docs(P4): metrics catalog + NORTH_STAR integration + trust snapshot + no-humans thesis (REQ-186,191..195,204,210..213)

P4 (Wave 3, docs) — REQ-186, 191, 192, 193, 194, 195, 204, 210, 211, 212, 213

New docs:
- docs/METRICS.md — canonical KPI catalog (grounded/derived/deferred)
- docs/metrics/*.md — 13 per-KPI definition-of-success docs (D-127)
- docs/METRICS_DEFERRED_ROADMAP.md — 8 deferred metrics + hot-path plan + re-eval triggers (REQ-210)
- docs/NO_HUMANS_THESIS.md — thesis defensibility brief (REQ-213)

New tools:
- core/metrics/trust_snapshot.py — 5 trust metrics + chain-integrity verdict + snapshot hash (REQ-211)
- scripts/check_north_star_diff.sh — CI check for NORTH_STAR strategic section changes (REQ-204)

Modified:
- .ciagent/config.json — strategic_direction_file: .ciagent/NORTH_STAR.md (REQ-186)

---ci---
project: acdl
phase: 4
milestone: v1.17
status: execute
---/ci---

docs/metrics/ai_decision_accuracy.md

@@ -0,0 +1,21 @@
+# AI Decision Accuracy — Definition of Success
+
+> KPI: AI Decision Accuracy
+> Target: ≥ 99.5% (no rollback, no follow-up incident within 5 min of action)
+
+**What this number means:** the percentage of AI decisions (confidence-
+gated policy engine outcomes) that were NOT followed by an apply failure
+or incident within 5 minutes. A high-confidence decision that later
+caused an incident does NOT count as accurate.
+
+**How it's computed:** `count(decisions WHERE outcome = 'succeeded' AND
+no incident within 5min)` ÷ `total decisions`. Correlation via
+`decision_id` → `run_id` → subsequent `apply.failed` or `incident.detected`
+events.
+
+**What "good" looks like:** ≥ 99.5% means fewer than 1 in 200 decisions
+cause a secondary failure. The 0.5% allowance is for novel edge cases.
+
+**D-122 honesty:** Nova's "AI" is the confidence-gated policy engine
+(confidence_signal + HITL gate), not an LLM planner. The Decision Ledger
+captures this real decision path — not a fabricated "AI agent."

docs/metrics/attestation_coverage.md

@@ -0,0 +1,18 @@
+# Attestation Coverage — Definition of Success
+
+> KPI: Attestation Coverage
+> Target: 100% of prod/dr promotions attested by a human
+
+**What this number means:** every production and disaster-recovery
+promotion has a recorded human attestation (approver identity, 8-concern
+matrix result, separation-of-duties check on prod). This is the
+"autonomy in operations, human in accountability" proof.
+
+**How it's computed:** `count(prod/dr promotions with attestation.recorded
+event) ÷ count(total prod/dr promotions)`. Sourced from the Decision
+Ledger (`attestation.recorded` events) + `hitl_gates.py` + outbox
+`approver_*` attributes.
+
+**What "good" looks like:** 100% means no prod/dr promotion ever lands
+without a human sign-off on record. The absence of an operator is never
+the absence of a record (NORTH_STAR Anti-Goal #3).

docs/metrics/confidence_gate_halt_rate.md

@@ -0,0 +1,16 @@
+# Confidence-Gate Halt Rate — Definition of Success
+
+> KPI: Confidence-Gate Halt Rate
+> Target: not a committed target (operational signal)
+
+**What this number means:** how often the confidence gate itself halted
+a run (band = block), independent of HITL blocks. The gate is the AI's
+self-halt; HITL is the human gate. This distinguishes the AI's
+self-regulation from human escalation.
+
+**How it's computed:** `count(runs WHERE confidence_band = 'block')` ÷
+`total runs`.
+
+**What "good" looks like:** a low but non-zero rate means the gate is
+working (catching genuinely uncertain runs) without being overly
+conservative (blocking everything).

docs/metrics/cost_savings.md

@@ -0,0 +1,18 @@
+# Cost Savings via Infracost Estimates — Definition of Success
+
+> KPI: Cost Savings via Infracost Estimates
+> Target: ≥ 25% on pilot estates (partial)
+
+**What this number means:** the pre-apply cost estimate from Infracost
+shows the delta between the planned infrastructure and the current
+state. Negative deltas = savings.
+
+**How it's computed:** `sum(fact_cost_estimate.delta_usd WHERE delta < 0)`
+per period.
+
+**What's grounded:** the pre-apply estimate (Infracost reads plan JSON,
+offline).
+
+**What's deferred:** actual-spend reconciliation from AWS CUR (D-096 —
+needs live AWS billing). The placeholder view
+`placeholder_live_cur_reconciliation.csv` has the schema ready.

docs/metrics/decision_ledger_coverage.md

@@ -0,0 +1,16 @@
+# Decision Ledger Coverage — Definition of Success
+
+> KPI: Decision Ledger Coverage
+> Target: 100% of AI actions with backfilled outcome
+
+**What this number means:** every AI decision (confidence-gated policy
+engine outcome) is captured in the Decision Ledger with its outcome
+backfilled from the subsequent apply.completed/failed event.
+
+**How it's computed:** `count(decision_ledger rows WHERE outcome ≠
+'pending') ÷ count(decision_ledger rows)`. Sourced from
+`metrics/decision_ledger.db`.
+
+**What "good" looks like:** 100% means no AI decision is ever lost or
+left without an outcome. The ledger is the trust substrate (NORTH_STAR
+Objective #2).

docs/metrics/deployment_frequency.md

@@ -0,0 +1,13 @@
+# Deployment Frequency — Definition of Success
+
+> KPI: Deployment Frequency
+> Target: not a committed target (operational signal)
+
+**What this number means:** the rate of infrastructure state updates
+deployed safely per day. A DORA-adjacent metric for infrastructure.
+
+**How it's computed:** `count(run.completed WHERE exit_code = 0)` per
+day.
+
+**What "good" looks like:** multiple deploys per day (vs. weekly/monthly
+for human ops teams).

docs/metrics/fte_hours_saved.md

@@ -0,0 +1,17 @@
+# FTE Hours Saved (Toil Reallocation Value) — Definition of Success
+
+> KPI: FTE Hours Saved
+> Target: ≥ 70% of pre-Nova FTE allocation (derived)
+
+**What this number means:** the engineering hours saved by automated
+operations, valued at the blended engineering rate. This is what those
+hours were spent on instead (the "toil reallocation" — capital freed
+up from ops to feature development).
+
+**How it's computed:** `run count × manual baseline minutes per run ÷ 60
+× blended hourly rate`. The manual baseline is the estimated time a
+human team would take for the same operation (e.g., 30 min/ticket).
+
+**Honesty caveat:** computed on N internal runs today; the production-
+denominator activates post-pilot. The formula is grounded; the
+production numbers are not yet.

docs/metrics/human_escalation_frequency.md

@@ -0,0 +1,18 @@
+# Human Escalation Frequency — Definition of Success
+
+> KPI: Human Escalation Frequency
+> Target: < 0.1% of platform actions (Post-Pilot)
+
+**What this number means:** how often the AI platform was forced to fall
+back or escalate to a human operator due to low confidence. This is the
+inverse of Touchless Resolution Rate, scoped to operational escalations
+only.
+
+**How it's computed:** `count(runs WHERE hitl_block = 1 AND reason =
+'confidence')` ÷ `total runs`. Attestation sign-offs are excluded.
+
+**What "good" looks like:** < 0.1% means fewer than 1 in 1000 runs
+require human intervention. Near-zero is the goal.
+
+**What would be "gamer metrics":** counting attestation sign-offs as
+escalations (they're not — they're designed controls).

docs/metrics/mttr.md

@@ -0,0 +1,19 @@
+# MTTR (Platform-Run) — Definition of Success
+
+> KPI: MTTR (p95)
+> Target: < 60 seconds
+
+**What this number means:** the time from a platform-run failure
+(apply.failed) to a successful retry. This is platform-run MTTR, not
+infra-incident MTTR (which requires an incident detection system that
+Nova doesn't have yet — deferred).
+
+**How it's computed:** p95 of `successful_retry.time − failed_run.time`
+across all runs that failed then succeeded.
+
+**What "good" looks like:** < 60 seconds means the platform recovers
+from a failed run in under a minute, 95% of the time.
+
+**What's deferred:** infra-incident MTTR (anomaly detected → healed)
+requires an incident detection/remediation system (self-healing
+velocity). That's a future emitter.

docs/metrics/platform_roi.md

@@ -0,0 +1,15 @@
+# Platform ROI — Definition of Success
+
+> KPI: Platform ROI
+> Target: ≥ 250% measured annually (derived)
+
+**What this number means:** the total financial value delivered (labor
+savings + cloud cost optimization + avoided downtime losses) vs. the
+platform's operational/licensing cost.
+
+**Formula:** `(FTE hours saved × blended rate + cloud savings + avoided
+downtime) ÷ platform op cost`.
+
+**Honesty caveat:** computed on N internal runs today; the production-
+denominator activates post-pilot. The formula is grounded; the
+production numbers are not yet.

docs/metrics/policy_compliance_rate.md

@@ -0,0 +1,14 @@
+# Zero-Trust Policy Compliance Rate — Definition of Success
+
+> KPI: Zero-Trust Policy Compliance Rate
+> Target: not a committed target (operational signal)
+
+**What this number means:** the percentage of infrastructure assets
+continuously verified as compliant with security baselines and policies.
+
+**How it's computed:** `1 − count(assets WHERE last_scan.status ≠ pass)
+÷ count(assets)`. Sourced from `fact_policy_check` (Checkov results).
+
+**What "good" looks like:** 100% means every resource passed every
+policy check. The Nova tagging standard (nova_tagging.py, hard mode) is
+the primary check.

docs/metrics/provisioning_lead_time.md

@@ -0,0 +1,13 @@
+# Provisioning Lead Time — Definition of Success
+
+> KPI: Provisioning Lead Time
+> Target: not a committed target (operational signal)
+
+**What this number means:** the time from intent received (run.started)
+to apply completed (run.completed). Measures how fast Nova provisions
+compliant environments.
+
+**How it's computed:** `run.completed_at − run.started_at` per run.
+
+**What "good" looks like:** minutes, not days. The reduction from days
+(human ops) to minutes (autonomous) is the velocity proof.

docs/metrics/touchless_resolution_rate.md

@@ -0,0 +1,23 @@
+# Touchless Resolution Rate — Definition of Success
+
+> KPI: Touchless Resolution Rate
+> Target: ≥ 99% across production estates (Post-Pilot)
+
+**What this number means:** the percentage of platform runs that complete
+end-to-end without an operational HITL block. An operational HITL block
+is a confidence-driven escalation (the AI's confidence was too low to
+proceed). Attestation gates (qa/prod/dr sign-offs) are NOT counted as
+escalations — they are designed controls, not autonomy failures.
+
+**How it's computed:** `runs WHERE hitl_block = 0 AND environment = 'dev'`
+÷ `total runs` (dev environment only, where attestation gates don't apply).
+For production estates: `runs WHERE hitl_block = 0` ÷ `total runs`
+excluding attestation-gate sign-offs.
+
+**What "good" looks like:** ≥ 99% means fewer than 1 in 100 runs require
+human intervention due to low confidence. The 1% allowance is for
+genuine edge cases (novel failure modes, blast-radius exceedances).
+
+**What would be "gamer metrics":** counting attestation gates as
+"touchless" (they're not — they're human by design) or counting only
+dev runs (cherry-picking the easiest environment).