ORACLE Pre-Conversion Intent v1.0 — STEP 0 Verify-Before-Build Report

Date: 2026-08-18 Session: oracle-preconv (cloud; master prompt v1.2 — confirmed running v1.2 after the owner supplied it; v1.1 executed first, v1.2 patches folded in) Plan of record: docs/ORACLE_PRECONVERSION_INTENT_v1.0.md — sha256 bd8096f786d0e31d356f82ab30d0d26266d93c71dbf093cb4f296ff31dd3fa66 (20,774 bytes, byte-identical to the Drive prompt-board copy 1OM0ZhFuEozRTgkzXIE_EHLzjKAFZFQe5) Authority order applied: WIRING.md > OFFERING_MAP v2.2 > .claude/rules/03-canonical-architecture.md > plan of record > master prompt. Rule applied throughout: where the registry/tree disagrees with the plan, the registry wins — the delta is recorded here and bindings adjusted; the plan is not forced.

claim_label for everything in this lane: built, pre-benchmark. Nothing below is a performance claim.

0. Coordination preconditions

Check	Result
Ledger overlap	None. Session board + `CLAUDE.md` inbox + `.claude/memory/inbox/2026-08.md` read before build; no live claim touches cells 33–35 / creative vectors / bridges. The open `cell37-external-kg-v1.2-build` claim's `src/shared/virtuoso_models` overlap is landed history (PR #716 merged, `d22bb99`).
Shared memory	`shared-memory/memory.py recall` ×2 (oracle preconversion / cell33 behavioral schema) = 0 matches — no prior or duplicate build.
Claim recorded	`claim: oracle-preconversion-intent-v1.0` recorded 2026-08-18T20:10:54Z, landed on `origin/main` (memory-only commit `5bbf95e`, ancestry-verified; survived the `e929cda` theirs-merge intact — re-grepped on `origin/main`). Session-board row added.
Branch discipline	Build branch `claude/oracle-preconversion-intent-v1.0` — local only, no push until in-session `APPROVED: MERGE` (GATE 1). The only pushes so far are memory/coordination commits on `claude/new-session-b63rxg` (claim visibility per rule 02; precedent: the cell37 lane's GATE-1 claim landed the same way).

0b. v1.2 patch — standing-gate script verification (FALSE-GREEN FOUND)

The v1.2 patch mandates verifying each standing gate's actual filename and invocation, because "a gate that errors on a missing script is a FAILED gate, never a passed one." Verifying surfaced a false green in v1.1's gate list:

python3 scripts/mizoki_canon.py --check is a NO-OP that always exits 0. scripts/mizoki_canon.py is a library — no argparse, no __main__, no --check handling (grep-verified; its own docstring names the real consumers). Running it with --check imports the module and exits 0 without checking anything. This matches the y3m3cq pre-flight review's finding F1 (docs/reports/ORACLE_PRECONV_PROMPT_REVIEW_2026-08-18.md, HIGH). Any prior "canon clean via mizoki_canon --check" claim is void.
The verified standing-gate commands, used at every gate hereafter: 1. Full pytest — per cell, fresh venv. 2. Canon lint: python3 scripts/skill_sync.py --audit (exit 1 on HIGH; the real canon gate — replaces the no-op). 3. Skill parity: python3 scripts/skill_sync.py --check and python3 scripts/skills_sync.py --check (BOTH exist — the first checks 14 skills + derived copies + registry; the second checks skill-copy byte-identity/version parity) and python3 scripts/ontology_skills_sync.py --check. 4. Hygiene greps (rule 03 V1 SRDAL / V2 retired-model / V3 PA-API). 5. python3 scripts/claude_memory.py check --strict.
Measured results on the rebased branch: skill_sync.py --audit exit 0 (no HIGH; only pre-existing LOW findings in skill files this build never touched), skill_sync.py --check / skills_sync.py --check / ontology_skills_sync.py --check all exit 0, hygiene greps clean on changed source, memory strict valid.

0c. v1.2 patch — Cell 37 canon delta (recorded)

The v1.2 patch requires recording Cell 37's actual registry identity vs OFFERING_MAP v2.2 and queueing an OFFERING_MAP correction if it drifts. Measured:

OFFERING_MAP v2.2 names Cell 37 "Data Injector". The deployed Cloud Run service is market-signal-ingest (src/cells/cell37/). docs/architecture/CELL_REGISTRY.md row 37 (the number authority) already reconciles both: name market-signal-ingest, canonical role "Data Injector & External Intelligence Gateway" (owner ruling 2026-08-11), external-intelligence scope largely [IN BUILD], with an explicit "stated role vs. built reality" note.
Assessment: this is already-recorded canon, not new unrecorded drift. CELL_REGISTRY.md (which outranks OFFERING_MAP on cell numbers/identity) already carries the reconciliation with its evidence. This build uses Cell 37 only as external-evidence corroboration for latent bridges (Workstream D), which matches the canonical role exactly.
OFFERING_MAP correction queued (end-of-run report): OFFERING_MAP v2.2's bare "Data Injector" label for Cell 37 could be tightened to match CELL_REGISTRY.md's fuller "Data Injector & External Intelligence Gateway (market-signal-ingest; external-intel scope [IN BUILD])". This is a docs-only tightening for the canon-owner; this build does not touch OFFERING_MAP. Recorded so the delta does not survive this run unrecorded.

1. Live Cell Registry — ground truth

gcloud is not installed in this cloud sandbox (which gcloud = exit 1; rule 02 records this as a known environment fact). The Cloud-Run half of the registry check is therefore not runnable from this session; the registry files + live-state records are the ground truth used, and every claim below about serving state is a repo-recorded state, not a fresh probe:

docs/architecture/CELL_REGISTRY.md (number authority, 39 registered cells): cells 33–36 = intent-signal-ingest / intent-scoring-api / intent-graph / intent-causal — all deployed, IAM-locked, dispatch-only (deploy-intent-platform.yml); cell 36 live-verified 2026-08-11 (run 31494599562). Cell 34 rev 00018-c7b serving per lii-deploy-smoke row (2026-08-12).
Cell 37 market-signal-ingest (Data Injector & External Intelligence Gateway): deployed, partially operational; latest revision 00011-tfm per register item 25; extract leg not green (operator-held). Its MarketSignal layer (market_signal.py, market_ingest_gate) is the external-evidence path Workstream D treats as corroboration only.
Plan deltas (registry wins): the plan's header says "37-cell platform" (registry: 39); its C1 table calls Cell 35 "Neo4j intent graph" (registry: Firestore-journal + in-memory serving; Neo4j retired 2026-08-09, neo4j-uri NXDOMAIN by choice) and Cell 37 "data-injector" (registry name: market-signal-ingest, role per owner ruling 2026-08-11).

2. JourneyEvent schema + `response_schema_hash` baseline

Schema: src/shared/virtuoso_models/schemas/journey-event.json (byte-identical vendored twin under services/virtuoso-models-service/virtuoso_models/). No behavioral_signal block exists today — plan delta recorded; Workstream A adds it additively.
response_schema_hash = sha256(canonical(journey-event.json)), computed live in journey_event.schema_hash() and auto-stamped into provenance ("never hand-write" — package CLAUDE.md). Baseline hash at Step 0, measured in-session (python3 -c "…schema_hash()" on this tree): 686738a05b369f325ca9e2924a622a44db6575ebb66aa6f5b2122860a2cd9161 (corroborated: matches the 686738a05b36 schema pin recorded on the session board by the cre-outreach-p3 lane's mapper union merge)
The bump mechanism is therefore: change the schema file additively → the stamped hash moves everywhere automatically. No test pins a literal hash (both suites assert stamped == computed), so the deliberate bump needs no pin edits; the build adds an explicit old-events-still-validate test (rollback level 3).
Two envelopes, both governed, neither bypassed: cell33's IntentSignal is a constrained projection of mizoki_contracts.envelope.CanonicalEventEnvelope (the R2 "no parallel schema" ruling) — that is the consent-gated intent door. The virtuoso_models JourneyEvent is the Sense-phase journey layer (MAPPERS → ingest_gate, MERGE on event_id, CAS on source_payload_hash). The plan's A.3 language maps to the JourneyEvent side; the behavioral extension lands on both, additively, with the same rejection gates.

3. Cell 34 model lane

Plan says	Tree says (wins)
"matrix factorization + boosted tree, nightly batch"	BQML `BOOSTED_TREE_CLASSIFIER` (`model/bqml.sql:105`, `intent_bqml_v1`), hourly single-writer batch (`workflows/intent_scoring_schedule.yaml`, cron `7 * * * *`), plus an offline fv2 logistic surrogate (`backtest/model.py`, `fv2-logit-1`) and a memory-mode heuristic scorer. No realtime inference path exists.
"existing `LII_REALTIME` feature flag"	Does not exist (repo grep: doc/skill prose only). Per Step 0 instruction it is created in this build: per-cell env flag, default OFF, source-literal asserted by test. No central flag registry is used by cells 33–36 (family convention = per-cell `Settings` + env; `src/shared/flags.py` is legacy-cell-only).
"circuit breaker unchanged: on outage, SDK serves cached stale=true"	No circuit breaker exists in cell34 (grep = 0 hits). The analogous machinery is readiness verdicts (`ready\|degraded\|blocked`) and Cell 33's freshness verdict gate. Honored by architecture instead: the transformer path is additive beside the batch path; any transformer-path failure falls back to the existing serving behavior. Nothing existing is modified.
"LEARN loop writes `realized`/`realized_at` to `unified.intent_predictions`"	Neither the table nor the columns exist in code (repo-wide grep = skill/doc prose only). The real label loop: `mizoki_intent.intent_outcomes` joined at horizon → `sp_build_training_examples` → `mizoki_intent.intent_training_examples` (leakage gate `available_to_model_at <= label_cutoff`, `scoring.py:118-134`). The transformer's label stream binds to this loop.
"128-dim matching `unified.intent_vectors`"	`unified.intent_vectors` does not exist (docs only). 128 stands as the design dimension for the new surfaces. Spec-level `unified.` names resolve to dataset `mizoki_unified_data`* (ADR-NY-001 §4, recorded in `bigquery/schemas/measurement_rails_ddl.sql` header).
shadow table "`unified.intent_predictions_shadow`"	Bound to the real scores table's shape instead: `mizoki_intent.intent_scores_shadow` (shadow twin of `intent_scores`, additive DDL, operator-applied).

fv2 lockstep constraint honored: the five fv2 homes (scoring.py offline twin, bqml.sql, bqml_procedures.sql, ScoreStore.bqml_predictions, pipeline FEATURE_COLS) are not modified — the transformer is a separate, flag-off lane. Dependency note: cell34's image has no numpy/sklearn/torch; the transformer adds numpy with an exact pin (pure-CPU inference; latency benchmark test included).

4. Shopify Web Pixel (P1) — capture surface

Measured state (this repo): - The pixel JS artifact is not in this repository — it ships with the Shopify app project (docs/reports/SIGNAL_SHOPIFY_LANE_STATUS_2026-08-12_R3.md:68-71, docs/roadmap/P1_BUILD_PLAN.md:92-95). - The server side that exists here: webPixelCreate activation (services/service-marketing-connectors/shopify_install.py:651-737, settings allowlist {ingest_url, shop_domain}) pointing at {SHOPIFY_PIXEL_EXTENDER_URL}/pixel/events (main.py:838), deferring loudly while unset. - /pixel/events does not exist yet on intent-shopify-extender (routes today: /v1/downstream/shopify, /health, /readyz). The P1 record says the receiving endpoint "is designed together with that artifact".

Binding for Workstream A (no second collector): this build ships (1) the extender's /pixel/events receiving endpoint — flag-off, consent-gated, forwarding into the single cell33 door — and (2) the shared batching/VDI capture core as a versioned JS artifact + settings contract in this repo, which the Shopify-app pixel wraps (that wrap is [IN BUILD] in the app project, out of this repo's reach). No parallel ingest path is created; the endpoint fans into POST /v1/signals exactly as the webhook extender does.

Consent: evaluate_consent (src/shared/mizoki_intent/signal.py:103) — fail-closed, affirmative purpose-specific intent_processing: true, EU/EEA/UK + unknown regions require explicit basis. Gate order in cell33 (orchestrator.py:230-246): audio → consent (denied = dropped+counted, never persisted) → envelope → look-ahead floor → closed-contract validation → idempotent persist. Scope form ("analytics-only does not qualify") lives in libs/mizoki_lii/governance.py (CONSENT_SCOPE="intent_modeling") with both-direction tests.
Consent suites: src/cells/cell33/tests/test_ingest_api.py (ConsentGateTests/AbsentConsentGateTests + region sweep), test_outcomes_api.py, libs/mizoki_lii/tests/test_governance.py, extender test_extender.py::test_no_consent_never_forwarded.
Erasure: registry IDENTITY_LINKED_TABLES (cell33 storage.py:493 — 6 tables) minus ERASURE_EXCLUDED_TABLES must equal the union of every cell's SUBJECT_STORES — enforced by the divergence guard src/cells/cell34/tests/test_erasure_coverage.py. Cascade coordinator: cell34 POST /v1/intent/subject/{id}:erase. Graph leg: cell35 DELETE /v1/identity/{id} (memory sweep + Firestore journal subject_key stamps; SUBJECT_OWNED_LABELS, count→erase→recount receipts).
Erasure suites the new surfaces join: cell33/tests/test_subject_rights.py, cell34/tests/test_subject_rights.py + test_erasure_coverage.py, cell35/tests/test_graph_erasure.py, test_erasure_api.py, test_erasure_cascade.py, test_subject_rights.py, cell36/tests/test_subject_rights.py, plus the four test_dsar_readiness.py.
Deny-list: dual enforcement (write-time 422 in validate_signal_payload; boot refusal in load_taxonomy; read-path scrub scrub_deny_listed). BASE_DENY_TERMS in code can only be extended by YAML, never shrunk.

6. Plan-of-record artifact

docs/ORACLE_PRECONVERSION_INTENT_v1.0.md committed in this branch, byte-identical to the Drive prompt-board original (size match 20,774 B; base64-faithful transfer; sha256 above recorded in the coordination claim as well).

7. Cell 35 graph namespace — Workstreams C/D binding

Tree facts: namespace is frozen allowlists (graph_store.py:128-157) — labels {IntentIdentity, IntentTopic, SignalAggregate, IntentExplanation}, edges {HAS_INTENT, HAS_EXPLANATION, CITES, EXHIBITED}, composite tenant-first identity keys, mandatory edge metadata {confidence, source, timestamp, tenant_id, provenance}, durable-journal subject stamps (SUBJECT_OWNED_LABELS, edge_envelope endpoint rule). There are no Topic nodes and no PRECEDES edges (plan delta; the shipped co-occurrence substrate is HAS_INTENT edges + intent_transitions).

Binding: the plan's (:Topic) → IntentTopic; (:Customer) → IntentIdentity (composite key). New labels Creative (key (tenant_id, creative_id)) and LatentBridge (key bridge_id = tenant-embedded hash, brg_… per the agg_/exp_ convention); new edges RESONATED_WITH (IntentIdentity→Creative), IN_STATE (IntentIdentity→LatentBridge), EVIDENCES (IntentTopic→LatentBridge). All three new-edge classes are subject-owned where they start at IntentIdentity and join the erasure sweep + journal stamps + subject snapshot; Creative/LatentBridge nodes are tenant-scoped, identity-free (aggregate class — deliberately out of erasure's reach, same rule as SignalAggregate, published to auditors via the retained-nodes list).

8. Model roles

Role.CREATIVE_MM (→ gpt-5.6-sol) and Role.DATA_CAUSAL (→ gemini-3.6-flash, GA flag on) both exist in model_registry.py — no registry changes needed; all LLM access in this build goes through virtuoso_call(Role.…); zero hardcoded model strings.

9. Out of scope (restated, binding)

No activation of any kind (hesitation retargeting, aesthetic serving, bridge targeting), no holdout registration, no dashboard surfacing (O-5 open), no change to causal-credit math, no deploys, no DDL application (operator-applied per house convention), no site-visible files. All new capability flags default OFF with source-literal tests.