AUDIT.md — MIZ OKI 3.5 Integrated Bundle

Full-session audit record · 2026-07-25

Scope: every artifact produced in this build session, re-reviewed from the top with fresh eyes and re-tested with the real toolchain. Proof of health: python selfcheck.py → 20/20 checks pass (exit 0).

A. Issues found in THIS audit pass (all fixed, all regression-tested)

#	Severity	Where	Bug	Fix
1	Critical (test infra)	every prior test run	`pip install jsonschema` ran without `--break-system-packages`, failing silently on Ubuntu 24 — so all earlier "validation PASS" results used the weak fallback validator, never real JSON Schema validation	jsonschema installed correctly; selfcheck _t5 requires the real validator; MIGRATION documents the flag
2	Critical	`virtuoso_models/schemas/journey-event.json`	schema (`additionalProperties:false`) had no top-level `source_payload_hash`, but every mapper emits it → with real jsonschema, every mapped event failed validation; ingestion would be dead on arrival in production. Masked by #1	property added to schema; all 4 mappers re-validated with real jsonschema; `schema_hash` auto-updated
3	High	`mizoki_media/mdes.py` `_action_rank`	4 of 5 declared high-blast-radius actions (`channel_launch`, `channel_kill`, `audience_strategy_change`, `bid_strategy_migration`) ranked as "never allowed" → the operator-gate path for them was unreachable at any band	high-blast actions rank at expansion level: reachable only at the expand band, always behind the hard gate; unknown actions still never authorizable
4	High	`mizoki_media/contracts.py` `MediaEvent`	`event_id` hashed only kind+channel+timestamp → distinct payloads in the same second collide, breaking ingest idempotency	payload sha256 folded into the id; collision regression test added
5	Medium (correctness)	`mizoki_cre/simulation.py`	break-even occupancy formula double-counted occupancy (`ads·occ²/NOI` instead of `ads·occ/NOI`) → understated break-even	formula corrected; sanity band asserted in selfcheck
6	Medium (honesty)	`mizoki_cre/simulation.py`	with zero history rows the gate still named `historical_bootstrap` primary while `simulate()` silently ran Gaussian → the passport misrepresented its own method	zero-history → Gaussian primary with `history_status:"none_gaussian_proxy"`; proxy-substituted baselines carry an explicit `"proxy"` label; source-of-truth rule text synced
7	Low (toothless control)	`mizoki_finance/contracts` `EvidenceBundle`	`explanation_ready` checked `contradicting_sources is not None`, but the field defaulted to `[]` — the "contradiction search must have run" control could never fail	default is now `None` (= search never ran → not ready); the retrieval pipeline always sets a real list; regression test added
8	Hygiene	prior turns	no permanent test suite existed — MIGRATION.md told operators to "run the pytest suites" that were never shipped; all prior tests were inline throwaways	`selfcheck.py` shipped at bundle root: 20 permanent checks incl. regression tests for #2–#7; MIGRATION step 2 now runs it

B. Issues found and fixed in EARLIER passes (for the complete record)

vendor_adapters.py called assert_fallback_not_primary without importing it → NameError on every call (caught in first audit round).
virtuoso_call/get_fallback/GLOBAL_FALLBACK missing from __init__ exports → documented API didn't import.
get_model() mutated the shared registry entry when the 3.5-Pro flip env var was set (sticky flip) → returns a copy via dataclasses.replace.
model_registry.py used dataclasses.replace without importing dataclasses.
CRE Monte Carlo produced complex numbers on deep-loss runs (fractional root of a negative equity multiple) → sign-aware IRR capped at −100%.
WIRING.md documented only 3 keys per cell → ANTHROPIC_API_KEY required everywhere for the global fallback; CLAUDE.md lacked the cross-vendor fallback rule (rule 5).
Whitepaper/messaging: fabricated citation ("FinInvest-GTCN") removed pre-build; AVM QC rule rescoped OUT of CRE; HMAC tokens reclassified as pseudonymized personal data — all validation-driven corrections, encoded in code.

C. What was checked and found SOUND this pass

Role lock + reversible 3.5-Pro flip; cross-vendor failover incl. fallback=False re-raise; legacy-string guard; ingest idempotency (insert→duplicate→updated); finance look-ahead guard, no-bypass ValidationLab (weak signal FAILs: PBO 0.83/DSR 0.0), end-to-end pipeline chain integrity; media consent envelope (EEA TCF-native, IP-matching refusal, erasure propagation), incrementality-gated expansion, score-alone-never-unlocks, OPE clipping + divergence warning; CRE NOI chain, SR 11-7 validator independence, sequential phase locks, multi-year Phase B, weight-calibration guards; three-domain conformance.

D. Standing limits — unchanged and stated plainly

0% deployed. Nothing here has been pushed to GitHub, built into an image, or served on Cloud Run. No live API call to any model vendor has ever been made from this environment; the failover test stubs the dispatch layer.
Verified in isolation. Deterministic stubs stand in for trained models (regime engines, TGNN, uplift), real market/ad/deal data, Neo4j/BigQuery sessions, and a real CMP. The interfaces are production-final; the internals behind hooks are Phase-3 prototypes by design.
Internal repo claims remain unverified (Chrome extension six-tool registration) — verify via CLAUDE_CODE_PROMPT.md against the actual repo.
gemini-3.5-pro flip target is provisional until GA; flip is env-gated and reversible. Claude Fable 5 remains uncallable via API (suspended) — routing correctly avoids it.
Path to live: PUSH.md → CLAUDE_CODE_PROMPT.md in Claude Code → staging with live-key proofs → production behind observe/recommend autonomy (MIGRATION.md).