MIZ OKI SIGNAL — FEATURE CATALOG AND DEMO PLAN r2.1

Merges: FEATURE_CATALOG_v1.md (customer feature catalog) + DEMO_PLAN_r2.1.md (live demo arc) into one reference, plus the extended-capability additions landed in MIZOKI_MARKETING_AUTOMATION_WHITEPAPER_r2.0_SEP2026.md §13 (originally amendment r1.3, now folded into r2.0). Base files stand; this is a merge for reference convenience, not a supersession. Sources: whitepaper r1.2 · r3.5.1 + r3.5.2 · Growth Control r2.0 · completion/activation/OPT-GOV build reports · F4_GO_LIVE_2026-08-24 · PILOT_REPORT_SPEC_v1 · SHOPIFY_INSTALL_ACCEPTANCE_v1 · docs/reports/PRODUCT_LAUNCH_V5.3_COMPLETION_2026-08-25.md. Statuses honest; nothing here upgrades a label.

Part 1 — Live demo access

https://mizoki3.com/demo — HTTP 200, six interactive demo experiences confirmed live in the served markup (live-verified, 2026-08-25):

Demo	Maps to
Nexus Run	full SRPVDAL pipeline walkthrough
Signal Factory	live public demo — raw connector events → canonical → gate → all seven SRPVDAL stages, including one deliberate guardrail block caught in red; no signup
Counsel Room	governed-vertical generalization — a legal-domain scenario walkthrough presented strictly as roadmap/architectural proof-of-concept [PARTIAL — roadmap framing only]; no legal-domain data path, connector, or decision job is built, piloted, or shipped
Estate Room	CRE/real-estate decisioning scenario
Capital Desk	treasury/F5 covenant-veto walkthrough
Risk Sentinel	guardrail/refutation watch scenario
Executive Demo (`/media/demo`; `/media/executive-demo` 308s there)	the five-act arc (Part 3) as an interactive product surface — Lie · Signals · Proof · Profit · Restraint · 90 Days, single file, presenter talk track, 5-min cut, tenant scale; SYNTHETIC watermark; every panel status-labeled from OFFERING_MAP v2.3 §B.6 — LIVE once the homepage lane deploys it; until then merged, not deployed

Part 2 — Customer feature catalog

A. Shipped / operating — sell today, labeled

#	Customer outcome	How	Why it works	Status
1	Stop paying for conversions that would have happened anyway	Qualified + geo holdouts, ITT, CATE learners (X-Learner/DR-Learner), DoWhy refutation	Caused vs anticipated split; platforms grade themselves — we don't let them	PARTIAL
2	Know your real profit per campaign, not platform ROAS	Net Yield: revenue − refunds/returns − COGS − fees − fulfillment	4× ROAS at 40% returns can lose money; only commerce truth reveals it	IN BUILD
3	See buying intent before the click	Cells 33–36: VDI, dwell, sequences → SessionOutcomeForecast + hesitation head	Micro-behavior predicts state transitions; content-free, session-bounded, explained — never a bare score	IN BUILD/shadow
4	Show the ad each visitor actually responds to	CreativeSemanticProfile + aesthetic-affinity vectors → matched creative rank	Interaction history maps to creative embedding space; alignment beats rotation	IN BUILD/shadow
5	A system that refuses to act on thin evidence	DCP: passport → DEL score → Clipped-ReLU authorization → WITHHOLD verdicts	Zero authority below evidence threshold; hard clip at granted ceiling	PARTIAL
6	Always-on incrementality without blackout tests	F4 micro-geo calibration — LIVE since 2026-08-24	Continuous small experiments replace quarterly big ones; every reservation human-approved until 2 clean cycles	LIVE
7	Never advertise what you can't ship	F3 inventory sync: stockout → throttle-rec, overstock → accelerate-rec; stockout also enters the DEL veto directly	Recommendation-only by invariant; spend follows stock reality	PARTIAL
8	Spend within CFO-approved liquidity, automatically	F5 treasury gating: floors/covenants tighten caps; breach → named veto; atomic reservations	Two simultaneous approvals can't jointly breach a floor	PARTIAL
9	Spend to the profit-maximizing dollar, not the budget	Continuous dosage curves: τ(d)·margin = marginal cost, with CIs and support bounds	Marginal thinking on real margins; refuses to extrapolate beyond observed spend	PARTIAL
10	Privacy as a security asset	O-1: keystroke/audio/gaze schema-rejected; consent fail-closed; erasure cascade; session-bounded retention	A capability the schema refuses can't be subpoenaed, breached, or demanded	LIVE
11	Every decision reconstructable for your board	ValidationPassport chain (KMS-signed once wired): evidence, alternatives, rejected options, authority, rollback	Audit-grade defensibility; the analyst never needs the person who decided	PARTIAL
12	90 days from install to governed control	Pilot: Observe → Validate → Recommend; autonomy only after Brier ≤.20, AUC ≥.72, 2 clean cycles	Trust is earned in sequence — the sequence is the product	PARTIAL
13	Shopify: install → Day-1 in one sitting	OAuth install, pixel extension, commerce sync, COGS worksheet, honest `not_configured`	All net-yield inputs already live in Shopify	per Part D — OAuth install flow and Pub/Sub topic remain open
14	Integration without vendor lock-in	8 shipped `intent_*` tools — explain, score, compare/cohort, history, taxonomy, consent stats, incrementality, status — deterministic queries over the decision graph	Any BI tool, notebook, or pipeline calls the same typed endpoints the platform's own dashboards use	LIVE

B. Latent — already in the machinery; needs surfacing, not building

#	Latent product	Source (exists)	To present it
L1	Profit-Leak Report — SKUs advertised into a loss after returns	net yield + returns + J-03 registry	One generated report view
L2	Overcharge Statement — conversions billed as caused that we classify anticipated	Causal Credit Ledger	Quarterly statement template
L3	Hesitation Segment — high-VDI-no-click visitors	I-02 hesitation head (shadow)	Surface as segment + posture rec
L4	Checkout-Waste Alarm — always-on CPA-rise monitor	SIG-042 harness (OPT-GOV P4) + J-02	Schedule against live telemetry
L5	Restraint Ledger — decisions stopped this month, with reasons	WITHHOLD verdicts + passports	Report section; no competitor can show refusals
L6	Emerging-Demand Clusters	Cell-35 (LatentBridge), hypothesis-only	Digest view, deny-list screened
L7	CFO Mode — treasury posture on the weekly report	F5 + report spec §5	One report block
L8	Creative Fatigue Early-Warning	creative vectors + joint-DML ext	Threshold + report line
L9	LTV-Steered Bidding Preview (shadow)	F2 + LTV bid-weighting (flag-dark)	Shadow comparison view
L10	Autopilot Pacing (built dark)	growth-scheduler	Roadmap slide only

C. Platform machinery behind the catalog

SRPVDAL decision flow: Sense → Reason → Plan → Validate → Decide → Act → Learn — every LLM call routes through virtuoso_call(Role.X), no hardcoded model strings outside the registry.
13 execution adapters (+1 alias) across meta_ads and google_ads, as last measured on revision 00022-c5l (2026-08-16) — execution disabled everywhere (enabled=false, every EXECUTION_ADAPTER_* flag false), Stage-4 human-gated.
Causal inference: meta-learners (X-Learner/DR-Learner), DoWhy refutation battery, causal discovery, and RL-adjacent dosage curves (continuous cross-fitted R-learner).
Intent scoring (LII/ORACLE): Cells 33–36 — ingest, scoring, graph, causal credit — observe-only, model quality gated on real forward labels (see Completeness Audit v1.0 §2).
F4 growth control: geo-SCM scheduler, synthetic-control donor selection, perturbation-window scheduling — live for the mycocoons pilot.
GraphRAG/CausalRAG: retrieval over the intent/decision graph respecting causal edges; powers explanation payloads accompanying every score.
Governed build fleet: claim-first coordination ledger, skill parity checks, canon linter, adversarial verifiers, typed approval strings — the same gates applied to the platform's own development process.
8 intent tools: see Part 2A row 14.

Part 3 — The 20-minute demo arc (five acts)

Sources & provenance. All figures shown are the synthetic demo tenant's, watermarked SYNTHETIC on screen — including Act 1's ROAS-vs-contribution contrast. Act 2's session trace is the presenter's own consented device session, stated aloud. F4 demo reservations run only against a sandboxed demo tenant with zero-spend caps and demo-only geos — never the pilot tenant, never real spend. Demo tenant constants verified in the Tier 2 smoke run: MEASUREMENT_WRITEBACK=false, NET_YIELD_WRITEBACK=false, F4_CALIBRATION=false (demo tenant, distinct from the mycocoons pilot's true), EXECUTION_ADAPTERS_ENABLED=false, autonomy L0, all spend caps zero, demo geos DEMO-XX/DEMO-YY disjoint from pilot US-CA/NY/FL/TX.

Act 1 — The Lie (3 min). Platform dashboards vs bank account: same campaign, 4× reported ROAS, negative contribution after returns. Close: we optimize the number your bank account reports.

Act 2 — The Anticipation (4 min). Live session trace → VDI/dwell tokens → forecast with attention summary → hesitation flag → matched-creative rank (logged, not applied — say so). Then the O-1 moment: submit a keystroke-dynamics payload on stage → DISALLOWED_KEYSTROKE_DYNAMICS.

Act 3 — The Proof (4 min). Register a holdout (declared estimand, power) → CATE estimate → refutation battery with one test failing and the estimate WITHHELD.

Act 4 — The Profit (4 min). Net-yield waterfall with one row excluded and the missing cost named → dosage curve ("your profit-maximizing spend is $X, not your budget," CIs on screen) → J-03 losers list.

Act 5 — The Restraint (4 min). A decision runs the DCP: passport → DEL score plotted on the Clipped-ReLU curve → below threshold → WITHHOLD with named reason. Then the live F4 reservation halting at approval [LIVE — labeled]. Close: Restraint Ledger + the deploy-time canon-correction story (the gate corrected the platform's own whitepaper's stale Neo4j reference — "even our docs pass through the gate").

Leave-behind (1 min): sample weekly pilot report (ILLUSTRATIVE watermark) + 90-day pilot one-pager. A 5-minute cut = Acts 1+5 only for exec audiences.

Build deltas already closed since r1.0

SIG-042 passport harness now exists (OPT-GOV Phase 4) and is wired to the demo surface — Tier 1 deterministic replay verified in the Tier 2 smoke run (seal intact, chain valid, DEL score 86.0, not executed).
F4 is live — the restraint act shows a real reservation halting at real L2 approval.
Rejection trio (O-1, consent-off, sensitive bridge) verified live: 3/3 rejected.
Demo-tenant isolation proof exists: 16/16 isolation tests green (config, allowlist, actuator, adapter defaults, F4 isolation, geo disjointness).

Guardrails (retained verbatim)

Label everything; no fabricated realized outcomes; no hidden happy path; failure is content; ledger discipline. Anything shown from shadow/dark surfaces carries its flag state on screen; the anticipation act uses canon framing — anticipatory intent evidence from content-free signal, never mind-reading language; content_qa gates the demo script itself.