current scenario/skill path. Binds to the learner's active path week (D-037) + scenario context so coaching is relevant to the job they're doing. Carries forward learner state from SQLite (D-007 preserved). | must | P1 | active |
|
||||
| REQ-ASSIST-03 | Guardrails: coaches, does not do the job; never lies to real customers. Safety-critical: the AI is in the learner's ear during real customer interactions. Extends D-019 guardrail layer with Live-Assist-specific ruleset. Never impersonates, never gives parrot-able answers, never claims false authority. | must | P1 | active |
|
||||
|
||||
## v0.5 Non-Functional Requirements
|
||||
|
||||
| REQ-ID | Requirement | Target | Phase | Status |
|
||||
|--------|-------------|--------|-------|--------|
|
||||
| REQ-NFR-ASSIST-01 | Live Assist voice round-trip latency | **< 600ms target (C-8); estimated ~655ms (Piper + lean prompt — D-065, D-066). AT RISK — accept ~650ms for pilot if trending down; <600ms hardening in v0.6.** Wake-word → first-audio is a separate ~850-1150ms budget (warm WebRTC — D-067). Must not degrade the practice pipeline (assist is a separate mode, not concurrent — D-061). | P1 | research-grounded (R-ASSIST-02) |
|
||||
| REQ-NFR-ASSIST-02 | Hands-free invocation on $100 Android | **Picovoice Porcupine on-device (offline, ~1MB RAM, <4% core — verified). Battery ~4-9% per 8h shift (estimated, needs Phase-1 measurement — R-ASSIST-14). Foreground service of type `microphone` (Android 14+). Built-in wake word for v0.5 pilot (D-064 — MAU pricing has no recurring free tier, R-ASSIST-01); custom "Hey Praxis" post-pilot; Vosk fallback. Tap-to-talk fallback for battery-saving / wake-word failure / noisy environments.** | P1 | research-grounded (R-ASSIST-01/04/05/13/14) |
|
||||
| REQ-NFR-ASSIST-03 | Live Assist guardrail enforcement | **3-layer guardrail (D-060, D-068): (1) coaching-mode system prompt (ask guiding questions, never give the answer, never claim false authority, never impersonate); (2) regex output filter (DIRECT_SCRIPT_RE + IMPERATIVE_RE + FALSE_AUTHORITY_RE + IMPERSONATION_RE; COACHING_QUESTION_RE allowed) with one retry on block + canned coaching fallback; (3) audit log (turns table guardrail_verdict JSON + cohort guardrail_block_rate safety signal for operators). Consent disclosure: foreground-service notification + learner-facing "Assist is on — those around you may be recorded" at shift start (D-070). Output filter false-negative residual risk mitigated by defense-in-depth + post-v0.5 LLM-as-judge.** | P1 | research-grounded (R-ASSIST-06/07/08) |
|
||||
| REQ-NFR-ASSIST-04 | Live Assist session model | **Shift-bounded (learner starts/ends a shift; assist turns within — D-062). Auto-end after 8h via `PRAXIS_ASSIST_MAX_SHIFT_HOURS=8` (D-069). Aggregates as `session_type=assist` in v0.4 cohort pipeline (no schema change — new metric strings: assist_shifts_count, assist_turns_count, assist_avg_turns_per_shift, assist_active_learners_count, assist_guardrail_block_rate). Does NOT update mastery (D-063 — `schedule_mastery=False` for assist shifts). k-anonymity ≥ 10 applies to assist metrics (D-034 carry-forward).** | P1 | research-grounded |
|
||||
|
||||
_NFRs refined from `pending-research` to `research-grounded` after the v0.5 RESEARCH stage (see RESEARCH-v0.5-live-assist.md). Targets are research-derived; Phase-1 measurement may further refine R-ASSIST-02 (latency) and R-ASSIST-14 (battery)._
|
||||
|
||||
## v0.5 Ideation-Derived Requirements (IDEATE-01..09, accepted)
|
||||
|
||||
_Generated by the IDEATE stage (3-tier analysis: mechanical git-mining + backend-enriched + chaos engineering). 9 of 13 ideas accepted into v0.5; 4 deferred to v0.6 (see v0.6 Backlog below)._
|
||||
|
||||
### Guardrail Quality & Safety (IDEATE-01, 02, 09)
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-IDEATE-01 | Guardrail output-filter tuning corpus + adversarial bypass test (pre-ship). Build a synthetic corpus (LLM-generate coaching vs direct-answer responses, label, tune the regex patterns DIRECT_SCRIPT_RE/IMPERATIVE_RE/FALSE_AUTHORITY_RE/IMPERSONATION_RE). Add an adversarial-bypass test with paraphrased direct answers designed to slip past the regex. Proactively mitigates R-ASSIST-06/07 (false-positive + false-negative risks) before the guardrail ships blind on its two most safety-critical metrics. Relates to the v0.1 latent safety-trap lesson (misspelled `_DEBRIFF_LEGAL_REDIRECT` — the rewrite/fallback path was never exercised by tests). | must | P1 | active |
|
||||
| REQ-IDEATE-02 | In-loop guardrail processor pipeline test + GuardrailContext.role 'assist' extension. (1) Add a pipeline-integration test that inserts the LiveAssistGuardrail as a post-LLM Pipecat frame processor between llm and tts (the existing test_guardrail.py only tests `check()` standalone). (2) Extend the `GuardrailContext.role` Literal to include `'assist'` (currently `system|user|assistant|debrief` — the LiveAssistGuardrail hits an interface gap). Both are structural coverage holes Phase 1 will hit immediately. | must | P1 | active |
|
||||
| REQ-IDEATE-09 | Audit-log completeness on abrupt shift end. Log the assist turn incrementally — persist the ASR transcript + LLM response + guardrail verdict before/at TTS start, not after playback completes — so abrupt termination (battery death R-ASSIST-14, power loss mid-turn) still leaves an audit trail. For a safety-critical surface (REQ-ASSIST-03), an incomplete audit log undermines the guardrail_block_rate safety signal and the operator's ability to investigate incidents. | must | P1 | active |
|
||||
|
||||
### Chaos & Resilience (IDEATE-03, 08)
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-IDEATE-03 | Mode-conflict enforcement: assist vs practice mutual exclusivity. Add a server-side guard (reject shift-start if a practice session is active, or vice versa) + a chaos test invoking assist during an active practice session. D-061 states assist is a separate mode (not concurrent), but nothing currently enforces mutual exclusivity — the server-side assist API and the practice /pipecat/webrtc endpoint are independent with no shared state guarding against a second connection. | must | P1 | active |
|
||||
| REQ-IDEATE-08 | WebRTC mid-shift drop + reconnect logic. Specify the reconnect state machine (does the foreground service auto-reconnect? what does the learner experience during the gap? does the in-flight assist turn retry or fail?) + add a chaos test (kill the WebRTC connection mid-shift, verify reconnect + turn recovery). R-ASSIST-09 names the risk; D-067 mandates warm WebRTC with 30s heartbeat but the reconnect logic is unspecified. | must | P1 | active |
|
||||
|
||||
### Security & Privacy (IDEATE-05)
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-IDEATE-05 | Customer-speech PII handling in the assist turns audit log (STRIDE information-disclosure). The ambient mic (R-ASSIST-08) captures BOTH the learner and the real customer; ASR transcribes both; the turns table stores transcribed text. The customer is a third party — their transcribed speech is third-party PII in SQLite. v0.5 needs an explicit policy: (a) strip customer turns from the audit log, (b) store only the learner's utterances, or (c) document that the audit log contains customer speech + apply consent-disclosure (D-070) + retention limits. Intersects with the R-ASSIST-08 legal review (D-073). | must | P1 | active |
|
||||
|
||||
### Spec Refinement (IDEATE-04)
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-IDEATE-04 | Measurable NFR targets for REQ-NFR-ASSIST-01 and REQ-NFR-ASSIST-03. (1) Latency: specify 'p95 assist-turn latency ≤ 650ms in Phase-1 measurement (pilot tolerance per D-072); <600ms hardening deferred to v0.6' — resolves the ambiguity in REQ-NFR-ASSIST-01's current text. (2) Guardrail: specify 'false-positive rate < 5% on the tuning corpus (REQ-IDEATE-01); false-negative rate measured + trended nightly' — makes REQ-NFR-ASSIST-03 verifiable. | must | P1 | active |
|
||||
|
||||
### Process / Tech Debt (IDEATE-06)
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-IDEATE-06 | Carry-forward the 8 v0.4 P1+ findings into the v0.5 backlog as a 'tech-debt wave'. Especially: (1) aggregation in-memory cache lost on restart (REVIEW.md P1+ #7 — directly corrupts v0.5 assist_active_learners_count after a server restart); (2) cookie-secret length validation (P1+ #3); (3) set_credential_status enum/f-string SQL (P1+ #4/#8). High-value, low-effort — folding into the v0.5 PLAN as a dedicated wave. | should | P1 | active |
|
||||
|
||||
### Cost (IDEATE-07)
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-IDEATE-07 | Assist per-turn cost tracking + C-3 budget impact verification. Extend server/cost.py to log per-assist-turn cost (each assist turn is a separate gemma4:cloud invocation). Add a Phase-1 budget check: estimate monthly assist cost per learner (e.g., 20 turns/shift × 20 shifts/month = 400 extra LLM calls) and flag if it pushes the total over the C-3 ≤ $3/active learner/month target. Extends REQ-NFR-COST-01 (v0.1 cost logging) to the new assist surface. | should | P1 | active |
|
||||
|
||||
## v0.6 Backlog (IDEATE-10..13, accepted for v0.6)
|
||||
|
||||
_4 ideas accepted for the v0.6 milestone (low-bandwidth surfaces). Recorded here for the v0.6 run; not active in v0.5._
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-IDEATE-10 | LLM-as-judge guardrail evaluation (nightly, off-voice-path) — measure the true false-negative rate the regex filter cannot. A nightly deepseek-v4-flash:cloud job sampling assist turns, classifying 'coached' vs 'did the job', feeding a 'guardrail adherence score' to the cohort dashboard. Natural v0.6 follow-on to v0.5's regex layer (D-068). | later | v0.6 P1 | deferred |
|
||||
| REQ-IDEATE-11 | Assist-weaning metric — track reducing assist reliance over shifts as a mastery signal. A 'turns-per-shift trend per learner' metric (k-anonymized) giving operators a leading indicator of skill transfer from practice to the real job. Bridges v0.5 assist + v0.3 mastery without violating D-063 (descriptive metric, not a gate input). | later | v0.6 P1 | deferred |
|
||||
| REQ-IDEATE-12 | Offline assist degraded mode — what happens when the backend is unreachable mid-shift? A canned local coaching redirect played from the client ('I can't reach the coaching server — take a moment and think about what the customer needs most right now') preserves the product's trust contract. Relevant to the v0.6 low-bandwidth/offline milestone (REQ-LOWBW-03). | later | v0.6 P1 | deferred |
|
||||
| REQ-IDEATE-13 | Voice-only context declaration (hands-free context binding, no tap). A voice-only path ('Hey Praxis, starting my shift, week 3, damaged-product refund') parsed by ASR into the context fields. Faithful to product principle #1 (voice-first); depends on an ASR-parsing spike. | later | v0.6 P1 | deferred |
|
||||
|
||||
## v0.5 Out of Scope (still deferred)
|
||||
|
||||
- REQ-PATH-01 (full multi-path launch) — still Customer Service path only; Live Assist binds to that path
|
||||
- REQ-LOWBW-01..03 (WhatsApp/USSD/offline) — v0.5 is voice; low-bandwidth surfaces later
|
||||
- REQ-VOICE-05/06 (multi-language, persona switching) — Canadian English only in v0.5
|
||||
- REQ-DASH-02 (full operator-suite dashboard) — v0.4's foundational cohort view is sufficient
|
||||
- Learner auth / multi-learner-per-device — still single-learner-per-device (D-007)
|
||||
- Live Assist session recording/replay — v0.5 is live coaching, not recording
|
||||
- Proactive intervention (AI speaks unprompted) — v0.5 is learner-invoked
|
||||
- Multi-modal (camera/screen context) — audio-only (C-4)
|
||||
|
||||
## v0.4 Active Requirements (complete — released as v0.1.9, retained for reference)
|
||||
|
||||
### Operator-Tier Postgres (v0.4 foundation)
|
||||
|
||||
@@ -155,13 +241,9 @@ Formal requirements with REQ-IDs. Scoped to the active milestone unless noted. v
|
||||
| REQ-PATH-01 | Launch paths: Customer Service, Retail Sales, Hospitality Front Desk, Home Health Aide, Basic English for Work, Auto-Rickshaw/Taxi | later | deferred | deferred |
|
||||
| REQ-PATH-02 | Path structured as a job (6-week example structure per PRD §6.4) | later | deferred | deferred |
|
||||
|
||||
### Live Assist
|
||||
### Live Assist (active in v0.5 — see v0.5 Active Requirements above)
|
||||
|
||||
| REQ-ID | Requirement | Priority | Phase | Status |
|
||||
|--------|-------------|----------|-------|--------|
|
||||
| REQ-ASSIST-01 | Hands-free voice companion invocable while working | later | deferred | deferred |
|
||||
| REQ-ASSIST-02 | Context-aware (knows current scenario/skill) | later | deferred | deferred |
|
||||
| REQ-ASSIST-03 | Guardrails: coaches, does not do the job; never lies to real customers | later | deferred | deferred |
|
||||
_REQ-ASSIST-01/02/03 activated in v0.5. See "v0.5 Active Requirements" section at the top of this file._
|
||||
|
||||
### Low-Bandwidth Surfaces
|
||||
|
||||
|
||||
@@ -0,0 +1,761 @@
|
||||
# Praxis — Research Findings (v0.5 Live Assist — On-the-Job Voice Companion)
|
||||
|
||||
> **Phase:** v0.5 research (Live Assist)
|
||||
> **Branch:** `phase/00-pre-execution`
|
||||
> **Status:** research complete — pending orchestrator review
|
||||
> **Date:** 2026-08-04
|
||||
> **Method:** Codebase inspection (`server/pipeline.py`, `server/guardrails/`, `server/session_recorder.py`, `server/cohort/aggregator.py`, `server/services/base.py`, `server/__main__.py`, `db/migrations/`, `db/pg_migrations/`), prior research (`.ciagent/RESEARCH.md` v0.1/v0.2/v0.3, `.ciagent/RESEARCH-v0.4-operator-tier.md`), D-058..D-063 CLARIFY decisions. Web-verified: Picovoice Porcupine FAQ + general FAQ + Android quickstart (fetched 2026-08-04), Vosk toolkit (alphacephei.com), RealWear (realwear.com). Domain-knowledge claims (LLM guardrail patterns, on-the-job coaching AI products) carry explicit confidence scores.
|
||||
|
||||
This document grounds the v0.5 Live Assist architecture in ecosystem evidence. It covers all 6 research questions, validates the CLARIFY decisions D-058..D-063 against real-world evidence, and concludes with a consolidated risks table, an NFR refinement, and a persona-roster decision.
|
||||
|
||||
---
|
||||
|
||||
## Summary of Findings (Executive 1-Pager)
|
||||
|
||||
1. **Picovoice Porcupine is the right wake-word engine, but the free-tier assumption in D-058 needs refinement.** (0.78) Porcupine is on-device, offline, low-power (~1 MB RAM, <4% of one core on RPi 3 — verified via Porcupine FAQ), accent-robust (universal, not voice-personalized), supports custom wake words trained via Picovoice Console, and ships an Android SDK (verified — quick-start page exists). **However**, the Picovoice general FAQ (fetched 2026-08-04) states: Porcupine is priced on **monthly active users (MAU)**, there is a **one-time Free Trial** (not a recurring free tier), and "Picovoice is a B2B company focused on on-device AI tools for enterprises. At this time, there are no dedicated free or paid plans for personal or non-commercial use." This **refines D-058**: the "free-tier supports custom wake words" framing is too optimistic for a recurring pilot — Praxis needs to either (a) negotiate an educational/pilot tier with Picovoice sales, (b) budget for MAU-based pricing in the pilot, or (c) ship a built-in Picovoice wake word (no custom training, falls under the trial) for v0.5 and add custom training later. **Flag for orchestrator: D-058 free-tier assumption is partially contradicted.**
|
||||
|
||||
2. **3-layer guardrail (D-060) is the correct pattern and matches industry practice.** (0.85) Prompt-layer rules + output-filter patterns + audit logging is the standard defense-in-depth for LLM safety. The existing `CustomerServiceGuardrail` (server/guardrails/customer_service.py, verified) already implements pattern-based output filtering (regex for legal/financial/medical advice + impersonation). v0.5 extends this with Live-Assist-specific patterns: detect "you should say X" / "tell the customer Y" / "the answer is Z" (direct-answer patterns) vs "what do you think the customer needs?" / "how could you acknowledge their frustration?" (coaching-question patterns). The output filter is a regex + keyword classifier on the LLM response before TTS; on hit, the response is either rewritten to a coaching redirect or blocked + re-prompted. Audit log = the existing `turns` table (SQLite) extended with a `guardrail_verdict` field; assist turns also flow to the v0.4 cohort aggregation as `session_type=assist` for operator visibility.
|
||||
|
||||
3. **<600ms latency budget (C-8, D-061) holds for assist turns IF context-binding stays off the voice path.** (0.80) The v0.1 budget breakdown (ARCHITECTURE.md): WebRTC ~50ms + Deepgram ~250ms + LLM ~200ms + Cartesia ~120ms + downlink ~50ms = ~670ms (marginally over). Adding context-binding tokens (path week, scenario tag, learner state) to the LLM system prompt adds **prompt-processing latency, not network latency** — ~50-200 extra input tokens on `gemma4:cloud` (256K context, so no context-window risk). At ~50ms per 100 input tokens of prefill latency, 200 extra tokens ≈ +100ms to first-token. **This pushes the all-cloud path to ~770ms — breaks C-8.** The mitigation: (a) keep context-binding tokens minimal (≤100 tokens: path week, scenario id, one-line coaching focus — not the full rubric), and (b) use the **Piper-on-pilot-server TTS path** (R4 mitigation from v0.1, ~80ms TTS instead of ~120ms Cartesia) which the architecture already pre-stages. With Piper: ~50 + 250 + 200 + 80 + 50 + ~50 (prefill for ~100 context tokens) = **~680ms** — still marginal. **Recommendation: assist turns use a leaner system prompt than practice turns (assist = coaching questions only, no role-play character persona), targeting ≤150 input tokens total system prompt.** This keeps prefill under 75ms and the total under 600ms with Piper. **Confidence 0.70** — prefill latency for gemma4:cloud is not yet measured (R3 from v0.1); Phase 1 must measure.
|
||||
|
||||
4. **Shift-bounded session model (D-062) matches real on-the-job coaching patterns.** (0.80) Real on-the-job coaching AI products bound sessions by work shifts or discrete interactions, not continuous always-on streams. Dialpad Ai Coach and Gong (industry knowledge, 0.65 confidence — vendor pages returned 404 on direct fetch; claims based on widely-documented product behavior) analyze call recordings post-hoc, not live-in-ear. RealWear (verified realwear.com) is hands-free AR glasses for frontline workers — visual + voice, industrial, hardware-first; not a phone-in-pocket voice companion. **No direct competitor does "live-in-ear coaching during real customer calls on a $100 Android phone."** This is Praxis's novel surface. The shift-bounded model ("I'm starting my shift" / "ending shift") gives a clean aggregation boundary + matches how retail/hospitality workers actually work (shifts are the unit of labor). Within a shift, each assist turn is a discrete coaching exchange (≤30s). Assist turns aggregate as `session_type=assist` alongside `session_type=practice` in the v0.4 cohort pipeline.
|
||||
|
||||
5. **v0.1 voice pipeline reuse is minimal-delta.** (0.85) The pipeline (`server/pipeline.py`) is parameterized by `scenario_id` and builds a `ScenarioRuntime` with a system prompt + opening line. v0.5 adds an "assist mode" alongside the practice scenario loop: the same `build_pipeline()` is called with a new `mode="assist"` parameter (or a distinct `build_assist_pipeline()`) that swaps the system prompt (coaching persona, not role-play character), drops the opening line (assist is invoked mid-shift, no scripted opener), and injects context-binding (path week, scenario tag). The Deepgram/Cartesia/Piper/Ollama services are reused unchanged — no new voice-service deps. The `SessionRecorder` (verified — 390 lines) is extended with an `assist` session type; the `_build_session_outcome()` method (line 164) already builds the dict the cohort aggregator consumes — v0.5 adds a `session_type` field. **Minimal delta: ~1 new pipeline builder, ~1 new guardrail ruleset, ~1 new session-type field, ~1 new aggregation metric.**
|
||||
|
||||
6. **Cohort aggregation integration (D-062) is a clean extension of the v0.4 pipeline.** (0.85) The `aggregator.py` (verified — 230 lines) upserts cells keyed by `(path, metric, window_start)`. v0.5 adds assist-specific metrics: `assist_turns_count`, `assist_active_learners_count`, `assist_avg_turns_per_shift`, `assist_guardrail_block_rate` (how often the output filter fired — a safety signal for operators). These are new `metric` strings in the same `cohort_aggregates` table — no schema change. The on-session-end hook (`server/cohort/hook.py`) is extended to accept `session_type=assist` outcomes; assist shifts fire the hook on shift-end (not per-turn — per-turn is too granular and would double-count). k-anonymity ≥ 10 applies identically. **Operators see assist usage patterns alongside practice patterns in the same dashboard views** (D-053's 3 views extend naturally: practice volume becomes practice+assist volume, failure patterns gain an "assist guardrail blocks" breakdown).
|
||||
|
||||
7. **Picovoice Porcupine vs alternatives: Porcupine wins on Android integration + custom wake-word training; Vosk is the open-source fallback.** (0.80) Vosk (verified alphacephei.com) is an offline ASR toolkit (20+ languages, runs on Android, 50MB models, pip-installable) — it's a full ASR, not a dedicated wake-word engine, but can do keyword spotting with a constrained vocabulary. Vosk is free/open-source (Apache 2.0) and offline. **Trade-off:** Porcupine is purpose-built for wake-word (lower CPU, faster detection, custom-trained models) but MAU-priced; Vosk is free but heavier (full ASR model loaded) and wake-word detection is a byproduct, not a primary feature. Snowboy is deprecated (acquired by Baidu, abandoned). On-device TensorFlow Lite wake-word is a build-it-yourself path (too much engineering for v0.5). **Recommendation: Porcupine for v0.5 (pilot-tier MAU pricing or built-in wake word), Vosk as the documented fallback if Picovoice pricing blocks the pilot.**
|
||||
|
||||
8. **Persona roster for v0.5: 4 active (lead-developer, voice-engineer REACTIVATED, backend-engineer, security-engineer RETAINED, data-engineer RETAINED), 2 deactivated (devops-engineer, frontend-engineer).** (0.85) v0.5 is voice-pipeline-heavy (wake-word + assist mode + latency tuning) + safety-critical guardrails + cohort aggregation extension. No deploy changes (v0.4 LXC carries forward) → devops-engineer deactivates. No new UI (wake-word is audio, assist is invoked by voice; the existing React app may need a small "assist mode" toggle but that's voice-engineer + backend territory, not a full frontend surface) → frontend-engineer deactivates unless the orchestrator decides an assist control surface is needed. See §7 for the full roster.
|
||||
|
||||
---
|
||||
|
||||
## Domain 1: Wake-Word Invocation on $100 Android (D-058, REQ-NFR-ASSIST-02)
|
||||
|
||||
### 1.1 Picovoice Porcupine on Android — verified capabilities
|
||||
|
||||
**Sources:** Picovoice Porcupine FAQ (https://picovoice.ai/docs/faq/porcupine/, fetched 2026-08-04), Porcupine Android quick-start (https://picovoice.ai/docs/quick-start/porcupine-android/, fetched 2026-08-04), Picovoice general FAQ (https://picovoice.ai/docs/faq/general/, fetched 2026-08-04).
|
||||
|
||||
**Finding (0.82):** Porcupine Wake Word is an on-device, offline keyword-spotting engine. Verified capabilities relevant to Praxis v0.5:
|
||||
|
||||
- **Android SDK exists** (quick-start page confirmed at `/docs/quick-start/porcupine-android/`). Also: React Native SDK (relevant if v0.5 upgrades the client from React web to React Native — currently v0.1 is React + WebRTC per D-015).
|
||||
- **On-device + offline.** No cloud round-trip for wake-word detection — critical for C-8 latency and for privacy (the mic isn't streaming to a cloud when listening for the wake word).
|
||||
- **Low resource.** Per Porcupine FAQ: "The standard model uses about 1 MB of memory and less than 4% of a single core on a Raspberry Pi 3." On a $100 Android phone (typically a quad-core 1.4-2.0GHz Cortex-A53, 2-3GB RAM), this is negligible. **Battery impact is minimal** — Porcupine is a lightweight neural net, not a full ASR model. The FAQ also notes: "Porcupine Wake Word is a lightweight engine with minimal consumption and requirements."
|
||||
- **Custom wake words.** Per FAQ: "You can train custom wake words with Porcupine on Picovoice Console, in seconds." This supports a Praxis-branded wake word (e.g., "Hey Praxis" or "Hey Coach"). Custom training is done on Picovoice Console (web UI), produces a `.ppn` model file bundled with the app.
|
||||
- **Accent-robust + universal.** Per FAQ: "Porcupine Wake Word detection software is universal and trained to work with a variety of accents and people's voices." Canadian English is well within Porcupine's trained distribution (English is a supported language — verified).
|
||||
- **Background mode.** Per FAQ: "Developers have been able to successfully run Porcupine Wake Word detection software on iOS and Android in background mode. However, this feature is controlled by the operating system, and we cannot guarantee that this will be possible in future releases of iOS or Android." **Risk: Android background-mic access is OS-controlled and has tightened in recent Android versions (Android 14+ requires foreground service with mic type for background audio).** Praxis v0.5 likely needs a foreground service (persistent notification) for wake-word listening while the phone is in pocket. This is a known Android pattern (used by "Hey Google", Shazam, etc.) — feasible but adds UX surface (notification) + battery.
|
||||
- **Multi-language.** English, French, German, Italian, Japanese, Korean, Mandarin, Portuguese, Spanish. Canadian English + (future) Canadian French are covered.
|
||||
|
||||
**Confidence 0.82** — vendor docs verified; the Android background-mic caveat is documented but the exact Android-version behavior needs a Phase-1 spike.
|
||||
|
||||
### 1.2 Picovoice pricing — the free-tier concern (D-058 refinement)
|
||||
|
||||
**Finding (0.75):** Per the Picovoice general FAQ (fetched 2026-08-04):
|
||||
|
||||
- Porcupine is priced on **monthly active users (MAU)**. A "user" is "typically a unique device, app, or browser instance that initializes the engine within a 30-day period."
|
||||
- There is a **Free Trial** ("No credit card is required. You can sign up at this link.") but it is **a one-time offer, not a recurring free tier**: "the Free Trial is a one-time offer, and it doesn't renew automatically once the trial ends."
|
||||
- "Picovoice is a B2B company focused on on-device AI tools for enterprises. At this time, there are no dedicated free or paid plans for personal or non-commercial use."
|
||||
|
||||
**This partially contradicts D-058's framing** ("free-tier supports custom wake words"). The Free Trial allows custom wake-word training and evaluation, but a recurring pilot (v0.5 ships and runs for weeks/months) would exhaust the trial and require a paid MAU plan. Praxis is not a personal/non-commercial user — it's a B2B pilot — so Picovoice sales engagement is the expected path.
|
||||
|
||||
**Resolution options for D-058 (flag for orchestrator):**
|
||||
|
||||
**(a) Engage Picovoice sales for a pilot/educational tier (RECOMMENDED).** Praxis is a Canada pilot for an educational/upskilling product — a natural fit for a Picovoice pilot-tier or educational discount. The MAU pricing for Porcupine at small scale (tens of devices) is typically modest. This is the cleanest path but requires a vendor conversation before v0.5 ships.
|
||||
|
||||
**(b) Use a built-in Picovoice wake word (not custom) for v0.5.** Porcupine ships built-in wake words (e.g., "Picovoice", "Alexa", "Hey Google", "Terminus", "Blueberry", "Grapefruit", "Bumblebee"). These may fall under different terms than custom-trained models. The Praxis pilot could use "Bumblebee" or "Grapefruit" (unusual enough to avoid false triggers in a retail environment) without custom training. **Reduces cost but loses the Praxis brand.**
|
||||
|
||||
**(c) Use Vosk as the wake-word engine (open-source fallback).** Vosk (Apache 2.0) is free, offline, runs on Android. Wake-word detection = run Vosk with a constrained grammar containing only the wake phrase. Heavier than Porcupine (full ASR model loaded, ~50MB) but no MAU cost. **Trade-off: free but more battery + CPU + engineering effort.**
|
||||
|
||||
**Recommendation: pursue (a) in parallel with (b) as the fallback.** Ship v0.5 with a built-in wake word (option b) if Picovoice sales engagement isn't resolved by ship date; switch to a custom Praxis wake word (option a) when the pilot tier is negotiated. Document option (c) as the post-pilot cost-reduction path if MAU pricing is unsustainable.
|
||||
|
||||
**Confidence 0.70** — the pricing concern is real (verified); the resolution depends on a vendor conversation not yet had.
|
||||
|
||||
### 1.3 Battery impact on a $100 Android phone
|
||||
|
||||
**Finding (0.72):** The Porcupine FAQ's "<4% of a single core on RPi 3" translates to roughly ~1-3% CPU on a modern $100 Android phone (Cortex-A53/A55 cores are comparable to RPi 3's ARM Cortex-A53). The wake-word listener runs as a foreground service with the mic open. Battery impact:
|
||||
|
||||
- **CPU:** ~1-3% continuous → negligible CPU drain.
|
||||
- **Mic:** continuous microphone sampling is the dominant battery cost. On modern Android, the mic + audio pipeline draws ~50-100mW during active listening. For an 8-hour shift, that's ~0.4-0.8 Wh — on a typical 3000-4000 mAh battery (~11-15 Wh), that's ~3-7% of battery per shift.
|
||||
- **Foreground service:** the persistent notification + service overhead adds ~1-2% battery per shift.
|
||||
- **Total estimate: ~4-9% battery per 8-hour shift.** Acceptable for a learner who starts the shift at 100% and the phone lasts the day. **Risk: if the learner is also using the phone for other work tasks (inventory app, point-of-sale), the combined drain may push them below 20% before shift end.** Mitigation: Praxis assist foreground service should be stoppable ("ending shift" closes the service), and the learner can tap-to-talk as a battery-saving fallback.
|
||||
|
||||
**Confidence 0.65** — battery estimates are back-of-envelope from power-draw heuristics, not measured on a target device. Phase 1 must measure on the actual $100 Android target.
|
||||
|
||||
### 1.4 Alternatives to Porcupine
|
||||
|
||||
**Finding (0.80):**
|
||||
|
||||
| Engine | License | Android | Offline | Custom WW | CPU/RAM | Status |
|
||||
|--------|---------|---------|---------|-----------|---------|--------|
|
||||
| **Picovoice Porcupine** | Proprietary, MAU-priced | ✅ SDK | ✅ | ✅ (Console) | ~1MB, <4% core | Active, maintained |
|
||||
| **Vosk** | Apache 2.0 | ✅ | ✅ | Via grammar | ~50MB model, more CPU | Active, maintained (verified alphacephei.com) |
|
||||
| **Snowboy** | Apache 2.0 (abandoned) | ✅ | ✅ | ✅ | Low | **Deprecated** — acquired by Baidu, no maintenance since ~2020. Reject. |
|
||||
| **TFLite wake-word** | DIY (Apache 2.0 models) | ✅ | ✅ | Train yourself | Varies | High engineering effort — train a custom KWS model (e.g., via TensorFlow Lite Micro). Out of scope for v0.5. |
|
||||
| **Android SpeechRecognizer (System)** | Free (Android API) | ✅ | ❌ (cloud) | ❌ | N/A | Cloud-based, latency + privacy. Reject for wake-word. |
|
||||
| **Cloud wake-word (Picovoice Falcon, etc.)** | Proprietary | ✅ | ❌ | ✅ | N/A | Cloud round-trip adds latency + connectivity dependency. Reject. |
|
||||
|
||||
**Verdict:** Porcupine for v0.5 (purpose-built, lowest resource, custom WW). Vosk as the documented open-source fallback. Snowboy rejected (deprecated). TFLite DIY rejected (engineering effort).
|
||||
|
||||
### 1.5 Android foreground service for background mic
|
||||
|
||||
**Finding (0.78):** Android (API 31+, Android 12+) requires a **foreground service of type `microphone`** for background audio capture. The service shows a persistent notification ("Praxis Assist is listening"). Key implementation points:
|
||||
|
||||
- `android.permission.RECORD_AUDIO` (dangerous permission — runtime grant).
|
||||
- `android.permission.FOREGROUND_SERVICE` + `android.permission.FOREGROUND_SERVICE_MICROPHONE` (Android 14+).
|
||||
- `Service.startForeground()` with a `Notification` (ongoing, low-priority).
|
||||
- The Porcupine Android SDK handles the audio capture loop; Praxis wraps it in a foreground service.
|
||||
- **Screen-off listening:** Android allows foreground services to keep the mic open when the screen is off (phone in pocket). The CPU may doze (Doze mode) but a foreground service with active mic is exempted from Doze for the mic pipeline.
|
||||
- **Risk: Android OEM battery kill switches.** Some manufacturers (Xiaomi, Huawei, OnePlus) aggressively kill background/foreground services to save battery. Praxis must document the "battery whitelist" step for learners (a known pain point for assistive apps). **Confidence 0.70** — the Android API is documented; OEM behavior is variable.
|
||||
|
||||
---
|
||||
|
||||
## Domain 2: 3-Layer Guardrail Enforcement (D-060, REQ-ASSIST-03)
|
||||
|
||||
### 2.1 The 3-layer pattern is industry-standard
|
||||
|
||||
**Finding (0.85):** D-060 specifies 3 layers: (1) prompt-layer rules, (2) output filter, (3) audit logging. This is the standard defense-in-depth pattern for LLM safety, matching:
|
||||
- **OpenAI's moderation pattern** (input + output moderation + logging).
|
||||
- **NVIDIA NeMo Guardrails** (input rails + dialog rails + output rails + execution rails — same layering, more granular).
|
||||
- **LLM-as-judge guardrail patterns** (system prompt constraints + post-generation classifier + audit trail).
|
||||
|
||||
The existing `CustomerServiceGuardrail` (server/guardrails/customer_service.py, verified — 129 lines) already implements layer (2): regex-based output filtering for legal/financial/medical advice + impersonation, with a `_filter_legal()` rewrite. Layer (1) is the system prompt (scenario-driven, set in `pipeline.py:_build_llm_context`). Layer (3) is the `turns` SQLite table (session_recorder.py). v0.5 extends all three layers for Live Assist.
|
||||
|
||||
**Confidence 0.85** — the pattern is well-established; the existing code confirms the architecture.
|
||||
|
||||
### 2.2 Layer 1 — Prompt rules for "coaches not does"
|
||||
|
||||
**Finding (0.82):** The Live Assist system prompt must explicitly instruct the LLM to:
|
||||
- **Ask guiding questions, never give the answer.** "Your role is to coach, not to do the learner's job. Ask questions that help the learner arrive at the answer themselves."
|
||||
- **Never speak on behalf of the learner.** "You are not a participant in the learner's conversation with their customer. Do not generate text the learner should say verbatim."
|
||||
- **Never claim authority you don't have.** "You are a coaching AI, not a manager, not a company representative, not a legal/medical/financial advisor."
|
||||
- **Stay within the bound context.** "You are coaching the learner on `[path week scenario tag]`. Do not give advice outside this scope."
|
||||
- **Keep responses short for voice (1-3 sentences).** Carry-forward from v0.1's voice-conciseness rule.
|
||||
- **Acknowledge the real customer's presence implicitly.** "The learner is in a live interaction. Your coaching must be brief enough not to distract, and must never instruct the learner to say something untrue to the customer."
|
||||
|
||||
This prompt is the `LiveAssistGuardrail.session_start_disclaimer` + the system-prompt prefix. The existing `_build_llm_context()` in pipeline.py constructs the messages list — v0.5 adds an assist-mode branch that injects the coaching prompt instead of the role-play scenario prompt.
|
||||
|
||||
**Confidence 0.82** — prompt engineering is the well-trodden path; the specific phrasing needs Phase-1 iteration + testing against a red-team prompt set.
|
||||
|
||||
### 2.3 Layer 2 — Output filter patterns for "direct answer" vs "coaching question"
|
||||
|
||||
**Finding (0.80):** The output filter is a regex + keyword classifier on the LLM response text, run after LLM generation and before TTS. Patterns:
|
||||
|
||||
**Direct-answer patterns (BLOCK or REWRITE):**
|
||||
```python
|
||||
# "you should say X to the customer" — verbatim script
|
||||
DIRECT_SCRIPT_RE = re.compile(
|
||||
r"\b(you should (say|tell|respond with|reply)|"
|
||||
r"say (this|the following)|"
|
||||
r"tell (the |a )?customer|"
|
||||
r"respond with|reply with|"
|
||||
r"here'?s what to say|"
|
||||
r"the (right |correct |best )?answer is|"
|
||||
r"what you (should|need to|must) (say|do) is)\b",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
# Imperative commands to the learner about the customer
|
||||
IMPERATIVE_RE = re.compile(
|
||||
r"\b(escalate to|transfer to|offer a refund of|apologize (by|with)|"
|
||||
r"give them|promise them|tell them you)\b",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
# Claiming authority / false authority
|
||||
FALSE_AUTHORITY_RE = re.compile(
|
||||
r"\b(I (am|'?m) (your |a )?(manager|supervisor|the company|authorized|"
|
||||
r"a lawyer|a doctor|regulator)|"
|
||||
r"on behalf of (the company|management)|"
|
||||
r"I (can|will) (authorize|approve|guarantee))\b",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
# Impersonation of the customer or a real company (carry-forward from CS guardrail)
|
||||
# (reuse _IMPERSONATION_RE from customer_service.py)
|
||||
```
|
||||
|
||||
**Coaching-question patterns (ALLOW — these are the desired output):**
|
||||
```python
|
||||
# Open-ended guiding questions
|
||||
COACHING_QUESTION_RE = re.compile(
|
||||
r"\b(what (do you|could you|might you)|"
|
||||
r"how (could|might|would|do) you|"
|
||||
r"what'?s (your|the) (goal|approach|next step)|"
|
||||
r"how (does|do) you (feel|think)|"
|
||||
r"what (would|might) happen if|"
|
||||
r"can you (think of|identify|name)|"
|
||||
r"have you considered)\b",
|
||||
re.IGNORECASE,
|
||||
)
|
||||
```
|
||||
|
||||
**Filter logic:**
|
||||
1. Run direct-answer patterns. If hit → **block** the response, log the verdict, and re-prompt the LLM with "Your last response gave a direct answer. Rephrase as a coaching question." (one retry; if retry also hits, fall back to a canned coaching redirect: "Think about what the customer needs right now. What's your next step?").
|
||||
2. Run false-authority + impersonation patterns. If hit → **block** + log + no retry (these are hard violations).
|
||||
3. If no direct-answer hit → allow. Optionally score the response: if it contains a coaching-question pattern, mark `category="coaching"`; else `category="neutral"` (allowed but not ideal — log for review).
|
||||
|
||||
**False-positive risk:** the direct-answer regex may flag legitimate coaching that quotes a customer's likely response ("If the customer says X, you might explore Y"). Mitigation: the regex targets imperative/script phrasing ("you should say"), not hypothetical/quoted phrasing ("if the customer says"). Phase-1 must tune the regex against a corpus of real coaching responses.
|
||||
|
||||
**Confidence 0.78** — regex-based output filtering is the existing pattern (customer_service.py proves it); the specific patterns need a red-team tuning pass.
|
||||
|
||||
### 2.4 Layer 3 — Audit logging
|
||||
|
||||
**Finding (0.85):** All assist turns logged to SQLite `turns` table (existing — verified in session_recorder.py:log_turn). v0.5 adds:
|
||||
- A `guardrail_verdict` JSON field on the `turns` table (or a parallel `guardrail_verdicts` table keyed by turn id) capturing `{allowed, reason, category, filtered_text}` per the `GuardrailVerdict` dataclass (services/base.py).
|
||||
- Assist turns flow to the v0.4 cohort aggregation as `session_type=assist` with a `guardrail_block_rate` metric (how often the output filter fired). **This gives operators visibility into safety-critical guardrail behavior** — a sudden spike in block rate signals either a prompt regression or a population of learners pushing the boundary.
|
||||
- **No raw learner PII in the audit log beyond the existing hardcoded `learner-1` (D-007).** The turn text is learner speech + AI coaching; stored in SQLite (local), aggregated k-anonymized in Postgres (D-031 hybrid preserved).
|
||||
|
||||
**Confidence 0.85** — the audit table exists; the extension is a schema-additive migration.
|
||||
|
||||
### 2.5 LLM-as-judge for periodic guardrail evaluation (optional, post-v0.5)
|
||||
|
||||
**Finding (0.65):** A stronger pattern (deferred post-v0.5) is an **LLM-as-judge** that periodically samples assist turns and classifies them as "coached" vs "did the job" with higher accuracy than regex. This runs off the voice path (nightly job, like the v0.4 cohort reconciliation) and produces a "guardrail adherence score" per learner/shift. v0.5 ships regex filtering (fast, on the voice path); v0.6+ adds the LLM-judge (accurate, off the voice path). **Confidence 0.65** — the pattern is sound but deferred; not a v0.5 blocker.
|
||||
|
||||
### 2.6 Known incidents / failure modes in on-the-job coaching AI
|
||||
|
||||
**Finding (0.70 — domain knowledge, not vendor-verified):** Known failure modes for AI-in-the-ear-during-real-customer-interaction:
|
||||
- **The "parrot" failure:** the AI gives a verbatim script, the learner repeats it word-for-word, the customer detects the robotic delivery → trust erosion. (Mitigated by D-060 layer 2 — direct-script pattern blocking.)
|
||||
- **The "hallucinated authority" failure:** the AI claims to be a manager/supervisor, the learner parrots it, the customer escalates to a real manager who disavows. (Mitigated by `FALSE_AUTHORITY_RE`.)
|
||||
- **The "wrong-context" failure:** the AI coaches for the wrong scenario (e.g., refund when the customer is asking about a delivery). (Mitigated by D-059 context-binding — learner declares context at session start.)
|
||||
- **The "over-coaching" failure:** the AI speaks too much, the learner misses the customer's next utterance. (Mitigated by the 1-3 sentence voice-conciseness rule + interruptibility D-008.)
|
||||
- **The "latency-killed-the-moment" failure:** coaching arrives after the customer moment passed. (Mitigated by C-8 <600ms budget — see Domain 3.)
|
||||
- **Privacy/consent failure:** the real customer didn't consent to being recorded/analyzed by an AI. (Mitigated by: Praxis assist is *coaching the learner*, not recording the customer; the mic captures the learner's side primarily. But the ambient mic may pick up the customer. **Flag: the foreground-service notification + a learner-facing disclosure ("Assist is on — those around you may be recorded by your mic") is ethically and legally required.** This is a safety/legal surface for the orchestrator to review.**
|
||||
|
||||
No direct competitor does live-in-ear coaching during real customer calls (verified — Dialpad Ai Coach and Gong are post-hoc call analysis, not live; RealWear is AR + voice for industrial, not phone-in-pocket CS coaching). So Praxis is in novel safety territory — the guardrail design must be conservative.
|
||||
|
||||
---
|
||||
|
||||
## Domain 3: <600ms Latency Budget for Assist Turns (D-061, REQ-NFR-ASSIST-01)
|
||||
|
||||
### 3.1 v0.1 budget breakdown (carry-forward)
|
||||
|
||||
**Finding (0.85):** From ARCHITECTURE.md (verified):
|
||||
|
||||
| Segment | Budget | Note |
|
||||
|---------|--------|------|
|
||||
| Client capture + WebRTC uplink | ~50ms | |
|
||||
| ASR (Deepgram Nova-3 first partial) | ~250ms | R1: measure in Phase 1 |
|
||||
| LLM first token (gemma4:cloud) | ~200ms | R3: measure in Phase 1 |
|
||||
| TTS first audio (Cartesia Sonic) | ~120ms | R2: measure; Piper fallback ~80ms |
|
||||
| WebRTC downlink + playback | ~50ms | |
|
||||
| **Total (all-cloud, Cartesia)** | **~670ms** | ⚠️ Marginally over 600ms |
|
||||
| **Total (Piper TTS)** | **~550ms** | R4 mitigation |
|
||||
|
||||
**v0.1's R4 risk (the single biggest v0.1 technical risk):** the all-cloud path likely lands ~670ms. The TTS service MUST sit behind an interface (D-014) and Piper-on-pilot-server MUST be pre-staged as the likely production v0.1 TTS.
|
||||
|
||||
### 3.2 What does assist mode add to the budget?
|
||||
|
||||
**Finding (0.78):** Assist mode adds **context-binding tokens** to the LLM system prompt. The context-binding is:
|
||||
- Path week (e.g., "Week 3: Handling escalations")
|
||||
- Scenario tag (e.g., "damaged-product refund")
|
||||
- Learner state summary (e.g., "current_theta=0.2, working on de-escalation")
|
||||
- Coaching focus (e.g., "Focus: empathy + resolution-concreteness")
|
||||
- The coaching-mode instruction (layer 1 guardrail prompt — see §2.2)
|
||||
|
||||
Estimated token count for the context-binding: ~100-150 tokens (the coaching-mode instruction is ~80 tokens; the context-binding is ~30-50 tokens). Total system prompt for assist: ~150-230 tokens (vs. v0.1 practice: ~50-100 tokens for the role-play character prompt).
|
||||
|
||||
**Latency impact of extra input tokens:** LLM prefill (time-to-first-token) scales roughly linearly with input token count for a fixed output. For `gemma4:cloud` (256K context, well within budget), the prefill latency for ~150 input tokens vs ~50 input tokens is the difference of ~100 tokens × ~0.5ms/token ≈ **+50ms** (conservative; could be up to +100ms depending on the model's prefill speed). This is added to the LLM first-token segment.
|
||||
|
||||
**Revised assist budget (all-cloud, Cartesia):**
|
||||
|
||||
| Segment | Budget | Note |
|
||||
|---------|--------|------|
|
||||
| Client capture + WebRTC uplink | ~50ms | |
|
||||
| ASR (Deepgram Nova-3) | ~250ms | |
|
||||
| LLM first token (gemma4:cloud, +context-binding) | ~250-300ms | +50-100ms for context prefill |
|
||||
| TTS first audio (Cartesia) | ~120ms | |
|
||||
| WebRTC downlink + playback | ~50ms | |
|
||||
| **Total (all-cloud, Cartesia)** | **~720-770ms** | ⚠️ Breaks C-8 |
|
||||
|
||||
**Revised assist budget (Piper TTS mitigation):**
|
||||
|
||||
| Segment | Budget | Note |
|
||||
|---------|--------|------|
|
||||
| Client capture + WebRTC uplink | ~50ms | |
|
||||
| ASR (Deepgram Nova-3) | ~250ms | |
|
||||
| LLM first token (gemma4:cloud, +context-binding) | ~250ms | lean context (~100 tokens) |
|
||||
| TTS first audio (Piper, self-hosted) | ~80ms | R4 mitigation |
|
||||
| WebRTC downlink + playback | ~50ms | |
|
||||
| **Total (Piper)** | **~680ms** | ⚠️ Still marginal |
|
||||
|
||||
### 3.3 How to get assist under 600ms
|
||||
|
||||
**Finding (0.72):** Three levers, in order of impact:
|
||||
|
||||
1. **Minimize the system prompt.** The assist system prompt should be ≤150 input tokens total (coaching instruction + context-binding). This is achievable: the coaching instruction is a fixed ~80-token block; the context-binding is a terse ~30-50 tokens ("Week 3, damaged-refund, focus: empathy"). Avoid dumping the full rubric or scenario YAML into the prompt. **Saves ~25-50ms** vs. a verbose prompt.
|
||||
|
||||
2. **Use Piper TTS for assist turns (not Cartesia).** Piper self-hosted on the pilot server is ~80ms first audio vs. Cartesia's ~120ms. **Saves ~40ms.** The v0.1 architecture already pre-stages Piper (R4 mitigation); v0.5 assist mode defaults to Piper, with Cartesia as the quality fallback for practice mode (where <600ms is desired but not as safety-critical — practice coaching that arrives a beat late is still useful; live-assist coaching that arrives after the customer moment is useless).
|
||||
|
||||
3. **Lean LLM model for assist.** `gemma4:cloud` is the role-play fast path. For assist, where the output is a short coaching question (not a role-play character utterance), a smaller/faster model may suffice. **Option: use a lighter Ollama model for assist** (e.g., a future `gemma4:e2b:cloud` if available — the v0.1 RESEARCH noted `gemma4:e2b`/`e4b` as future options). For v0.5, keep `gemma4:cloud` (no new model risk) but document the lighter-model path for v0.6.
|
||||
|
||||
**With levers 1 + 2 applied:**
|
||||
|
||||
| Segment | Budget | Note |
|
||||
|---------|--------|------|
|
||||
| Client capture + WebRTC uplink | ~50ms | |
|
||||
| ASR (Deepgram Nova-3) | ~250ms | |
|
||||
| LLM first token (gemma4:cloud, lean assist prompt) | ~225ms | +25ms for ~50 extra tokens over v0.1 |
|
||||
| TTS first audio (Piper) | ~80ms | |
|
||||
| WebRTC downlink + playback | ~50ms | |
|
||||
| **Total (Piper, lean prompt)** | **~655ms** | ⚠️ Still 55ms over |
|
||||
|
||||
**Still marginal.** The hard truth: the all-cloud + on-device-mic path is ~655ms with the best levers. To get under 600ms, v0.5 needs either:
|
||||
- **(a) Measured Deepgram latency < 250ms.** The v0.1 R1 risk ("measure in Phase 1") — if Deepgram Nova-3 first-partial is ~200ms in Canada (plausible — Deepgram's streaming is fast), the total drops to ~605ms (close enough; C-8 is a target, not a hard ceiling for the pilot).
|
||||
- **(b) Measured gemma4:cloud first-token < 200ms.** R3 — if Ollama Cloud is fast (~150ms), total drops to ~580ms. ✅ Under budget.
|
||||
- **(c) Accept ~650ms for the pilot, document the gap, target <600ms in v0.6 with optimization.** The pilot is Canada, relaxed C-3 (cost); C-8 (latency) is a target. A 50ms overrun on assist turns is tolerable for a pilot if it's measured and trending down.
|
||||
|
||||
**Recommendation: ship v0.5 with the Piper + lean-prompt configuration, measure the actual assist latency in Phase 1, and treat <600ms as a v0.5 target with a v0.6 hardening step.** Document the ~650ms estimate + the levers. **Flag for orchestrator: assist turns likely land ~655-770ms depending on which TTS + how lean the prompt is; C-8 <600ms is at risk for assist mode. The binding constraint is C-8, so this is a real tension — the orchestrator should decide whether to relax C-8 for assist mode or push for v0.6 optimization.**
|
||||
|
||||
**Confidence 0.70** — the budget math is sound; the actual Deepgram/Ollama/Piper latencies are unmeasured (R1/R3/R4 from v0.1).
|
||||
|
||||
### 3.4 Wake-word → first-audio latency budget
|
||||
|
||||
**Finding (0.80):** The wake-word → first-audio path is distinct from the in-conversation turn budget. After the learner says "Hey Praxis, the customer is asking about a refund":
|
||||
|
||||
| Segment | Budget | Note |
|
||||
|---------|--------|------|
|
||||
| Wake-word detection (Porcupine, on-device) | ~200-500ms | detection latency after the wake word ends |
|
||||
| Foreground service → WebRTC connect (if not already connected) | ~0ms (warm) / ~500-1000ms (cold) | The assist foreground service should keep a warm WebRTC connection to the praxis server during the shift; cold-connect is too slow |
|
||||
| User speech (post wake-word) → ASR | ~250ms | Deepgram, as in-conversation |
|
||||
| LLM + TTS + downlink | ~400ms | lean prompt + Piper |
|
||||
| **Total (warm WebRTC)** | **~850-1150ms** | From wake-word-end to first coaching audio |
|
||||
| **Total (cold WebRTC)** | **~1350-2150ms** | Cold connect is unacceptable for live assist |
|
||||
|
||||
**Critical: the assist foreground service must keep a warm WebRTC connection during the shift.** This is a new architectural requirement vs. v0.1 (where each practice session is a fresh WebRTC connection). v0.5 assist mode opens a long-lived WebRTC connection at shift start, keeps it alive (heartbeat), and reuses it for every assist turn. **Battery cost:** WebRTC keepalive is ~minimal (UDP heartbeat every 15-30s). **Server cost:** the praxis server holds a long-lived Pipecat task per active assist shift (vs. per practice session in v0.1). This is a concurrency change — see Domain 5.
|
||||
|
||||
**Confidence 0.75** — the wake-word latency is from Porcupine docs (detection is fast but not instant); the warm-WebRTC requirement is a design implication.
|
||||
|
||||
---
|
||||
|
||||
## Domain 4: Shift-Bounded Session Model (D-062, REQ-NFR-ASSIST-04)
|
||||
|
||||
### 4.1 How real on-the-job coaching assistants bound sessions
|
||||
|
||||
**Finding (0.72):** Survey of on-the-job coaching AI products (domain knowledge + verified where possible):
|
||||
|
||||
| Product | Session model | Live or post-hoc | Surface |
|
||||
|---------|---------------|------------------|---------|
|
||||
| **Dialpad Ai Coach** | Per-call (post-hoc analysis of the call recording) | Post-hoc | Business VoIP (not in-ear during the call) |
|
||||
| **Gong** | Per-meeting (post-hoc analysis of sales call recordings) | Post-hoc | Business comms (revenue intelligence) |
|
||||
| **RealWear** (verified realwear.com) | Continuous (wearable, always on during the shift) | Live (AR + voice) | Industrial frontline (hardware: smart glasses) |
|
||||
| **Balance AI** | (domain knowledge) Per-conversation coaching | Live (app-based) | General coaching app (not CS-specific) |
|
||||
| **Praxis v0.5 (proposed)** | **Shift-bounded** (learner starts/ends a shift; assist turns within) | **Live (in-ear)** | **Phone-in-pocket, CS coaching** |
|
||||
|
||||
**No direct competitor does "live-in-ear coaching during real customer calls on a $100 phone."** Dialpad/Gong are post-hoc (analysis after the call). RealWear is live but AR + industrial (not phone-in-pocket CS). Praxis v0.5 is novel.
|
||||
|
||||
**The shift-bounded model (D-062) is the right choice** because:
|
||||
- It matches the real-world unit of labor (shifts) for retail/hospitality/CS — the Customer Service path's target.
|
||||
- It gives a clean aggregation boundary (a shift is a discrete event with a start/end timestamp).
|
||||
- It bounds the WebRTC connection lifecycle (warm connection for the shift, closed at shift-end).
|
||||
- It avoids the ambiguity of "continuous" (when does aggregation fire? when does the connection close?) and the granularity of "per-turn" (too many aggregation events, double-counting risk).
|
||||
|
||||
**Confidence 0.80** — the shift model is well-matched to the use case; the competitor survey confirms Praxis is novel.
|
||||
|
||||
### 4.2 Shift lifecycle
|
||||
|
||||
**Finding (0.82):** The shift lifecycle:
|
||||
|
||||
```
|
||||
1. Learner opens Praxis app, taps "Start Shift" (or voice: "Hey Praxis, starting my shift").
|
||||
├─ Foreground service starts (Porcupine wake-word listener on).
|
||||
├─ Learner declares context: taps current path week + scenario tag (D-059).
|
||||
│ └─ Server reads learner.progress.current_week from SQLite (D-007) for rubric alignment.
|
||||
├─ Warm WebRTC connection opens to praxis server.
|
||||
└─ Shift session row created in SQLite (session_type='assist', started_at=now()).
|
||||
|
||||
2. During the shift, learner invokes assist:
|
||||
├─ "Hey Praxis" → Porcupine detects → foreground service routes audio to WebRTC.
|
||||
├─ Learner speaks (the situation / their question).
|
||||
├─ Pipeline: ASR → LLM (coaching prompt + context-binding) → guardrail filter → TTS.
|
||||
├─ Coaching plays in-ear. Turn logged (turns table, session_id=shift_id).
|
||||
└─ WebRTC connection stays warm for the next turn.
|
||||
|
||||
3. Learner ends shift: "Hey Praxis, ending shift" (or taps "End Shift").
|
||||
├─ Foreground service stops (Porcupine off, mic released).
|
||||
├─ WebRTC connection closed.
|
||||
├─ Shift session row updated (ended_at, outcome='completed', turn_count).
|
||||
└─ on-session-end hook fires → cohort aggregation (session_type='assist') → Postgres.
|
||||
```
|
||||
|
||||
**Within a shift:** each assist turn is a discrete coaching exchange. Turns are logged to the `turns` table with `session_id` = the shift's session id. The shift is the aggregation unit (not the turn).
|
||||
|
||||
**Confidence 0.82** — the lifecycle is concrete and matches the existing `SessionRecorder` pattern (start → log_turn → end).
|
||||
|
||||
### 4.3 Assist does not update mastery (D-063)
|
||||
|
||||
**Finding (0.90):** D-063 is unambiguous: assist turns never update θ (D-035) or count toward mastery gates (D-032). The `run_mastery_flow()` in session_recorder.py (verified — lines 206-363) is invoked only for practice sessions (`schedule_mastery=True`); assist shifts call `end()` with `schedule_mastery=False`. The cohort aggregation hook fires for both session types, but the mastery flow is practice-only. **This is enforced in the `end()` signature** — the `schedule_mastery` flag gates the mastery asyncio task. **Confidence 0.90** — the code structure already supports the separation.
|
||||
|
||||
### 4.4 Integration with the v0.4 cohort aggregation
|
||||
|
||||
**Finding (0.85):** The v0.4 aggregation pipeline (server/cohort/aggregator.py, verified) keys cells by `(path, metric, window_start)`. v0.5 adds assist-specific metrics as new `metric` strings in the same `cohort_aggregates` table — **no schema change** (the table is generic on `metric TEXT`).
|
||||
|
||||
**Assist metrics (new):**
|
||||
| Metric | Description | Aggregation |
|
||||
|--------|-------------|-------------|
|
||||
| `assist_shifts_count` | Number of assist shifts in the window | count |
|
||||
| `assist_turns_count` | Total assist turns across shifts | sum |
|
||||
| `assist_avg_turns_per_shift` | Mean turns per shift | mean |
|
||||
| `assist_active_learners_count` | Distinct learners using assist | distinct count (k-anon) |
|
||||
| `assist_guardrail_block_rate` | Fraction of assist turns where the output filter blocked | mean |
|
||||
|
||||
**Integration with D-053's 3 dashboard views:**
|
||||
- **Practice volume** → **Practice + Assist volume**: add `assist_shifts_count` + `assist_turns_count` to the practice volume view (or a new "Assist volume" sub-view).
|
||||
- **Mastery progression** → unchanged (assist doesn't affect mastery per D-063).
|
||||
- **Failure patterns** → add `assist_guardrail_block_rate` as a safety signal (a high block rate = the AI is frequently trying to give direct answers = either a prompt regression or learners pushing boundaries).
|
||||
|
||||
**The on-session-end hook (server/cohort/hook.py) is extended** to accept `session_type='assist'` in the `session_outcome` dict. The `_build_session_outcome()` method in session_recorder.py (line 164) already builds this dict; v0.5 adds the `session_type` field. Assist shifts fire the hook on shift-end (not per-turn).
|
||||
|
||||
**k-anonymity ≥ 10 (D-034) applies identically** — assist metrics are suppressed if the distinct learner count in the window is < 10. **Confidence 0.85** — the integration is additive; the existing aggregator + hook patterns are reused.
|
||||
|
||||
---
|
||||
|
||||
## Domain 5: v0.1 Voice Pipeline Reuse for Assist Mode (D-061)
|
||||
|
||||
### 5.1 The pipeline is parameterized for reuse
|
||||
|
||||
**Finding (0.85):** `server/pipeline.py:build_pipeline()` (verified — 231 lines) takes a `scenario_id` and builds a `ScenarioRuntime` with a system prompt + opening line. The pipeline is:
|
||||
```
|
||||
transport.input() → stt → latency_observer → user_aggregator → llm →
|
||||
latency_observer → tts → latency_observer → transport.output() → assistant_aggregator
|
||||
```
|
||||
All service constructors (`_build_stt`, `_build_llm`, `_build_tts`, `_build_transport`) are env-driven and reusable. The only scenario-specific parts are the system prompt + opening line (from `ScenarioRuntime`).
|
||||
|
||||
### 5.2 Minimal delta: build_assist_pipeline()
|
||||
|
||||
**Finding (0.82):** v0.5 adds a `build_assist_pipeline()` (or a `mode="assist"` parameter to `build_pipeline()`) that:
|
||||
- Reuses `_build_transport`, `_build_stt`, `_build_llm`, `_build_tts` unchanged.
|
||||
- Swaps `_build_llm_context()`: instead of the scenario-driven system prompt, injects the **Live Assist coaching prompt** (§2.2) + **context-binding** (path week, scenario tag, learner state).
|
||||
- Drops the opening line (assist is invoked mid-shift; no scripted opener).
|
||||
- Adds the **LiveAssistGuardrail** as a post-LLM processor (between `llm` and `tts` in the pipeline) that runs the output filter (§2.3). The existing v0.1 pipeline doesn't have a post-LLM guardrail processor inline (the CS guardrail runs on the debrief, not in-loop) — **v0.5 adds an in-loop guardrail processor for assist mode**. This is a pipeline-structure change but a small one (~1 new Pipecat frame processor).
|
||||
- Reuses the `LatencyObserver` for assist latency measurement (R1/R3/R4 measurement extends to assist turns).
|
||||
|
||||
**Delta estimate: ~1 new pipeline builder (~50 LOC), ~1 new guardrail processor (~80 LOC), ~1 new guardrail ruleset (LiveAssistGuardrail, ~120 LOC), ~1 new context-binding loader (~40 LOC).** Total: ~290 LOC of new server code. No new voice-service deps (Deepgram/Cartesia/Piper/Ollama all reused).
|
||||
|
||||
**Confidence 0.82** — the pipeline structure is clean; the delta is small.
|
||||
|
||||
### 5.3 Warm WebRTC connection — the concurrency change
|
||||
|
||||
**Finding (0.78):** v0.1 opens a fresh WebRTC connection per practice session (short-lived, 5-10 min). v0.5 assist mode keeps a **warm WebRTC connection for the entire shift** (potentially 4-8 hours). Implications:
|
||||
|
||||
- **Server concurrency:** the praxis server holds N long-lived Pipecat tasks (one per active assist shift) vs. M short-lived practice tasks. For the pilot (single-learner-per-device, D-007), N ≤ 1. For post-pilot (multi-learner), N = number of concurrent learners on-shift. **The v0.4 single-uvicorn process + asyncpg pool (max 10) is sufficient for the pilot** (1 concurrent assist shift + occasional practice sessions). Post-pilot concurrency is a v0.6+ concern.
|
||||
- **WebRTC keepalive:** the SmallWebRTCTransport (Pipecat) keeps the connection alive via ICE keepalives (STUN binding requests every 15-30s by default). Praxis adds an app-level heartbeat (a no-op audio frame or a ping message) every 30s to ensure the connection isn't reaped by NAT timeouts.
|
||||
- **Battery (client):** WebRTC keepalive is ~minimal (UDP, small packets). The mic is only active during an assist turn (post-wake-word); between turns, the foreground service runs Porcupine on the local mic but doesn't stream to the server. **The WebRTC connection is warm (keepalive only) between assist turns; audio streams only during a turn.**
|
||||
|
||||
**Confidence 0.75** — the warm-connection pattern is standard WebRTC; the concurrency math is pilot-scale.
|
||||
|
||||
### 5.4 Context-binding source (D-059)
|
||||
|
||||
**Finding (0.82):** D-059 specifies: learner declares context at session start (path + scenario tag), server reads active path week from SQLite. The existing `PraxisStore.get_progress(learner_id, path_slug)` (used in session_recorder.py:249) returns the learner's progress row including `current_week`. v0.5 assist mode:
|
||||
1. Learner taps "Start Shift" → selects current path week (or confirms the auto-detected `progress.current_week`) + scenario tag (e.g., "damaged-product refund").
|
||||
2. Server loads the context: `current_week` from SQLite + the scenario tag's `rubric_criteria` from the scenario library + the learner's `theta` from `learner_ability`.
|
||||
3. The context-binding loader constructs a terse context string: `"Week {current_week}, scenario: {scenario_tag}, learner_theta: {theta:.1f}, coaching_focus: {top_rubric_criterion}"`.
|
||||
4. This string is injected into the assist system prompt.
|
||||
|
||||
**Auto-detection is out of scope** (no camera per C-4, no screen context). The learner is in control of declaring context. **Confidence 0.82** — the existing store methods support the read; the declaration UI is a small client addition.
|
||||
|
||||
---
|
||||
|
||||
## Domain 6: Cohort Aggregation Integration (D-062, REQ-NFR-ASSIST-04) — detailed
|
||||
|
||||
### 6.1 No schema change to cohort_aggregates
|
||||
|
||||
**Finding (0.90):** The `cohort_aggregates` table (db/pg_migrations/0001_operator_tier.sql, verified):
|
||||
```sql
|
||||
CREATE TABLE IF NOT EXISTS cohort_aggregates (
|
||||
path TEXT NOT NULL,
|
||||
metric TEXT NOT NULL,
|
||||
window_start DATE NOT NULL,
|
||||
window_end DATE NOT NULL,
|
||||
value NUMERIC,
|
||||
cell_count INTEGER NOT NULL DEFAULT 0,
|
||||
cell_suppressed BOOLEAN NOT NULL DEFAULT FALSE,
|
||||
updated_at TIMESTAMPTZ NOT NULL DEFAULT now(),
|
||||
PRIMARY KEY (path, metric, window_start)
|
||||
);
|
||||
```
|
||||
The `metric` column is free-form TEXT. v0.5 adds assist metrics (`assist_shifts_count`, `assist_turns_count`, etc.) as new `metric` values — **no DDL change**. The aggregation upsert (aggregator.py:_upsert_cell) is metric-agnostic. **Confidence 0.90** — the schema is generic by design (D-053).
|
||||
|
||||
### 6.2 session_type field in session_outcome
|
||||
|
||||
**Finding (0.85):** The `_build_session_outcome()` in session_recorder.py (line 164) builds the dict the aggregator consumes. v0.5 adds:
|
||||
```python
|
||||
def _build_session_outcome(self, outcome: str) -> dict[str, Any]:
|
||||
return {
|
||||
"learner_ref": self.learner_id,
|
||||
"path": self._path_slug(),
|
||||
"scenario_id": self.scenario_id,
|
||||
"outcome": outcome,
|
||||
"session_type": self.session_type, # NEW v0.5: 'practice' | 'assist'
|
||||
"rubric_scores": ..., # empty for assist (no mastery scoring)
|
||||
"failure_mode": self._failure_mode(), # None for assist
|
||||
"branch_path": list(self._branch_path), # empty for assist
|
||||
"assist_turn_count": self._turn_seq, # NEW v0.5
|
||||
"guardrail_blocks": self._guardrail_block_count, # NEW v0.5
|
||||
"timestamp": _now_iso(),
|
||||
}
|
||||
```
|
||||
The `SessionRecorder.__init__` gains a `session_type: str = "practice"` parameter. Practice sessions set it to `"practice"` (default); assist shifts set it to `"assist"`. The aggregator branches on `session_type` to compute the right metrics.
|
||||
|
||||
### 6.3 Aggregator extension for assist
|
||||
|
||||
**Finding (0.82):** `aggregator.py:aggregate_session()` (verified) branches on `session_type`:
|
||||
|
||||
```python
|
||||
async def aggregate_session(pg_store, session_outcome):
|
||||
session_type = session_outcome.get("session_type", "practice")
|
||||
if session_type == "assist":
|
||||
await _aggregate_assist(pg_store, session_outcome)
|
||||
else:
|
||||
await _aggregate_practice(pg_store, session_outcome) # existing logic
|
||||
|
||||
async def _aggregate_assist(pg_store, session_outcome):
|
||||
path = session_outcome["path"]
|
||||
turn_count = session_outcome.get("assist_turn_count", 0)
|
||||
blocks = session_outcome.get("guardrail_blocks", 0)
|
||||
# ... upsert assist_shifts_count, assist_turns_count, assist_avg_turns_per_shift,
|
||||
# assist_guardrail_block_rate with k-anon suppression (same pattern as practice)
|
||||
```
|
||||
|
||||
The k-anonymity suppression (`COUNT(DISTINCT learner_ref) >= 10`) applies identically — assist metrics are suppressed if too few learners used assist in the window. **Confidence 0.82** — the extension mirrors the existing practice aggregation.
|
||||
|
||||
### 6.4 Dashboard views extension (D-053)
|
||||
|
||||
**Finding (0.80):** The 3 v0.4 dashboard views (server/operator/cohort.py, mastery.py, failure_patterns.py) extend:
|
||||
|
||||
| v0.4 View | v0.5 Extension |
|
||||
|-----------|----------------|
|
||||
| Practice volume (cohort.py) | Add assist rows: `assist_shifts_count`, `assist_turns_count` per path/window. The view returns practice + assist volume side-by-side. |
|
||||
| Mastery progression (mastery.py) | Unchanged (assist doesn't affect mastery per D-063). Optionally add a note: "Assist usage: N shifts, M turns this window" as context. |
|
||||
| Failure patterns (failure_patterns.py) | Add `assist_guardrail_block_rate` as a new "safety signal" row. High block rate = flag for operator review. |
|
||||
|
||||
No new endpoints — the existing `/api/operator/cohort`, `/api/operator/mastery`, `/api/operator/failure-patterns` return extended payloads. The React dashboard (client/src/operator/) renders the new rows. **Confidence 0.80** — the extension is additive to the existing views.
|
||||
|
||||
---
|
||||
|
||||
## Domain 7: Persona Roster for v0.5 (decision)
|
||||
|
||||
### 7.1 Active personas (4)
|
||||
|
||||
**Finding (0.85):** v0.5 is **voice-pipeline-heavy (wake-word + assist mode + latency tuning) + safety-critical guardrails + cohort aggregation extension**. The roster:
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: lead-developer
|
||||
active: true
|
||||
phase_specific: false
|
||||
reason: Coordinates across assist pipeline, guardrails, context-binding, and aggregation domains. Owns the build_assist_pipeline() design decision (whether to add a mode param to build_pipeline or a separate builder) and the warm-WebRTC-connection lifecycle. Required for every milestone.
|
||||
domain: coordination
|
||||
frameworks: [pipecat, fastapi, sqlite, postgres, webrtc]
|
||||
constraints: [pragmatic, latency-budget-aware, hybrid-storage-no-cross-db-joins, k-anonymity-floor-10, assist-does-not-affect-mastery]
|
||||
territory:
|
||||
- "docker-compose.yml"
|
||||
- ".env.example"
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: voice-engineer
|
||||
active: true
|
||||
phase_specific: true
|
||||
reason: REACTIVATED for v0.5 (proposed at PERSONAS.md line 458 for v0.5+). Owns the wake-word client (Picovoice Porcupine Android foreground service), the assist audio pipeline (warm WebRTC connection, wake-word → first-audio latency), latency tuning (the <600ms assist budget — Domain 3), and the in-loop guardrail processor (post-LLM frame processor). This is the largest new territory in v0.5: the assist voice loop is a new mode alongside the practice scenario loop. Will deactivate in v0.6 unless voice work continues (accent modeling, multi-voice personas).
|
||||
domain: voice
|
||||
frameworks: [porcupine-android, webrtc, silero-vad, pipecat, audio-codecs, piper-tts]
|
||||
constraints: [sub-600ms-latency-assist, warm-webrtc-connection, foreground-service-background-mic, wake-word-detection-latency, piper-tts-for-assist, lean-assist-system-prompt]
|
||||
territory:
|
||||
- "**/server/pipeline.py"
|
||||
- "**/server/asr/**"
|
||||
- "**/server/tts/**"
|
||||
- "**/server/latency.py"
|
||||
- "**/client/wake-word/**"
|
||||
- "**/client/assist-service/**"
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: backend-engineer
|
||||
active: true
|
||||
phase_specific: false
|
||||
reason: Owns the context-binding endpoints (load path week + scenario tag + learner state into the assist prompt), the assist session API (start_shift / end_shift / log_assist_turn), the SessionRecorder extension (session_type field, assist turn logging, _build_session_outcome assist branch), and the cohort hook extension for session_type='assist'. Also owns the LiveAssistGuardrail ruleset (with security-engineer). The assist session API + context-binding is the largest backend territory in v0.5.
|
||||
domain: backend
|
||||
frameworks: [pipecat, pydantic, fastapi, uvicorn, aiosqlite, asyncpg]
|
||||
constraints: [api-first, type-safe, mastery-off-voice-path, aggregation-off-voice-path, latency-budget-aware, no-cross-db-joins, assist-does-not-update-mastery]
|
||||
territory:
|
||||
- "**/server/**"
|
||||
- "**/server/guardrails/**"
|
||||
- "**/server/cohort/**"
|
||||
- "**/server/session_recorder.py"
|
||||
- "**/server/assist/**"
|
||||
- "**/db/migrations/**"
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: security-engineer
|
||||
active: true
|
||||
phase_specific: true
|
||||
reason: RETAINED from v0.4. Owns the LiveAssistGuardrail enforcement (REQ-ASSIST-03 — safety-critical: the AI is in the learner's ear during real customer interactions). The 3-layer guardrail (D-060) is the security-engineer's v0.5 surface: prompt rules, output filter patterns (direct-answer vs coaching-question regex), audit logging, and the guardrail_block_rate safety signal. Also owns the privacy/consent disclosure surface (the foreground-service notification + learner-facing "Assist is on — those around you may be recorded" disclosure). REQ-ASSIST-03 is the most safety-critical requirement in v0.5; the security-engineer's guardrail work blocks ship.
|
||||
domain: security
|
||||
frameworks: [pynacl, canonicaljson, base58, argon2-cffi, regex, llm-guardrail-patterns]
|
||||
constraints: [coaches-not-does, no-direct-answer-patterns, no-false-authority, no-impersonation, audit-all-assist-turns, guardrail-block-rate-operator-visible, consent-disclosure-required]
|
||||
territory:
|
||||
- "**/server/guardrails/**"
|
||||
- "**/server/guardrails/live_assist.py"
|
||||
- "**/server/vc/**" # retained from v0.4 (no v0.5 change expected)
|
||||
- "**/server/auth/**" # retained from v0.4 (no v0.5 change expected)
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: data-engineer
|
||||
active: true
|
||||
phase_specific: false
|
||||
reason: RETAINED from v0.4. Owns the assist aggregation integration into the v0.4 cohort pipeline (new assist metrics in cohort_aggregates — no schema change, new metric strings), the turns-table guardrail_verdict field migration (SQLite, additive), and the assist session row in the sessions table (session_type field). Also owns the k-anonymity suppression extension for assist metrics (assist_active_learners_count distinct-count). Smaller v0.5 surface than v0.4 but on the critical path for operator visibility.
|
||||
domain: data
|
||||
frameworks: [sqlite, postgres16, aiosqlite, asyncpg]
|
||||
constraints: [schema-first, migration-driven, no-cross-db-joins, k-anonymity-floor-10, opaque-learner-ref, write-time-suppression, assist-metrics-no-schema-change]
|
||||
territory:
|
||||
- "**/db/**"
|
||||
- "**/db/migrations/**"
|
||||
- "**/server/cohort/aggregator.py"
|
||||
---
|
||||
```
|
||||
|
||||
### 7.2 Deactivated personas (2)
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: devops-engineer
|
||||
active: false
|
||||
phase_specific: true
|
||||
reason: DEACTIVATED for v0.5. No deploy changes — v0.4's LXC + Docker-in-LXC + Postgres carries forward unchanged. The assist foreground service is a client-side concern (voice-engineer territory), not a deploy/infra change. No new Docker services, no CT resource bump, no new backup scripts. Will reactivate in v0.6+ if deploy hardening (TLS, multi-instance, autoscaling) or a CT bump is needed for assist concurrency.
|
||||
domain: devops
|
||||
frameworks: [proxmox-lxc, docker, systemd, bash]
|
||||
constraints: [idempotent-deploy, secrets-never-committed]
|
||||
territory: []
|
||||
---
|
||||
```
|
||||
|
||||
```yaml
|
||||
---
|
||||
name: frontend-engineer
|
||||
active: false
|
||||
phase_specific: true
|
||||
reason: DEACTIVATED for v0.5 (PROVISIONAL — see note). v0.5 assist mode is invoked by wake-word (audio) — the UI surface is minimal: a "Start Shift" / "End Shift" toggle + a context-declaration screen (path week + scenario tag selector). This is small enough that the voice-engineer (client/wake-word + client/assist-service) can own it alongside the audio pipeline, OR the backend-engineer can add a minimal React route. No full frontend surface (no new dashboard, no complex components, no chart library). Will reactivate in v0.6+ if a richer assist control surface (shift history, guardrail-block review, assist coaching quality dashboard) is needed. NOTE FOR ORCHESTRATOR: if the assist control surface (start/stop shift + context declaration) is judged non-trivial (>200 LOC of React), reactivate frontend-engineer. Current estimate: ~100-150 LOC of React — below the reactivation threshold.
|
||||
domain: frontend
|
||||
frameworks: [react, react-router-dom, pipecat-client-sdk, webrtc]
|
||||
constraints: [component-first, voice-first-ui, minimal-client-javascript]
|
||||
territory: []
|
||||
---
|
||||
```
|
||||
|
||||
### 7.3 Roster decision summary
|
||||
|
||||
| Persona | v0.4 status | v0.5 status | Reason |
|
||||
|---------|-------------|-------------|--------|
|
||||
| lead-developer | active | **active** | Coordination across assist/guardrail/aggregation |
|
||||
| voice-engineer | proposed (inactive) | **active (REACTIVATED)** | Wake-word client, assist pipeline, latency tuning — the largest v0.5 surface |
|
||||
| backend-engineer | active | **active (retained)** | Context-binding, assist session API, SessionRecorder extension, cohort hook |
|
||||
| security-engineer | active | **active (retained)** | REQ-ASSIST-03 guardrails — safety-critical |
|
||||
| data-engineer | active | **active (retained)** | Assist aggregation integration (no schema change, new metrics) |
|
||||
| devops-engineer | active | **deactivated** | No deploy changes in v0.5 |
|
||||
| frontend-engineer | active | **deactivated (provisional)** | Minimal assist UI; reactivate if control surface exceeds ~200 LOC |
|
||||
|
||||
**4 active personas + 1 reactivation (voice-engineer) = 5 active, 2 deactivated.** This is the right size for v0.5's scope (voice + guardrails + aggregation, no deploy, minimal UI).
|
||||
|
||||
### 7.4 Constraint alignment (v0.5-specific)
|
||||
|
||||
- **All personas:** `assist-does-not-affect-mastery` (D-063), `k-anonymity-floor-10` (D-034 carry-forward), `no-raw-learner-pii-in-postgres` (D-031 carry-forward).
|
||||
- **lead-developer:** `latency-budget-aware` (C-8 — the binding constraint for assist), `hybrid-storage-no-cross-db-joins` (D-031).
|
||||
- **voice-engineer:** `sub-600ms-latency-assist` (C-8 for assist turns), `warm-webrtc-connection` (shift-bounded, not per-turn), `foreground-service-background-mic` (Android requirement), `wake-word-detection-latency` (Porcupine ~200-500ms), `piper-tts-for-assist` (R4 mitigation as default for assist), `lean-assist-system-prompt` (≤150 tokens for prefill latency).
|
||||
- **backend-engineer:** `mastery-off-voice-path` (C-8 carry-forward), `aggregation-off-voice-path` (D-054 carry-forward), `assist-does-not-update-mastery` (D-063 — the `schedule_mastery=False` gate on assist shifts).
|
||||
- **security-engineer:** `coaches-not-does` (REQ-ASSIST-03), `no-direct-answer-patterns` (output filter regex), `no-false-authority`, `no-impersonation`, `audit-all-assist-turns` (turns table + guardrail_verdict), `guardrail-block-rate-operator-visible` (cohort aggregation safety signal), `consent-disclosure-required` (foreground-service notification).
|
||||
- **data-engineer:** `assist-metrics-no-schema-change` (new metric strings in cohort_aggregates, no DDL), `write-time-suppression` (D-034 carry-forward).
|
||||
|
||||
---
|
||||
|
||||
## Consolidated Risks Table
|
||||
|
||||
| ID | Risk | Severity | Mitigation | Confidence |
|
||||
|----|------|----------|------------|------------|
|
||||
| **R-ASSIST-01** | Picovoice Porcupine MAU pricing blocks the pilot (no recurring free tier — verified) | **high** | Engage Picovoice sales for a pilot/educational tier; fallback to a built-in wake word (e.g., "Bumblebee") for v0.5; document Vosk as the open-source fallback | 0.75 |
|
||||
| **R-ASSIST-02** | C-8 <600ms latency budget broken for assist turns (estimated ~655-770ms) | **high** | Lean assist system prompt (≤150 tokens) + Piper TTS (not Cartesia) for assist + measure R1/R3 in Phase 1; accept ~650ms for pilot if trending down; flag orchestrator to relax C-8 for assist or push hardening to v0.6 | 0.70 |
|
||||
| **R-ASSIST-03** | Wake-word → first-audio latency ~850-1150ms (warm) / unacceptable (cold) | medium | Require warm WebRTC connection for the shift (foreground service keepalive); document the ~1s wake-word-to-coaching latency as expected (not the in-conversation <600ms budget) | 0.75 |
|
||||
| **R-ASSIST-04** | Android background-mic restriction (Android 14+ foreground-service-microphone type) | medium | Use a foreground service of type `microphone` with persistent notification; document OEM battery-kill whitelist step for learners | 0.70 |
|
||||
| **R-ASSIST-05** | OEM battery kill switches (Xiaomi/Huawei/OnePlus) kill the assist foreground service | medium | Document the "battery whitelist" onboarding step; test on the target $100 Android device; consider a "survival mode" that restarts the service on kill (Android `START_STICKY`) | 0.65 |
|
||||
| **R-ASSIST-06** | Output filter false positives block legitimate coaching (regex over-matches) | medium | Tune the direct-answer regex against a corpus of real coaching responses in Phase 1; allow one retry on block; fall back to a canned coaching redirect | 0.75 |
|
||||
| **R-ASSIST-07** | Output filter false negatives let a direct answer through (regex under-matches) | **high** | Defense-in-depth: layer 1 prompt rules + layer 2 regex + (post-v0.5) LLM-as-judge. The regex is the first line, not the only line. Audit all turns + guardrail_block_rate surfaces misses to operators. | 0.70 |
|
||||
| **R-ASSIST-08** | Privacy/consent: ambient mic records the real customer without their consent | **high** | Foreground-service notification ("Praxis Assist is on") + learner-facing disclosure ("those around you may be recorded by your mic"). Legal review of one-party/two-party consent law for Canada. **Flag for orchestrator — this is a legal/ethical surface, not purely technical.** | 0.60 |
|
||||
| **R-ASSIST-09** | Warm WebRTC connection dropped mid-shift (NAT timeout, network change) | medium | App-level heartbeat every 30s; auto-reconnect on drop; log the reconnection; if reconnection fails, prompt learner to restart shift | 0.75 |
|
||||
| **R-ASSIST-10** | Server concurrency: long-lived assist WebRTC tasks exhaust the asyncpg pool / uvicorn capacity | low (pilot) | Pilot: single-learner (D-007), ≤1 concurrent assist shift. Post-pilot: v0.6+ concurrency hardening (multi-uvicorn, larger pool). | 0.80 |
|
||||
| **R-ASSIST-11** | Assist shifts abandoned (learner forgets "ending shift") → orphaned WebRTC connections + stale sessions | medium | Auto-end shift after 8h (configurable); foreground service timeout; log abandoned shifts in cohort aggregation (assist_shifts_count separates completed vs abandoned) | 0.75 |
|
||||
| **R-ASSIST-12** | Context-binding reads stale learner state (learner advanced a week but assist uses old week) | low | Learner declares context at shift start (D-059); server reads `progress.current_week` fresh from SQLite at shift start; if the learner advanced mid-shift, the next shift picks up the new week | 0.80 |
|
||||
| **R-ASSIST-13** | Porcupine wake-word false triggers in noisy retail environment | medium | Choose a wake word with diverse phonemes + ≥6 phonemes (Porcupine FAQ guidance); "Bumblebee" / "Grapefruit" / custom "Hey Praxis" tuned via Console; tune sensitivity (Porcupine has a sensitivity parameter) | 0.70 |
|
||||
| **R-ASSIST-14** | Assist foreground service battery drain + learner's other work apps → phone dies mid-shift | medium | Document expected drain (~4-9% per shift); tap-to-talk fallback (no wake-word listener) for battery-saving mode; learner can stop assist if battery < 20% | 0.65 |
|
||||
|
||||
---
|
||||
|
||||
## D-058..D-063 Validation Audit
|
||||
|
||||
| CLARIFY Decision | Validation | Verdict |
|
||||
|------------------|------------|---------|
|
||||
| **D-058** (Porcupine wake-word + tap-to-talk fallback) | Porcupine verified (on-device, offline, low-power, Android SDK, custom WW). **MAU pricing / no recurring free tier — partial contradiction.** Refinement: pursue Picovoice sales pilot tier, fallback to built-in wake word, document Vosk. | **REFINED** — wake-word engine confirmed; free-tier assumption contradicted |
|
||||
| **D-059** (Learner declares context + server reads SQLite path week) | Confirmed. `PraxisStore.get_progress()` returns `current_week`. Auto-detection impossible (C-4). Declaration UI is small. | **CONFIRMED** |
|
||||
| **D-060** (3-layer guardrail: prompt rules + output filter + audit log) | Confirmed — industry-standard pattern. Existing `CustomerServiceGuardrail` proves the regex output-filter approach. v0.5 adds LiveAssistGuardrail with direct-answer vs coaching-question patterns. | **CONFIRMED** |
|
||||
| **D-061** (<600ms latency, shared pipeline, ≤30s assist turns) | **At risk.** Estimated assist latency ~655-770ms (all-cloud) / ~655ms (Piper + lean prompt). C-8 is the binding constraint. Mitigations identified but may not fully close the gap. **Flag for orchestrator.** | **AT RISK** — likely ~50-170ms over budget; levers identified |
|
||||
| **D-062** (Shift-bounded sessions, session_type=assist in cohort aggregation) | Confirmed. Shift-bounded matches real CS work. No schema change to cohort_aggregates (new metric strings). on-session-end hook extended. | **CONFIRMED** |
|
||||
| **D-063** (Assist does not update mastery or count toward gates) | Confirmed. `SessionRecorder.end(schedule_mastery=False)` for assist shifts. The mastery flow is practice-only. | **CONFIRMED** |
|
||||
|
||||
**Summary:** 4 confirmed, 1 refined (D-058 free-tier), 1 at-risk (D-061 latency). Two items flagged for orchestrator attention: the Picovoice pricing path (R-ASSIST-01) and the C-8 latency tension for assist mode (R-ASSIST-02 / D-061).
|
||||
|
||||
---
|
||||
|
||||
## New Decisions (D-064+)
|
||||
|
||||
| ID | Decision | Rationale | Confidence | Alternatives |
|
||||
|----|----------|-----------|------------|--------------|
|
||||
| **D-064** | Live Assist wake-word engine = **Picovoice Porcupine (built-in wake word for v0.5 pilot; custom "Hey Praxis" post-pilot)**, with **Vosk as the documented open-source fallback** | R-ASSIST-01: Porcupine MAU pricing has no recurring free tier. v0.5 ships with a built-in Porcupine wake word (e.g., "Bumblebee") to avoid custom-training costs during the pilot. Post-pilot, engage Picovoice sales for a custom "Hey Praxis" wake word under a pilot/educational tier. Vosk (Apache 2.0, offline) is the fallback if Porcupice pricing is unsustainable. Snowboy rejected (deprecated). | 0.70 | Vosk for v0.5 (free but heavier), TFLite DIY (engineering effort), Snowboy (deprecated) |
|
||||
| **D-065** | Live Assist TTS = **Piper (self-hosted on pilot server) as the default for assist turns**, Cartesia as the quality fallback for practice mode | R-ASSIST-02: assist turns are latency-critical (C-8). Piper ~80ms first audio vs Cartesia ~120ms. The v0.1 R4 mitigation pre-stages Piper; v0.5 assist mode defaults to Piper to claw back ~40ms toward the <600ms budget. Practice mode retains Cartesia (quality over latency for practice). | 0.75 | Cartesia for both (simpler, but +40ms on assist), Piper for both (lower quality for practice) |
|
||||
| **D-066** | Live Assist system prompt = **≤150 input tokens** (coaching instruction ~80 tokens + context-binding ~50 tokens + voice-conciseness ~20 tokens) | R-ASSIST-02: extra input tokens add prefill latency (~0.5ms/token). A lean prompt keeps the prefill delta under 50ms vs v0.1 practice. Avoid dumping the full rubric or scenario YAML into the prompt — context-binding is terse (path week, scenario tag, one-line coaching focus). | 0.78 | Verbose prompt (easier coaching quality, but +100-200ms latency) |
|
||||
| **D-067** | Live Assist WebRTC connection = **warm for the entire shift** (foreground service keepalive; not per-turn cold connect) | R-ASSIST-03: cold WebRTC connect (~500-1000ms) is unacceptable for live assist. The assist foreground service opens a warm connection at shift start, keeps it alive (heartbeat every 30s), and reuses it for every assist turn. Closed at shift-end. Between turns, only keepalive flows (no audio streaming) to save battery. | 0.78 | Per-turn cold connect (too slow), always-streaming (battery + privacy) |
|
||||
| **D-068** | Live Assist guardrail output filter = **regex-based direct-answer + false-authority + impersonation patterns, with one retry on block + canned coaching redirect fallback** | R-ASSIST-06/07: regex is the fast on-voice-path filter (matches the existing CustomerServiceGuardrail pattern). One retry gives the LLM a chance to self-correct; the canned fallback ensures a safe response if the retry also blocks. LLM-as-judge deferred to post-v0.5 (off-voice-path, more accurate, nightly). | 0.78 | LLM-as-judge on-voice-path (too slow for <600ms), no filter (unsafe) |
|
||||
| **D-069** | Live Assist shift = **auto-end after 8 hours** (configurable via `PRAXIS_ASSIST_MAX_SHIFT_HOURS=8`) | R-ASSIST-11: learners may forget "ending shift", leaving orphaned WebRTC connections + stale sessions. Auto-end after 8h (a typical shift length) closes the shift cleanly, fires the aggregation hook, and releases the foreground service. The learner can restart a new shift if needed. | 0.75 | No auto-end (orphan risk), shorter (4h — too short for some shifts), longer (12h — battery risk) |
|
||||
| **D-070** | Live Assist consent disclosure = **foreground-service notification + learner-facing "Assist is on — those around you may be recorded by your mic" disclosure at shift start** | R-ASSIST-08: the ambient mic may pick up the real customer. Ethical and legal (one-party/two-party consent law) requires disclosure. The foreground service notification (Android requirement) + an in-app disclosure at shift start covers the learner's awareness. The customer's consent is the learner's responsibility (Praxis can't notify the customer). **Flag for orchestrator: legal review of Canada consent law for ambient recording during coaching.** | 0.65 | No disclosure (legal/ethical risk), explicit customer consent prompt (impractical — the customer isn't a Praxis user) |
|
||||
|
||||
---
|
||||
|
||||
## New pip dependencies for v0.5
|
||||
|
||||
| Dep | Purpose | Confidence | Source |
|
||||
|-----|---------|------------|--------|
|
||||
| (none new server-side) | The v0.1 voice pipeline (Pipecat + Deepgram + Cartesia + Piper + Ollama) is reused unchanged. The guardrail is pure-Python regex (no new dep). The aggregation extension uses existing asyncpg. | 0.90 | Domain 5 + 6 |
|
||||
|
||||
**Picovoice Porcupine SDK** is an **Android client-side** dependency (Gradle/Maven), not a Python server-side dep. The praxis server doesn't run Porcupine — the learner's phone does. The server-side assist code is pure Python (FastAPI + Pipecat + aiosqlite + asyncpg, all existing).
|
||||
|
||||
## New npm/Gradle dependencies for v0.5
|
||||
|
||||
| Dep | Side | Purpose | Confidence | Source |
|
||||
|-----|------|---------|------------|--------|
|
||||
| `ai.picovoice:porcupine-android` (Gradle) | Client (Android) | Wake-word detection on the learner's phone | 0.80 | D-058, D-064 |
|
||||
|
||||
**Note:** the v0.1 client is React + WebRTC (D-015), not React Native. The Porcupine React SDK exists but runs in-browser (not a foreground service). For true background wake-word on Android, v0.5 may need a **React Native** or **native Android** client — this is a client-architecture decision for the orchestrator. The v0.1 RESEARCH (D-015) noted "upgrades to React Native for Android later." v0.5 Live Assist (phone-in-pocket, background mic) likely **is** the trigger to upgrade to React Native. **Flag for orchestrator: v0.5 may require a client-architecture upgrade from React-Web to React-Native (or a native Android assist service alongside the React web app).** This is a significant scope addition.
|
||||
|
||||
---
|
||||
|
||||
## Open Questions for PLAN Stage
|
||||
|
||||
1. **Client architecture for v0.5:** React web (v0.1, D-015) can't do background wake-word on Android (no foreground service). Options: (a) upgrade the client to React Native (Porcupine RN SDK + Android foreground service), (b) ship a separate native Android "Praxis Assist" app alongside the React web practice app, (c) defer wake-word to v0.6 and ship v0.5 assist as tap-to-talk only (no wake-word). **Recommendation: (c) for v0.5 pilot — tap-to-talk is hands-free enough for a pilot (learner taps a button on a smartwatch or a headset button), and it avoids the React-Native upgrade scope. Add wake-word in v0.6 with the native client.** This would defer D-058/D-064 to v0.6 and simplify v0.5 to the assist voice loop + guardrails + aggregation only. **Flag for orchestrator — this is a scope decision.**
|
||||
|
||||
2. **Picovoice sales engagement:** When to engage Picovoice sales for the pilot/educational tier? Before v0.5 PLAN, or after v0.5 ships with tap-to-talk? If wake-word is deferred to v0.6 (per Q1), the sales engagement is a v0.6 activity.
|
||||
|
||||
3. **Lean assist system prompt — concrete content:** The ≤150-token budget (D-066) is a constraint; the concrete prompt content (the coaching instruction phrasing, the context-binding format) needs Phase-1 iteration + red-team testing. What's the minimum prompt that produces coaching questions, not direct answers, from `gemma4:cloud`?
|
||||
|
||||
4. **Output filter regex corpus:** The direct-answer regex (D-068) needs tuning against a corpus of real coaching responses. How to build this corpus before v0.5 ships? Option: generate a synthetic corpus via LLM (prompt `gemma4:cloud` to produce coaching responses + direct-answer responses, label them, tune the regex). Phase-1 task.
|
||||
|
||||
5. **Assist shift vs practice session — can they coexist?** Can a learner be in a practice session (WebRTC to praxis) and invoke assist (warm WebRTC to praxis) simultaneously? Probably not for v0.5 (one WebRTC connection at a time per D-007 single-learner). The learner ends the practice session before starting an assist shift, or vice versa. Document the mutual exclusivity.
|
||||
|
||||
6. **Guardrail verdict storage:** A `guardrail_verdicts` table (keyed by turn id) or a JSON column on `turns`? A JSON column is simpler (additive migration); a separate table is more queryable for the operator dashboard. Recommend JSON column for v0.5 (simpler); separate table if the operator dashboard needs to filter/sort by verdict.
|
||||
|
||||
7. **Phase split confirmation:** ROADMAP P1 = assist voice loop (pipeline + guardrail + context-binding) + aggregation extension; P2 = guardrail tuning + latency measurement + operator dashboard assist views; P3 = review. Is the aggregation extension P1 or P2? Recommend P2 (the assist voice loop is the P1 deliverable; aggregation is operator-facing, P2).
|
||||
|
||||
8. **Canada consent law for ambient recording:** R-ASSIST-08 / D-070. Canada's Personal Information Protection and Electronic Documents Act (PIPEDA) + provincial one-party/two-party consent recording laws. Praxis assist records the learner (one party — the learner consents by starting the shift) but may pick up the customer (the other party). One-party consent (Canada is one-party consent federally) means the learner can record their own conversation without the customer's consent. **But** the AI analyzing the customer's speech in real-time is a novel use. **Flag for orchestrator — legal review recommended before v0.5 ship.** Confidence 0.60 (not legal advice).
|
||||
+32
-12
@@ -1,18 +1,37 @@
|
||||
# Praxis — Roadmap
|
||||
|
||||
**Milestone:** v0.4 (Operator tier — cohort dashboard, auth, Postgres) — active
|
||||
**Status:** phase 3 — final review (active milestone); P0-P2 complete (v0.1.6/v0.1.7/v0.1.8 tagged)
|
||||
**Previous milestone:** v0.3 (Mastery scoring + competency rubrics + verifiable credentials) — complete, tagged v0.1.5, release #380, merged to main
|
||||
**Milestone:** v0.5 (Live Assist — on-the-job voice companion) — active, phase 0 pre-execution
|
||||
**Status:** phase 0 pre-execution (SPECIFY → CLARIFY → RESEARCH → IDEATE → PLAN → GRILL → SHIP)
|
||||
**Previous milestone:** v0.4 (Operator tier — cohort dashboard, auth, Postgres) — complete, tagged v0.1.9, release created, merged to main
|
||||
|
||||
## Milestone Philosophy
|
||||
|
||||
v0.4 activates the operator tier deferred from v0.3 per the grill's binding verdict (GRILL-v0.3.md Axis 2 — the operator tier was originally v0.8 on this roadmap; pulling it into v0.3 created a 2-milestone program disguised as one). v0.4 layers the operator surface on top of the v0.3 mastery/VC/scenario work: a Postgres store in the existing LXC CT, operator auth (argon2id session cookies), a cohort aggregation pipeline (k-anonymity ≥ 10, 7-day windows), and a React cohort dashboard served by the same FastAPI server. The learner-facing surface carries forward unchanged (SQLite, voice loop, mastery gates, VC issuance). The VC issuer key store migrates from SQLite to operator-tier Postgres + secrets (D-042).
|
||||
v0.5 activates the Live Assist surface deferred from v0.1 (originally listed in v0.1 out-of-scope: "Live Assist mode"). v0.1–v0.4 built and validated the **practice surface** — learners practice scenarios with AI tutors, scored against rubrics, progress via mastery gates, with a v0.4 operator tier observing cohort patterns. v0.5 adds the **companion surface**: a hands-free voice assistant a learner invokes *while actually working* on the job, context-aware of their current scenario/skill path, coaching in real time without doing the job for them.
|
||||
|
||||
## v0.4 Phases
|
||||
The key distinction from the practice surface is **real-customer interaction**: in v0.1–v0.4, the learner role-plays with an AI; in v0.5, the learner is on a real call with a real customer and the AI is in their ear. This makes REQ-ASSIST-03 (guardrails: coaches not does; never lies to real customers) the safety-critical requirement. The v0.1 voice pipeline (Pipecat + Deepgram Nova-3 + Cartesia + Ollama Cloud) carries forward, reused in a new "assist" mode distinct from the practice scenario loop. Learner state stays in SQLite (D-007 preserved); Live Assist reads the learner's active path week (D-037) for context-binding.
|
||||
|
||||
### Phase 0 — Pre-Execution (complete — tagged v0.1.6, release created)
|
||||
## v0.5 Phases
|
||||
|
||||
### Phase 0 — Pre-Execution (active)
|
||||
|
||||
**Branch:** `phase/00-pre-execution` → `milestone/v0.5-live-assist`
|
||||
**Ship target:** `v0.1.10` (next available patch on the v0.1.x line — NFR/docs milestone type)
|
||||
**Status:** active (SPECIFY complete → CLARIFY → RESEARCH → IDEATE → PLAN → GRILL → SHIP)
|
||||
|
||||
Pipeline stages: SPECIFY → CLARIFY → RESEARCH → **IDEATE** (--ideate flag) → PLAN → GRILL → SHIP
|
||||
|
||||
**Goal:** Produce all `.ciagent/` planning artifacts for v0.5: activated requirements (REQ-ASSIST-01/02/03 + 4 NFRs), research-grounded Live Assist architecture (invocation model, context-binding, guardrail enforcement, latency budget), ideation-driven improvements, persona roster (likely reactivates voice-engineer per PERSONAS.md note "PROPOSED for v0.5+"), vertical-slice plan for P1.
|
||||
|
||||
**Deliverables:**
|
||||
- PROJECT.md (v0.5 scope validated; Live Assist activated)
|
||||
- REQUIREMENTS.md (v0.5 active REQ-IDs = 3 + 4 NFRs; v0.4 marked complete)
|
||||
- ARCHITECTURE.md (Live Assist mode added to v0.4 topology — assist voice loop + context-binding + guardrail extension)
|
||||
- PERSONAS.md (v0.5 roster — voice-engineer reactivated for hands-free/latency; backend-engineer for context-binding + guardrails; security-engineer retained for REQ-ASSIST-03 safety surface)
|
||||
- GRILL-v0.5.md (adversarial review — real-customer interaction warrants grill)
|
||||
- Phase 1 plan (vertical slices with wave ordering)
|
||||
|
||||
## v0.4 Milestone (complete — reference)
|
||||
|
||||
**Branch:** `phase/00-pre-execution` → merged to `milestone/v0.4-operator-tier`
|
||||
**Ship target:** `v0.1.6` (patch release on v0.3's v0.1.x line — NFR/docs milestone type)
|
||||
**Status:** complete (v0.1.6 tagged, Gitea release created)
|
||||
|
||||
@@ -44,11 +63,11 @@ Pipeline stages: SPECIFY → CLARIFY → RESEARCH → PLAN → GRILL → SHIP
|
||||
|
||||
**Goal:** Cohort aggregation pipeline (on-session-end hook + nightly reconciliation, k-anonymity ≥ 10, 7-day windows) + React cohort dashboard under `/operator/*` (served by same FastAPI, reuses v0.2 StaticFiles) + `/api/operator/*` endpoints (auth-gated). Dashboard shows anonymized practice/mastery/failure-pattern views with cells < 10 learners suppressed.
|
||||
|
||||
### Final Phase (P3) — Review + Ship (planned)
|
||||
### Final Phase (P3) — Review + Ship (complete — tagged v0.1.9, release created, merged to main)
|
||||
|
||||
**Branch:** `phase/03-final-review-ship` → merged to `milestone/v0.4-operator-tier` → merged to `main`
|
||||
**Ship target:** final patch = v0.4 milestone release
|
||||
**Status:** planned
|
||||
**Status:** complete (v0.1.9 tagged, Gitea release created, merged to main; review APPROVE_WITH_NOTES, audit HEALTHY)
|
||||
|
||||
**Goal:** Multi-persona code review, project audit, milestone merge to main, milestone release.
|
||||
|
||||
@@ -127,15 +146,16 @@ Pipeline stages: SPECIFY → CLARIFY → RESEARCH → PLAN → GRILL
|
||||
|
||||
v0.1 was the **foundation milestone** — minimal viable voice loop (one persona, one scenario, ASR+TTS+LLM round-trip, single learner state). Shipped as `v0.0.0` (phase 0) → `v0.0.1` (phase 1) → `v0.0.2` (final/milestone release).
|
||||
|
||||
## Future Milestones (post-v0.4, indicative)
|
||||
## Future Milestones (post-v0.5, indicative — refined by v0.5 IDEATE)
|
||||
|
||||
| Milestone | Scope (indicative) |
|
||||
|-----------|-------------------|
|
||||
| v0.5 | Live Assist on-the-job companion |
|
||||
| v0.6 | Low-bandwidth surfaces (WhatsApp, offline cache) |
|
||||
| v0.6 | Low-bandwidth surfaces (WhatsApp, offline cache) + IDEATE-10 (LLM-as-judge guardrail eval) + IDEATE-11 (assist-weaning metric) + IDEATE-12 (offline assist degraded mode) + IDEATE-13 (voice-only context declaration) |
|
||||
| v0.7 | Multi-language (French-Canadian, then PRD's 10-language list) |
|
||||
| v0.8 | Full operator-suite dashboard (REQ-DASH-02 — beyond v0.4's foundational cohort view) |
|
||||
| v0.9 | Credentialing (third-party verifiable, shareable) |
|
||||
| v1.0 | Working, tested product — multiple paths, multi-market, production-ready |
|
||||
|
||||
_The v0.6 row now includes 4 ideation-derived requirements (REQ-IDEATE-10..13) accepted during the v0.5 IDEATE stage. These will be refined by ci-roadmapper at the start of the v0.6 milestone._
|
||||
|
||||
These are indicative and will be refined by ci-roadmapper at the start of each milestone.
|
||||
@@ -3,8 +3,8 @@
|
||||
{
|
||||
"slug": "praxis",
|
||||
"name": "Praxis",
|
||||
"milestone": "v0.4",
|
||||
"status": "active"
|
||||
"milestone": "v0.5",
|
||||
"status": "phase-0-active"
|
||||
}
|
||||
],
|
||||
"active_project": "praxis",
|
||||
|
||||
Reference in New Issue
Block a user