LII Anticipatory Intent — Phase 0 Discovery Report

Date: 2026-08-18 Phase: P0 (read-only discovery, no code) Claim: Latent Intent Inference integration plan audit — cells 33–36 existing LII, lift-engine causal modules, schema gaps, cell allocation, build gates Plan of record: docs/INTENT_API_PLATFORM_BLUEPRINT.md (the existing blueprint, v1.0, 2026-07-30) External plan: Boris's "How 'They Read My Mind'" anticipatory-intent plan (LII_ANTICIPATORY_INTENT_PLAN_v1.md) — NOT FOUND IN REPO; must be sourced before build work begins.

0. Plan Document Status

The external plan document (LII_ANTICIPATORY_INTENT_PLAN_v1.md) was not found at outputs/, docs/, or anywhere in the repository. A search across the Google Drive working directories also returned no match. The plan must be provided by the owner before any build-phase work can proceed.

This discovery report audits the existing LII infrastructure against the plan's known scope (new cells for latent intent, identity/household graph, incrementality/causal; behavioral micro-signals; interest graph; caused-vs-anticipated credit), as described in the tasking context.

1. Cell Registry — Ground Truth

Source: docs/architecture/CELL_REGISTRY.md v1.1 (the number authority, owner rulings 2026-08-11/2026-08-16)

• Current count: 39 registered cells (32 production + 4 LII + 1 gateway + 2 CRE prospecting)
• Constitution preamble: patched to 39 (owner ruling 2026-08-16)
• Next free indices: 40 onward — cells 1–39 are all assigned

Existing LII cells (33–36)

Cell 37 (market-signal-ingest) is the Data Injector & External Intelligence Gateway — partially operational, external-intelligence scope largely [IN BUILD]. Cells 38–39 are CRE Prospecting, deployed-CI.

Proposed assignments for the LII plan's new cells

The plan proposes three new logical components: (a) a latent-intent cell, (b) an identity & household graph cell, and (c) an incrementality/causal cell. However, much of this capability already exists in Cells 33–36 and the lift-engine service. Before allocating new cell numbers, the owner must decide whether the plan's components are extensions of existing cells or new cells.

If new cells are needed:

Owner decision required (Article VI): whether to allocate new cells (40–42) vs. extend existing cells (33–36). The build-debt approach of extending existing cells is lower-risk and follows the owner's ruling on CRE (extend cells inline rather than creating a privacy cell — ADR-LII-001 decision (a)).

2. Existing LII Code Inventory

Grand total: ~43,749 lines across Cells 33–37 + shared package + SDK + MCP server, with 1,023 test functions. All cells except Cell 37 use exact dependency pins.

2.1 Cell 33 — intent-signal-ingest (SENSE)

Location: src/cells/cell33/ Key modules: - ingest_cell/orchestrator.py — consent gate (fail-closed), audio prohibition, signal taxonomy validation, envelope construction, BigQuery persistence - ingest_cell/main.py — FastAPI routes: POST /v1/signals, GET /health, GET /readyz, GET /metrics - schema/intent_bigquery_ddl.sql — DDL for mizoki_intent tables including intent_signals, intent_outcomes, intent_consent_denials, intent_taxonomy_dim, intent_erasure_audit

Governance enforcement: audio prohibition (3 layers), consent fail-closed (region-aware, EU/EEA/UK deny-by-default), sensitive-topic deny-list (boot refusal), payload contract closed (ALLOWED_PAYLOAD_KEYS = {"signal_type", "topic_id", "strength", "identity_kind", "consent"})

Lines: 6,894 (source ~3,500, tests ~3,400) | Tests: 192 functions | Dependencies: all exact pins

Status: PRODUCTION-READY (deployed, IAM-locked, Shadow since 2026-07-31)

2.2 Cell 34 — intent-scoring-api (REASON)

Location: src/cells/cell34/ Key modules: - scoring_cell/orchestrator.py — scoring pipeline, cohort queries, transition detection, explain proxy, DSAR cascade coordinator - scoring_cell/scoring.py — compute_features() (9 fv2 features), heuristic_score() (fallback), FEATURE_VERSION = "fv2" - scoring_cell/main.py — FastAPI routes: GET /v1/intent/score/{identity_id}, POST /v1/intent/scores:batch, GET /v1/intent/cohort, GET /v1/intent/transitions, GET /v1/intent/explain/..., GET /v1/intent/taxonomy, POST /v1/intent/subject/{identity_id}:erase - backtest/ — offline evaluation harness with fv2 model, oracle ceiling analysis, bootstrap CIs - model/bqml.sql — production BOOSTED_TREE_CLASSIFIER definition

Model quality (from docs/reports/LII_BACKTEST_MODEL_QUALITY_2026-08-16.md): - Baseline (heuristic): AUC 0.6726 mean / Brier 0.2224 - fv2 model: AUC 0.6967 mean / Brier 0.1845 (crosses Brier ≤ 0.20 gate on 8/8 seeds) - Oracle ceiling: AUC 0.7277 (the 0.72 gate is unreachable on synthetic data) - Binding constraint: no forward labels — blocked on real merchant data (Shopify webhooks, GA4, Klaviyo)

Lines: 11,959 (source ~4,100, tests ~4,200, backtest ~2,740) | Tests: 275 functions | Dependencies: all exact pins

Status: PRODUCTION-READY (deployed, IAM-locked, observe-only; Phase A ranking gate FAILED and OPEN)

2.3 Cell 35 — intent-graph (REASON)

Location: src/cells/cell35/ Key modules: - graph_cell/graph_store.py — 1,759 lines; in-memory graph backend with frozen namespace - graph_cell/orchestrator.py — graph operations, explain paths, DSAR erasure leg - graph_cell/sync.py — score-to-graph synchronization - graph_cell/durable.py — optional Firestore write-through durability journal

Graph model (implemented):

What is NOT implemented: - PRECEDES co-occurrence edges (described in OFFERING_MAP F.4 and blueprint §3 Cell 35) - Household/cross-device linking nodes - Interest embeddings or vector indices - Two-tower candidate generation paths

Backend: in-memory; Neo4j permanently retired (2026-08-09 owner decision, NXDOMAIN by choice). Firestore durability journal available but off by default.

Lines: 11,127 (source ~4,990, tests ~5,740) | Tests: 321 functions | Dependencies: all exact pins

Status: PRODUCTION-READY (deployed, IAM-locked; DSAR live-verified 2026-08-10)

2.4 Cell 36 — intent-causal (VALIDATE/DECIDE)

Location: src/cells/cell36/causal_cell/ Key modules (3,043 lines total): - holdouts.py (300 lines) — holdout assignment, deterministic-identity-only gate - estimators.py (639 lines) — DR-Learner, CUPED estimators - refutation.py (333 lines) — arm-level permutation refutation battery - orchestrator.py (577 lines) — incrementality report generation - activation.py (112 lines) — uplift_export_cohort gated activation - storage.py (512 lines) — BigQuery storage, outcome ingestion, probabilistic identity exclusion

Causal capabilities: - ✅ Holdout assignment (deterministic hash, 10% permanent) - ✅ DR-Learner estimator - ✅ CUPED estimator - ✅ Refutation battery (arm-level permutation) - ✅ Probabilistic identity exclusion (3 enforcement points) - ✅ uplift_export_cohort → guardrails → DCP activation path - ❌ Reports held in-memory only (lost on restart; no BigQuery table) - ❌ No ghost-ad/bid-log integration - ❌ No caused-vs-anticipated credit classification yet

Lines: 4,705 (source ~3,060, tests ~1,640) | Tests: 91 functions | Dependencies: all exact pins

Status: PRODUCTION-READY (live-verified 2026-08-11)

2.5 Cell 37 — market-signal-ingest (SENSE / external intelligence)

Location: src/cells/cell37/ Lines: 1,309 (source only) | Tests: ZERO — no test directory exists | Dependencies: MIXED PINS (reproducibility concern per .claude/rules/03-dependency-reproducibility.md)

Cell 37 uses virtuoso_call indirectly via the /extract endpoint (delegates to virtuoso_models.extraction.market_extract.extract_market_signals). This is the only LII-adjacent cell with LLM calls.

Dependency concern: requirements.txt uses floor-only specs for most libraries (fastapi>=0.111, pydantic>=2.7, google-cloud-bigquery>=3.20, etc.) — only google-cloud-firestore, google-api-core, google-genai, and anthropic have upper bounds. This violates the dependency reproducibility rule (03-dependency-reproducibility.md, the Cell 03 GraphRAG cache incident).

2.7 Shared library — src/shared/mizoki_intent/

Key modules: - signal.py — signal validation, audio prohibition, consent evaluation, payload contract - taxonomy.py — taxonomy loader, deny-list enforcement, boot refusal - subject.py — GDPR subject access/erasure with count→delete→recount guarantee - client.py — IntentClient SDK, IntentScore.as_evidence()

2.8 SDK library — libs/mizoki_lii/ (4,089 lines)

Modules: client.py (890 lines), governance.py (439 lines), transport.py (405 lines), auth.py (167 lines), config.py (234 lines), errors.py (83 lines) Tests: 5 test modules (1,430 lines)

Provides LIIClient and LIIConfig for external consumers. Observe-only governance is enforced in the library.

2.9 MCP tool surface

mcp/mizoki-lii-mcp/ — 5 tools (observe-only):

Boss Agent shipped tools (8 intent_* tools): intent_score_get, intent_cohort_query, intent_transitions_recent, intent_explain, intent_taxonomy_list, intent_consent_stats, intent_incrementality_report, intent_status. The 5 lii_* stdio tools are built but Boss registration is owner-gated (open-work item 19).

3. Causal / Lift-Engine Inventory

3.1 Cell 36 causal modules (3,043 lines)

Covered in §2.4 above.

3.2 Lift-engine service (services/lift-engine/) — 12,716 Python lines

Core modules:

3.3 What the lift-engine CAN serve for the LII plan

3.4 Other causal modules in the repo

Total causal code in the platform: ~41,500 lines across all groups. The previously cited ~7,087 figure likely referred to lift-engine core only (src/core/ + src/evaluation/ = ~7,208 lines). The full inventory is nearly 6× larger. No causal module uses virtuoso_call — the entire causal codebase is pure statistical/ML (numpy, sklearn, scipy, econml, causalml, DoWhy).

4. Schema Gap Analysis

4.1 Canonical Event Envelope

Source: contracts/mizoki_contracts/envelope.py

The envelope payload field is Dict[str, Any] — an unconstrained dict. Schema enforcement happens at the cell level, not the envelope level.

4.2 IntentSignal payload contract (Cell 33)

Source: src/shared/mizoki_intent/signal.py

ALLOWED_PAYLOAD_KEYS = frozenset({
    "signal_type", "topic_id", "strength", "identity_kind", "consent",
})

This is a CLOSED contract. Any key not in this set is REJECTED (fail-closed). The behavioral micro-signals the LII plan requires are ALL missing:

Impact: Extending ALLOWED_PAYLOAD_KEYS requires a coordinated change across signal.py, BigQuery DDL (intent_signals table schema), the backtest harness, and the feature pipeline. This is the schema work the plan assumes exists.

4.3 JourneyEvent schema

Source: src/shared/virtuoso_models/schemas/journey-event.json

The context.ext field (additionalProperties: true) is the only extensibility point, but micro-signals are not defined there. The actor object carries user_id, email, phone_sha256, device_ifa, ip, ua — basic identity fields, no behavioral signals.

4.4 MarketSignal schema

Source: src/shared/virtuoso_models/schemas/market-signal.json

Market-level signals only (SERP position, trend index, sales rank, etc.). No person-level behavioral signals — by design (Cell 37 rejects person identifiers).

5. Graph Schema / Ontology Check

5.1 Cell 35 implemented graph model

Four node types, four edge types — frozen namespace (see §2.3 table).

Critical gaps for the LII plan's interest graph:

5.2 Ontology registry

Source: ontology/registry.yaml v1.2.0

No interest, topic, co-occurrence, affinity, embedding, or vector node/edge types exist in the governed ontology. Cell 35's graph types (IntentIdentity, IntentTopic, etc.) are cell-local constructs with no ontology mapping.

Required OCPs for the LII plan: - Architecture-level: LatentIntent as a governed concept (requires owner approval) - Steward-level: InterestEmbedding, TopicCoOccurrence, HouseholdLink, BehavioralMicroSignal - Domain module: marketing/intent domain module (planned but not materialized)

6. API Contract Audit

6.1 Existing LII endpoints (Cell 34 — the canonical Intent API)

6.2 GDPR/DSAR endpoints (all four cells)

6.3 Cell 36 causal endpoints

6.4 Proposed endpoints from the LII plan — status

6.5 Measurement-rails service (the causal credit backbone)

Service: measurement-rails (port 8045) | Auth: NO app-level auth — relies entirely on IAM posture | Status: NOT DEPLOYED by any CI workflow — operator RUNBOOK only | All flags shipped OFF

Critical finding: measurement-rails is NOT in production/service-registry.yaml and has no CI deploy path. The "caused vs anticipated" credit classifier — a central requirement of the LII plan — lives here, not in the lift-engine.

6.6 Lift-engine service API surface

The lift-engine (services/lift-engine/) exposes routes via src/api/routes.py and src/api/causal_routes.py. These are a separate service from the LII cells, not yet wired to intent data.

7. Build Gates

Critical corrections required before any LII plan implementation can proceed. Numbered in severity order.

Gate 1 — Plan document must be sourced

Boris's anticipatory intent plan (LII_ANTICIPATORY_INTENT_PLAN_v1.md) is not in the repository. Without the canonical plan, build scope is undefined. Owner action: provide the plan document.

Gate 2 — Cell allocation requires Article VI owner directive

If the plan calls for new cells (40+), an Article VI directive is required (CONSTITUTION.md VI.1). The same-commit coherence obligation applies: CELL_REGISTRY, CONSTITUTION preamble, source_of_truth, byte-identical twins, check_conformance(), skillpack re-sync.

Owner decision needed: extend existing cells (33–36) vs. allocate new cells (40–42).

Gate 3 — IntentSignal payload contract must be extended

ALLOWED_PAYLOAD_KEYS is a 5-field frozen set. Adding behavioral micro-signals (dwell, scroll, watch ratio, etc.) requires a coordinated change across: - signal.py (add to ALLOWED_PAYLOAD_KEYS) - intent_bigquery_ddl.sql (add columns to intent_signals) - backtest/ harness (add to feature pipeline) - scoring_cell/scoring.py (extend compute_features() fv3) - Migration of existing data (new columns have no historical values)

This is the largest schema change in the LII plan and must be designed as an OCP before implementation.

Gate 4 — All LLM calls through virtuoso_call(Role.X)

Any new LLM-powered component (interest embedding generation, natural-language intent explanations, causal narrative generation) must route through virtuoso_call. No hardcoded model strings. The existing LII cells comply; new code must inherit this.

Gate 5 — Graph namespace frozen — extension requires explicit unlock

Cell 35's ALLOWED_LABELS and ALLOWED_EDGE_TYPES are frozen sets (graph_store.py:128-133). Adding PRECEDES, household edges, or interest embeddings requires: - Extending the frozen namespace (code change + tests) - OCP for the new node/edge types - Deciding whether the new types live in Cell 35 or a new cell

Gate 6 — No vector index/embedding store exists

ADR-LII-001 explicitly confirms zero vector/embedding infrastructure in Cells 33–36. The plan's two-tower embeddings and ANN retrieval (Cell 34 Phase B / Vertex AI Vector Search) are entirely unbuilt. If added, the DSAR erasure cascade MUST be extended to cover the new store — documented obligation in ADR-LII-001 §d.

Gate 7 — Label stream blocked (register item 6c)

The Phase A ranking gate is FAILED and OPEN. The model cannot be evaluated on real data because there are zero forward labels. Unblocking requires: - Shopify webhooks (orders/create) — the label stream - GA4 export link — session/page depth signals - First registered holdout (register item 23) — required for the third autonomy criterion (stable lift ≥ 2 purchase cycles)

No activation is legal without these. The LII plan's advanced capabilities compound this: more features require more signal, which requires more data sources.

Gate 8 — Consent enforcement for behavioral micro-signals

Behavioral micro-signals are GDPR profiling (Article 22). The existing consent gate (evaluate_consent()) checks for intent_processing: True but the new signal types will carry richer personal data (browsing behavior, scroll patterns, dwell times). The consent basis and data-subject rights implications must be reviewed before any micro-signal is persisted.

Gate 9 — Identity resolution rules

The plan's household graph and cross-device linking introduce probabilistic identity resolution. Platform rule: probabilistic identities are excluded from causal math (3 enforcement points in Cell 36). The identity graph must: - Label every link deterministic or probabilistic with confidence - Enforce the deterministic-only gate on all causal paths - Never send person-level data to third-party enrichment (entity-spine scope)

Gate 10 — Dual refutation implementations

Cell 36 has arm-level permutation refutation (333 lines). The lift-engine has full DoWhy refutation (897 lines). The plan must decide which is canonical and whether to merge them. Running two parallel refutation batteries creates a consistency risk.

Gate 11 — Incrementality reports are ephemeral

Cell 36 holds reports in-memory only (bounded to MAX_REPORTS_KEPT = 100, lost on container restart). No BigQuery table exists for intent_incrementality_reports. The DDL defines no such table. Reports are not auditable after the fact. The plan's credit ledger requires durable storage.

Gate 12 — Observe-only is structural, not configurable

Nothing in the repository can promote LII out of observe-only automatically. The promotion criteria (Brier ≤ 0.20, AUC ≥ 0.72, stable lift ≥ 2 purchase cycles, human approval) are documented gates, not code thresholds. The plan must not assume any activation capability exists.

8. Existing Infrastructure — Surprises

Surprises (positive)

1. Cell 36 is live-verified with full refutation. The causal-credit cell is the most production-ready component in the LII stack. The plan's incrementality requirements compose with it, not replace it.

2. The lift-engine has 12,716 lines of causal code. Meta-learners (X/T/S), DoWhy refutation, causal discovery (PC/GES/LiNGAM), RL allocation (UCB/Thompson), Bayesian uplift, counterfactual evaluation — all built. The plan's causal requirements are largely a wiring exercise, not a build.

3. The fv2 model crosses the Brier gate. The backtest report (2026-08-16) shows the fv2 logistic model achieves Brier 0.1845 mean (≤ 0.20 on 8/8 seeds). The fallback scorer does not. This is not yet the production path but the model quality work is ahead of what the plan likely assumes.

4. DSAR is live-verified. The subject-rights cascade across all four cells has been end-to-end tested against real BigQuery, including the streaming-buffer 409 and clean 200 retry. Any new data store the plan adds must extend this cascade.

5. The libs/mizoki_lii SDK exists (4,089 lines). External consumers have a governed client library. The plan's "democratization" goal already has a foundation.

Surprises (negative)

1. PRECEDES edges are described in 3 documents but do not exist in code. The OFFERING_MAP, blueprint, and skillpack all describe co-occurrence reasoning via PRECEDES edges. Cell 35's frozen namespace does not include them and the code has never implemented them.

2. Zero behavioral micro-signal fields in any canonical schema. The plan assumes rich behavioral data (dwell, scroll, hover, watch ratio) can flow through the platform. The payload contract is a 5-field frozen set. Even the JourneyEvent schema has no such fields.

3. The ontology has no LII domain module. Cell 35's graph types are cell-local constructs with no governed ontology mapping. The plan's interest graph requires ontology work.

4. Three refutation engines with no shared interface. Cell 36, the lift-engine, and Cell 26 each have their own refutation implementation. This must be rationalized.

5. Cell 37 has zero tests. The only LII-adjacent cell with no test directory. Given its virtuoso_call usage (the /extract endpoint), this is a coverage gap.

6. Cell 37 dependency pins are floor-only. Most requirements use >= without upper bounds, violating the reproducibility rule derived from the Cell 03 GraphRAG cache incident (03-dependency-reproducibility.md).

7. measurement-rails is not in the service registry and has no CI deploy path. The "caused vs anticipated" credit classifier — central to the LII plan's causal credit requirement — lives in this unregistered service.

8. The boss_agent_skillpack.md quick-reference still references the never-built POST api.mizoki.ai/api/v1/causal/holdout at line 169 — contradicting the documented correction at line 35 of the same file (corrected 2026-08-11).

9. Service registry says cells 33–36 are "designed" but they are deployed and serving (Cell 36 live-verified). The registry is stale.

9. Owner Decisions Needed

10. Phase Summary

Report generated 2026-08-18 as Phase 0 of the LII Anticipatory Intent Integration. No code was written. Next phase requires the plan document and owner decisions D1–D8.