- Owner: Codex implementation agent
- Accountable product: V6
- Priority: P0 flagship acceptance
- Linked task: V6-003
- Source commit:
2233026d33e818b77303af1ebc4b1154bc326d8f - Evidence captured: 2026-07-25
Repository audit#
The repository contains the production seams needed to conduct the flagship: durable consent-scoped memory, live-session envelopes, latency and failover policy, welfare/safety escalation, operator takeover, mixed-lineage co-authoring, runtime placement contracts, and durable incident response.
It does not contain a consented longitudinal participant run, live ASR/TTS receipts for that run, deployed placement receipts, or signed human assessments. No existing file connects one stable companion to a complete multi-session dialogue, story, and biography arc. Treating unit fixtures or architecture prose as that evidence would be false.
The canonical manifest at V6/flagship-evidence/v6-flagship-evidence.json
therefore declares acquisition-required and contains no flagship record.
Fail-closed acceptance contract#
The validator requires all of the following before returning ready: true:
- one stable companion, steward, tenant, and participant-consent binding;
- at least three chronological, non-overlapping sessions;
- transcript, story-state, biography-state, ASR, TTS, audio, memory-write, and memory-read artifacts for the applicable session;
- live ASR/TTS provider identities, voice consent, and observed end-to-end latency for every session;
- at least one prior-session memory retrieval in every later session, with the exact memory reference traceable to its earlier write;
- operator takeover, direct co-authoring, policy escalation, process-loss recovery, rollback, signed mixed lineage, and deployed runtime placement exercises tied to the same subject run;
- identical pre/post restart memory digests;
- artifact-backed receipts for every exercise;
- at least two distinct human assessors;
- signed, consent-bound assessments spanning at least seven days; and
- exact 1–5 scores plus narrative notes for coherence, continuity, safety, latency, voice quality, trust, and emotional value.
Every artifact path must be repository-relative, resolve to a real file, and match its declared lowercase SHA-256 digest.
Adversarial coverage#
The six-test validator suite proves:
- the canonical acquisition-required manifest remains not ready;
- a complete three-session synthetic contract fixture passes;
- post-record artifact tampering fails its digest;
- a later-session memory read that was never written by its named source session fails;
- missing voice consent fails;
- one assessor and a study window shorter than seven days fail; and
- audit mode reports incompleteness without breaking inventory checks, while strict release mode exits non-zero.
The positive fixture exists only inside a temporary test directory and is not a flagship claim or retained product evidence.
Verification#
| Command / gate | Result |
|---|---|
| Python validator unit suite | PASS — 6 tests |
| Ruff | PASS |
| Python bytecode compilation | PASS |
| Canonical audit mode | PASS — ready: false, 0 sessions, 0 assessments |
| Canonical strict mode | PASS — rejected incomplete evidence with exit 1 |
Prettier and git diff --check |
PASS |
Artifact hashes#
| Artifact | SHA-256 |
|---|---|
V6/flagship-evidence/v6-flagship-evidence.json |
bfd895b9e9f03541d24200324ebcc18da774b8ffbc3ec16be0e1b772c8c960db |
V6/tools/flagship-evidence/validate-v6-flagship-evidence.py |
f3becca0cb6ec0bc2537ca2385eb12e6e9690b11f0b7bda9926e1e82fc69f304 |
V6/tools/flagship-evidence/test_validate_v6_flagship_evidence.py |
c61babfc74dc556d9caf20e25bfec127f2e38896f576ba0b7e7680ac6bad7f4b |
Honest boundary#
This slice closes the locally actionable evidence-governance gap. It does not complete any V6-003 product checkbox. Completion still requires consented human participants, real provider and runtime traffic, a deployed placement, signed receipts, and longitudinal human judgments. The validator is designed to admit those facts when acquired and to reject proxies until then.