MIZOKI SIGNAL — FINAL PRODUCT OVERVIEW v5.3

The complete offering: ORACLE anticipatory intent + causal proof + governed autonomous media acquisition, with the Shopify offshoot as first commercial instance

Date: August 11, 2026 · v5.3 change: Cell 37 scope expanded by owner — Cell 37 = Data Injector & External Intelligence Gateway: governed injection of non-streaming data PLUS processing of all external data sources and search intelligence (Google, Amazon, Meta, SERP, Wikidata, Schema.org). (v5.2: Cell 37 = Data Injector · v5.1: cells 33–36 canonical for LII.) Supersedes SIGNAL_OVERVIEW_v5.2_FINAL.md, v5.1, and v5.0 on this board. Incorporates: Master Document v4.0 FINAL (Shopify offshoot) + "Signal Intelligence Division — Marketing Capabilities Documentation" + L5 Autonomous Media Architecture blueprint corrections + Story Bank v1.2. This overview is the top-level product description; the Master v4.0 remains the Shopify build spec beneath it. Claim discipline: [Validated] / [Illustrative] / [Roadmap]; headline targets labeled verified/benchmarked/pilot/illustrative per deployment. Naming: MIZ OKI 3.5 (engineering) · MIZOKI3 (brand) · Signal (division) · ORACLE (flagship LII capability) · MIZOKI Signal for Shopify (commercial product). Naming supersession (owner ruling 2026-09-22, OPEN_ITEMS OR-3): "ORACLE" in this document is the internal program codename. Customer-facing copy names the capability Intent Engine v2 within the Causal Growth Control offer (OFFERING_MAP corrections register row 28). The body below is retained as written — a versioned artifact is bannered, never rewritten. Decision material: docs/reports/ORACLE_DOC_PUSH_2026-09-15.md §12.2.

1. Positioning — one sentence, three layers

"Anticipatory intent with proof of causal lift, executed under earned autonomy." - Anticipate (ORACLE): calibrated predictions of who is entering market, what they'll want next, and when — never "mind-reading," never audio, always with an explanation path. - Prove (the ledger): every conversion classified caused vs. anticipated with a confidence interval; the prediction never grades itself. - Act (governed execution): the clipped-ReLU DEL authorizes per action class; autonomy is a certified property earned one reversible class at a time — a certifiable autonomous media control system. Headline design targets [labeled, never guaranteed]: 40% CAC reduction · 35% ROAS increase · 67% ROI improvement.

2. The capability stack (full division map, Shopify offshoot inside it)

2.1 SENSE — omnichannel signal capture [Validated — platform architecture]

Google Ads (GAQL SearchStream + field validation + MCC traversal), OpenRTB bidstream, ESP/email, CRM & ecommerce (Shopify app: orders/refunds/inventory/fulfillments webhooks, bulk backfill, Web Pixel micro-signals), offline/CTV at geo level. Everything lands as a Canonical Event Envelope (tenant, provenance, governance labels, confidence, audit IDs) → BigQuery unified + the Firestore-backed intent graph (graph-store correction 2026-08-19: Neo4j retired 2026-08-09; apply the same fix to the Drive original) + immutable ledger. Consent-first: Cell 33's hard consent gate — no consent, no ingestion; consent state travels on every event. Cell 37 — Data Injector & External Intelligence Gateway [owner-stated; service binding verified in F7]: one governed doorway for everything that isn't a live first-party stream — (a) injection: backfills, bulk loads, partner/file feeds, synthetic training data; and (b) external intelligence: processing of all external data and search sources — Google (search/SERP intelligence, Knowledge Graph entities), Amazon (marketplace/catalog signals via the Creators API, PA-API 5 being deprecated), Meta (platform-exposed signals), Wikidata (CC0 canonical-identity foundation), Schema.org structured data, and SERP feeds. All of it enters through the same Canonical Event Envelope under the blueprint's provenance-aware ingestion rules: consent and provenance labels on every event, public-source agreement alone never commits to the graph, Wikidata anchors canonical identity, Google/Amazon graph access is never assumed (Freebase retired; Trends gated), and synthetic data never leaves the training plane. External intelligence informs targeting hypotheses; only the experiment stack substantiates lift.

2.2 RESOLVE — identity, deterministic-first [Validated]

Exact-match identity on authenticated identifiers first; probabilistic device/IP/behavioral extension only where governance permits; household graph with every link labeled deterministic vs. probabilistic so rollup confidence is always visible. High-stakes personalization prefers deterministic links.

2.3 ANTICIPATE — ORACLE / Latent Intent Inference [Validated infrastructure; outcome claims preview]

Two-stage candidate-generation + ranking (the Covington et al. RecSys '16 pattern): two-tower embeddings (BigQuery ML → Vertex AI Vector Search ANN retrieval) + sequence models for ranking and timing. Served sub-100ms via the Intent Scoring API: intent vectors, predicted next interests with calibrated probabilities (Brier-scored, not vanity scores), intent stage (awareness / consideration / in-market / purchase-imminent), purchase-timing windows, churn/expansion propensity where calibration supports it, and explanation paths through the graph. Intent graph: Customer/Household/Topic/Product/Campaign nodes, SHOWED_INTEREST + PRECEDES edges; dark-funnel inference from micro-signal shape (Gartner: buyers spend ~17% of the journey with suppliers — the rest is invisible; we infer it and say it's inference). Predictive audiences upgrade lookalikes: expansion candidates carry calibrated intent, not demographic resemblance. Alerts: "notify when account enters in-market," Pub/Sub + SSE, Boss-agent MCP workflows.

2.4 PROVE — causal measurement [the credibility core]

2.5 ACT — governed execution (Shopify offshoot = first commercial instance)

Black-box-era, lever-native: the four levers that remain in the ASC/PMax world — (1) value signal (E[NCM] per conversion via CAPI/Conversion Value Rules under the versioned NCM-v1 metric contract; margin and pLTV, never raw revenue), (2) exclusions (existing customers, high-return cohorts, owned-channel converters), (3) creative supply (fatigue detection, intent-stage → message fit, holdout-judged winners; generation out of GA scope), (4) budget & guardrails (covenant caps, working-capital-aware pacing off the Predictive Financial cell). Shop Campaigns and Klaviyo in the reallocation set under the same causal audit; feed enrichment from intent-graph language (L1-safe). Cadence: sense continuously, evaluate every 15 minutes, act dampened (cumulative ±20%/campaign/day; structural weekly).

2.6 GOVERN — the moat

SRPVDAL end-to-end; clipped-ReLU DEL per action class: authority_c = min(cap_c, max(0, DEL_score − threshold_c)); platform floor thresholds (raisable by the customer, never lowerable); deterministic denial below threshold; margin-proportional authority above; mechanical per-class demotion; adaptive envelopes as threshold shifts. Autonomy ladder L0→L5 certified per account × action class, reversible classes first. Model promotion gates [canonical]: Brier ≤ 0.20, AUC ≥ 0.72, stable lift across ≥ 2 purchase cycles — below the bar, models stay advisory (observe-only); degradation above Brier 0.20 demotes automatically. Trust features sold as features: consent gate, sensitive-topic deny-list, no audio, ever, GDPR Art. 22 human-in-the-loop, erasure cascade, immutable journal, one-tap kill switch, weekly plain-language digest. Attestation honesty: proofs verify computation ran, never that lift was causal. Fleet integrity: DP-aggregated priors only; no cross-merchant bid coordination by construction.

3. Who gets what (tiers × functions)

Shopify merchant tiers (Master §1.3): T1 <300 orders/mo (value feeds, feed enrichment, cold-start, priors — no incrementality claims; L0–L1) · T2 300–3K (ghost-bid + cohort holdouts, always-on rotating holdout with the measurement tax disclosed, NCM reallocation; L2–L3) · T3 3K+ (geo, lift-calibrated mini-MMM, cross-channel, covenants; L4–L5 per class). Marketing-function playbooks: Demand Gen (in-market alerts → concentrate spend → validate with ghost-bid/geo → cut anticipated-only channels) · Lifecycle (timing-triggered nurtures; suppress anticipated-anyway sends — Story 3) · Growth (always-on incrementality; optimize iROAS not platform ROAS) · Brand (geo-proven CTV/brand budgets — Story 4) · RevOps (Intent API + ledger as the one causal source of truth for sourced-vs-influenced pipeline). Regional views: geo/segment intent heatmaps on the same structure that powers geo experiments and per-segment uplift.

4. Access surfaces

Command Center (Intent Scores Grid, Predicted Journey Timeline, Incrementality Panel with refutation status + ledger) · Intent Signal API (sub-100ms) · Python SDK · Pub/Sub feed · SSE dashboards · Boss agent via MCP (query intent, pull ledger, draft plan, approval-gated activation) · Shopify embedded app + free Profit Truth Audit wedge.

5. Rollout — three clocks, one sequence

6. Reconciliations

  1. Cell registry — RESOLVED by owner (Aug 11): cells 33–36 online and canonical for LII (ingest / scoring / graph / causal); Cell 37 = Data Injector & External Intelligence Gateway (owner-stated: injection of non-streaming data + processing of all external/search sources — Google, Amazon, Meta, Wikidata, Schema.org, SERP — under provenance-aware ingestion rules). The marketing capabilities doc's 28/33/34/35 mapping is in error and must be corrected wherever it propagated; the "32-cell" count is stale (37 cells). F7 publishes CELL_REGISTRY.md, verifies service bindings from Cloud Run ground truth, marks unbuilt external-intelligence scope [IN BUILD], corrects the capabilities doc, and writes the registry reference into CLAUDE.md. Until it ships, external copy references services by name, not number. (Registry shipped 2026-08-11, v1.1 2026-08-16; the owner's 2026-08-16 ruling added cells 38–39 CRE Prospecting — 39 registered cells now; docs/architecture/CELL_REGISTRY.md remains the number authority.)
  2. Promotion thresholds canonized: Brier ≤0.20 / AUC ≥0.72 / ≥2 cycles wired into the autonomy ladder and demotion machinery.
  3. Headline targets aligned: 40% CAC / 35% ROAS / 67% ROI as labeled design targets — never bare.
  4. Ghost-bid citation upgraded (JMR 2017) with its availability caveat qualifying T2 promises.
  5. Trust features promoted to product surface (deny-list, no-audio, Art. 22, link labeling).
  6. Dark-funnel inference: inference, never observation.
  7. Benchmark sourcing rule: 0.27x / 1.5–3x / 17% are market context, never MIZOKI results.

7. Open owner decisions

Master §3.6's fifteen, plus: 16. ~~Cell registry~~ — CLOSED (33–36 canonical; 37 = Data Injector & External Intelligence Gateway). 17. Stage 1–4 benchmarks as contractual SLAs vs. internal targets. — RULED 2026-09-15 (D-17): SLA on process, never on outcome; see docs/product/PRICING_PACKAGING_v1.md PD-9 (owner ruling 2026-09-15, docs/reports/OWNER_RULINGS_2026-09-15_REGISTER_CLOSEOUT.md).

FINAL INTEGRATION PROMPT v2.3 (supersedes Part V of Master v4.0 — one-paste)

OPERATOR PROMPT — SIGNAL-SHOPIFY LANE: INTEGRATE OVERVIEW v5.3 (2026-08-11)

SOURCES (Drive MIZOKICloudRun prompt-board folder):
  1. SIGNAL_OVERVIEW_v5.3_FINAL.md  — top-level product overview (this doc)
  2. SIGNAL_SHOPIFY_MASTER_v4_FINAL.md — Shopify build spec (remains build canon)
  3. The L5 Autonomous Media Architecture blueprint — platform phases/gates
  4. "Signal Intelligence Division — Marketing Capabilities Documentation"
Hierarchy: overview (positioning) > master (Shopify build) > capabilities doc
(marketing-org reference); blueprint governs platform phases. On conflict,
record in ledger and escalate — do not silently pick.

You are the owner-designated coordinator for the Shopify/media/net-yield/
marketing-docs/JourneyEvent lanes.

CLAIM FEATURES FIRST (scripts/claude_memory.py record --tags coordination):
  [F1] master-doc-integration  [F2] roadmap-binding    [F3] relu-del-canon
  [F4] citation-verification   [F5] fleet-integrity    [F6] p1-kickoff
  [F7] cell-registry-reconciliation  [F8] promotion-gates-canon
(Features, not paths — PR #644 lesson.)

F1–F6: execute per Master v4.0 Part V (unchanged), with one amendment to F1:
  commit BOTH docs — docs/product/SIGNAL_OVERVIEW_v5.md and
  docs/product/SIGNAL_SHOPIFY_MASTER_v4.md; overview is the top-level index;
  CLAUDE.md lane index points at the overview first.

F7 — CELL REGISTRY PUBLICATION (owner rulings: 33-36 canonical; 37 = Data
     Injector & External Intelligence Gateway)
   *(Executed — CELL_REGISTRY.md published 2026-08-11, v1.1 2026-08-16. The
   "37 cells total" below was true at ruling time; the owner's 2026-08-16
   ruling added cells 38–39 CRE Prospecting — 39 registered cells now. The
   registry remains the number authority; note added by the 2026-08-21
   truth-debt sweep.)*
1. OWNER RULINGS 2026-08-11: cells 33-36 are the live LII cells (ingest /
   scoring / graph / causal); CELL 37 = DATA INJECTOR & EXTERNAL
   INTELLIGENCE GATEWAY — (a) governed injection of non-streaming data into
   the Canonical Event Envelope (backfills, bulk loads, partner/file feeds,
   synthetic training data) AND (b) processing of ALL external data and
   search sources: Google (SERP/search intelligence, Knowledge Graph
   entities), Amazon (Creators API; PA-API 5 deprecated), Meta platform-
   exposed signals, Wikidata (CC0 canonical identity), Schema.org, SERP
   feeds — under the blueprint's provenance-aware ingestion rules
   (public-source agreement alone never commits to the graph; Google/
   Amazon graph access never assumed; external intelligence = targeting
   hypotheses only, never lift substantiation).
   The marketing capabilities doc's 28/33/34/35 mapping is IN ERROR —
   correct it in that doc and anywhere it propagated.
2. Verify from ground truth (repo + Cloud Run service list): the 33-36
   service bindings AND Cell 37's service name + deploy status. If the
   deployed reality of Cell 37 differs from the stated dual role, record
   the discrepancy in the ledger and escalate to the owner — do not
   silently relabel. Mark unbuilt portions of the external-intelligence
   scope [IN BUILD] per truth discipline.
3. Publish docs/architecture/CELL_REGISTRY.md as the single canonical
   registry (cell number ↔ service name ↔ function ↔ deploy status),
   37 cells total (the "32-cell" figure is stale everywhere — fix it).
   Route via review PR.
4. UPDATE CLAUDE.md: add/refresh the platform-facts section — 37-cell
   platform, cells 33-36 = LII (ingest/scoring/graph/causal), cell 37 =
   Data Injector & External Intelligence Gateway (injection + Google/
   Amazon/Meta/Wikidata/Schema.org/SERP processing under provenance
   rules), link to CELL_REGISTRY.md as the number authority — so every
   future session inherits the corrected registry. Record in ledger.
5. Rule going forward: external copy references services by NAME; numbers
   are engineering-internal, sourced only from CELL_REGISTRY.md.

F8 — PROMOTION GATES CANONIZATION
1. Add to docs/architecture/DEL_AUTHORIZATION_FUNCTION.md (from F3): intent-
   model promotion gates Brier ≤ 0.20, AUC ≥ 0.72, stable lift ≥ 2 purchase
   cycles; models below the bar are advisory-only regardless of merchant
   autonomy level; Brier degradation > 0.20 auto-demotes the dependent
   action classes. [IN BUILD] markers where enforcement is not yet wired.
2. Bind the capabilities doc's Stage 1-4 deployment benchmarks (≥80%
   identity coverage; refutation-passing iROAS with CI clearing hurdle;
   2-cycle revert rules) into the P1/P2 exit criteria in
   docs/roadmap/SIGNAL_SHOPIFY_PHASE_BINDING.md (from F2). Flag owner
   decision 17 (SLA vs. internal target) in the status report.

REPORT: SIGNAL_SHOPIFY_LANE_STATUS_2026-08-11.md to the Drive folder when
F1-F5, F7, F8 complete: artifacts, verification results, the resolved cell
registry, and owner decisions still open (esp. 9, 11, 12, 14, 15, 17).
← All docsView source on GitHub →