Verifiable Autonomous Decision Intelligence
A nervous system for your business.
MIZ OKI 3.5 turns consented first-party signals into calibrated hypotheses, governed decisions, and auditable learning — anticipatory intent, with proof of causal lift.
Governing principle: Signals may propose. Causal evidence must qualify. Policy gates authority.
| Component | Type | Status |
|---|---|---|
| ENVELOPE | Evidence | consented |
| SESSION | Sequence | ephemeral |
| HORIZON | Forecast | illustrative · 0.74 ± 0.09 |
| HYPOTHESIS | Intent (TTL-bound) | illustrative · supported |
| VALIDATION | Passport | causal gate closed |
| DECISION | Action | WITHHOLD / NO ACTION |
Every edge carries provenance. Every hypothesis expires.
§01 · The system
Know what changed. Test why. Act within policy.
Not an AI assistant. A governed Decision Control Plane that turns fragmented signals into decisions an enterprise can inspect, reverse, and defend.
Causal credit
Incrementality before attribution. X-Learner / DR-Learner estimation with automated DoWhy refutation and ghost-ad holdouts where supported.
Calibrated confidence
Forecast quality is measured, never proclaimed. Promotion gates: Brier ≤ 0.20, AUC ≥ 0.72, stable across ≥ 2 purchase cycles.
Governed authority
Every decision traceable and permissioned. Observe-only by default; autonomy is certified per account and action class — never a global switch.
§02 · SRPVDAL
From signal to governed action.
One closed control loop. Seven accountable stages. No shortcut from prediction to execution.
SENSE
Capture consented, provenance-bearing first-party interaction telemetry. No consent, no event.
REASON
Forecast defined outcomes and form temporary, uncertainty-bearing intent hypotheses.
PLAN
Generate message, creative, timing, or no-action candidates — never direct activation.
VALIDATE
Test consent, quality, policy, calibration, causality, risk, and rollback. Non-bypassable.
DECIDE
Compare expected incremental value, uncertainty, budget, and downside risk. iROAS, not platform ROAS.
ACT
Execute only through the Decision Control Plane's governed execution adapters, after every gate passes and the exact mutation is authorized.
LEARN
Reconcile verified outcomes, recalibrate, expire hypotheses, preserve evidence.
| Stage | Status | Detail |
|---|---|---|
| SENSE | Complete | Consent + provenance verified |
| FORECAST | Complete | Evaluation hypothesis · example 0.74 |
| CANDIDATE | Complete | Surface risk-reduction evidence |
| CAUSAL | Pending | Experiment required |
| VERDICT | SHADOW / NO ACTION | Withhold execution |
§03 · ORACLE pre-conversion intent — in build
See interaction patterns emerge — before the click.
Session-level inference recognizes changing patterns without converting identity, biometrics, or raw input into a durable profile. Behavior becomes a sequence — not a dossier.
Passive attention
Viewport deceleration normalized against the active session baseline across declared page regions. Behavioral proxy — not physiology.
Session sequence forecast
Ordered, privacy-minimized events estimate a configurable near-term outcome horizon. Sequence-aware, calibration-gated, explanation-bearing.
Creative-semantic alignment
Versioned multimodal asset representations produce bounded creative-compatibility signals. Asset-level — not psychological profiling.
Latent intent bridges
Temporary graph links connect session evidence to a revocable intent hypothesis with explicit TTL. Provenance, uncertainty, expiry — corroborated by governed external evidence (Cell 37).
| T+ | Event | Detail |
|---|---|---|
| 00:01.420 | CONTENT REGION | Session baseline established |
| 00:03.000 | VDI SIGNAL | Technical diagram · deceleration vs. baseline |
| 00:05.110 | TAB HIDDEN | Content-free lifecycle event |
| 00:08.760 | RETURN | Price-region dwell |
| — | EVALUATION HYPOTHESIS | illustrative · 0.74 |
Cell bindings resolve from the live registry — never hard-coded. Incrementality is measured by the causal battery (Cell 36) before any capability claim. Preview · in development.
§04 · Decision Control Plane
Every recommendation carries its evidence.
The ValidationPassport binds evidence, consent, uncertainty, model versions, policy checks, approvals, and rollback into one inspectable record.
| Check | Result |
|---|---|
| CONSENT | PASS |
| DATA QUALITY | PASS |
| CALIBRATION | PASS |
| POLICY | PASS |
| CAUSAL EVIDENCE | EXPERIMENT REQUIRED |
| ELIGIBILITY | WITHHOLD EXECUTION |
Methods: X- / DR-Learner · DoWhy refutation · ghost-ad holdouts where supported · calibration gates (Brier ≤ 0.20, AUC ≥ 0.72).
§05 · Authorized rollout
Two gates. No hidden third.
Safe additive integration
Backward-compatible implementation. Evidence-backed tests. Kill switches. Immutable image. Zero traffic-changing authority.
Narrowly allowlisted evaluation
Consented scope. Strict isolation. Shadow comparison. Guardrails, audit, verified rollback. No customer-impacting mutation.
Production activation
Activation runs only through the Decision Control Plane's governed execution adapters, and requires: separate explicit owner approval per gate; a passing ValidationPassport; qualified causal evidence from a pre-registered holdout; successful experiment and guardrail results; an exact, authorized mutation; and verified rollback.
O-1 · Privacy locked
What the system refuses to learn matters.
Keystroke dynamics are excluded from ingestion permanently — alongside the standing bright lines: no audio, no gaze, no real-time location. Rejection happens at the schema, not in policy documents.
- form_started
- field_focused
- form_completed
- form_abandoned
Field references are coarse, allowlisted, and non-sensitive. Sensitive field classes emit nothing.
- Raw keyboard, input, or composition events
- Entered text, key identity, or edit history
- Dwell / flight time, cadence, repeat, or pressure
- Typing profiles, embeddings, fingerprints, baselines