Evidence + intent
Consented first-party events, provenance, quality, identity boundaries, and expiring session evidence. Gate: consent fail-closed and O-1 privacy lock.
MIZ OKI SIGNAL / CAUSAL GROWTH CONTROL
Know why performance moved.
Govern what happens next.
Signal connects consented media, site, inventory, CRM, margin, creative, geo, and enterprise-policy evidence in a governed Growth Decision Graph. It ranks the most supportable causal hypothesis, separates forecasts from qualified incrementality, and prepares bounded moves—including no action—inside explicit evidence and authority gates.
MIZ OKI — Operating Knowledge IntelligenceSignal — Causal Growth Control
Illustrative synthetic product experience · Status-aware build blueprint · Human authority remains required
Current posture · Recommend-onlyMEASUREMENT_WRITEBACK · OFFNET_YIELD_WRITEBACK · OFFIntent evidence · Shadow / no actionO-1 privacy lock · Enforced
00 / FOUR LOOPS · ONE CONTROLLER
Every signal enters a loop.
No loop outranks governance.
Evidence, causal measurement, decision control, and independently observed outcomes converge inside the seven-stage SRPVDAL Decision Control Plane.
Causal measurement
Treatment assignment, qualified holdouts, CATE, refutation, uncertainty, and calibration. Forecasts never grade their own incremental effect.
Decision + passport
Ranked hypotheses, chosen and rejected plans, DEL gates, approval, rollback, dispatch or named veto—packaged as a ValidationPassport.
Outcome + learning
Predicted and independently realized outcomes reconcile into decision memory. Activation requires a registered holdout.
One controller: SENSE → REASON → PLAN → VALIDATE → DECIDE → ACT → LEARN. Prediction never grades itself. A registered holdout is required before activation. A score cannot skip Validate or Decide; no-action always remains eligible.
01 / THE BUYER JOB
Four questions.
One accountable answer.
Growth teams do not need another chart. They need a defensible path from a performance change to the next authorized move.
Diagnose
Connect media, site, inventory, CRM, and margin evidence to rank what most plausibly explains the observed change.
Prove
Measure incremental contribution instead of treating platform attribution as causal truth.
Decide
Compare candidate moves against profit, risk, downstream effects, and no-action.
Govern
Apply policy, authority, approval, rollback, and audit controls before dispatch.
02 / GROWTH DECISION GRAPH
Connected data is context.
Connected decisions become memory.
The Growth Decision Graph is the Signal application view of the same temporal-causal, policy-aware MIZ OKI Decision Graph.
01Signal→
02Entity→
03Cause→
04Objective→
05Policy→
06Action→
07Outcome
Campaign → Landing page → SKU → Margin → Policy → Media action → Incremental outcome
Source, time, identity boundary, provenance, quality, consent, and confidence.
Campaign, landing page, SKU, margin, objective, constraint, policy, authority, and risk.
Hypothesis, counterfactual, plan, approval, action, rollback, observed outcome, and lesson.
A conventional graph explains what is connected. The MIZ OKI Decision Graph preserves why a decision was proposed, how it was governed, what was authorized, and what outcome was observed.
03 / DECISION EVIDENCE CONTRACT
Ten questions before
any action.
This is a decision-evidence lens applied after the versioned canonical event schema—not a replacement for that schema. It turns normalized evidence into an auditable decision record.
What changed?
Record the observed physical, digital, or economic state change with source time and provenance.
Which entities?
Bind the affected campaign, landing page, SKU, system, and identity boundary without silent stitching.
What may explain it?
Rank a named, refutable causal hypothesis and preserve plausible alternatives and uncertainty.
Which objective?
Declare the financial or operating outcome, population, and time horizon before choosing a move.
Which policy applies?
Attach consent, authority, budget, margin, legal, security, and action-class constraints.
What could we do?
Generate a bounded candidate set that always includes hold, investigate, and no-action.
What could go wrong?
Expose downside, counterfactual uncertainty, blast radius, reversibility, and rollback readiness.
What did we predict?
Store the ex-ante outcome distribution, assumptions, model version, calibration, and abstention state.
What occurred?
Measure the ex-post outcome against the same definition, population, and horizon used before action.
What must change?
Update decision memory only from qualified evidence, retaining the prior trace and version history.
GATE 01Identity + signature
GATE 02Policy alignment
GATE 03Context sufficiency
GATE 04Hard constraints
Authority contract: a missing required check or hard failure routes to hold, veto, or human review. L0–L5 authority is certified per account and action class; this product experience remains RECOMMEND / NO MUTATION.
04 / ANTICIPATORY INTENT EVIDENCE · PREVIEW · IN DEVELOPMENT
Recognize the sequence.
Do not invent a dossier.
Preview · in development. Intent Engine v2 is presented here as an intended, consented first-party evidence path for Sense and Reason. It never shortcuts Validate, policy authority, or controlled execution.
PassiveAttentionSequence
Consented, first-party viewport deceleration and coarse lifecycle events are normalized to the active-session baseline; raw sequence buffers purge when the session ends.
SessionOutcomeForecast
An ordered event stream estimates a configured near-term outcome with calibration, drift, and abstention metadata; state ends with the active session.
CreativeSemanticProfile
Versioned asset representations expose bounded visual and semantic compatibility without inferring a psychological identity, and expire with the creative lifecycle.
IntentHypothesis
Temporary graph bridges join provenance-bearing session evidence to a revocable hypothesis with explicit uncertainty, strict TTL, and erasure-cascade support.
ValidationPassport
Validated consent, quality, calibration, policy, causal evidence, authority, rollback, and disposition fields feed the governed decision trace.
00:01.420CONTENT REGIONsession baseline
→00:03.090VIEWPORT DECELERATIONcoarse region only
→00:05.110TAB HIDDENcontent-free event
→00:08.760RETURNsequence restored
→EXAMPLE TTLIntentHypothesisillustrative · expires
O-1 / PRIVACY LOCK
Keystroke dynamics are prohibited.
Raw keys · text · cadence · dwell or flight time · typing profiles · audio · gaze · fine individual geolocation · sensitive-trait inferenceDISALLOWED_KEYSTROKE_DYNAMICSVALIDATION PASSPORT
Prediction is advisory. Authority remains gated.
Consent failure means no processing. Sensitive-category inference is denied at ingest and hypothesis creation. Probabilistic identities do not enter causal-effect math.SHADOW / NO ACTIONRetention boundary: session buffers purge at termination; session forecast state is session-only; asset profiles expire on creative retirement; hypotheses expire by strict TTL; erasure cascades through every relevant store. No persistent psychological dossier is created.
05 / GOVERNED DECISION TRACE
Do not punish the wrong campaign.
Choose an illustrative incident. The trace changes, but the evidence and authority gates do not.
OBSERVED INCIDENT
Acquisition cost rises while conversion falls.
- Auction pressure remains stable
- Mobile checkout completion weakens
- Eligible inventory remains available
CAUSAL HYPOTHESIS
Checkout latency explains more of the observed loss than bid inflation.
SUPPORTED · EXPERIMENT PENDINGGOVERNED PLAN
Hold media changes. Repair the mobile path and run a randomized recovery test.
HUMAN APPROVAL REQUIREDSense
Capture the selected incident and its provenance-bearing cross-stack evidence.
Reason
Rank named causal hypotheses and expose uncertainty instead of assigning automatic blame.
Plan
Compare hold, repair, reallocate, suppress, investigate, and no-action candidates.
Validate
Check incrementality, economics, quality, consent, risk, policy, and rollback readiness.
Decide
Approve, route to an operator, keep in shadow, or veto within explicit authority.
Act
Dispatch only an exact, scoped, authorized mutation through the existing action pathway.
Learn
Compare predicted and verified outcomes and write the complete trace to decision memory.
Illustrative composite scenarios—not production telemetry or customer results, and never a guaranteed outcome.
05B / VALIDATIONPASSPORT · IN BUILD
Every proposal carries its evidence, alternatives, authority, and outcome contract.
The public example is synthetic. It shows the trace package being built—not live tenant data and not a fabricated realized outcome.
Hard rule: a DEL score never overrides a failed hard constraint. Missing evidence routes to hold, named veto, or human review.
Decision + job
decision_id · synthetic tenant · loop_id · job_id
Signals + hypotheses
signal_refs · ranked hypotheses · provenance · confidence and uncertainty
Selected + rejected plans
Chosen option, bounded alternatives, no-action, and the explicit reason each option advanced or failed.
DEL + authority
Gate results, DEL breakdown, autonomy level, approval record, named constraint, and rollback reference.
Dispatch or veto
action_dispatch_or_veto records the governed ending; no silent mutation and no authority inferred from a model score.
Prediction → independent outcome
Measurement window, realized outcome when independently available, learning-update reference, and SHA-256 integrity.
06 / HIGH-VALUE DECISION JOBS
Make the decision contract
explicit before the model runs.
Each job declares the signal, eligible decision, governing constraint, evidence required, and measurable outcome.
J-01
Incrementality
- Signal
- Attributed growth moves
- Decision
- Test contribution before reallocating
- Constraint
- Qualified support + holdout eligibility
- Evidence
- Treatment-effect estimate and refutation
- Outcome
- Incremental contribution
J-02
Waste prevention
- Signal
- CPA rises or conversion weakens
- Decision
- Hold, repair, suppress, or investigate
- Constraint
- No automatic campaign blame
- Evidence
- Cross-stack causal hypothesis
- Outcome
- Avoided inefficient spend
J-03
Margin control
- Signal
- Revenue and profit diverge
- Decision
- Model a margin-aware allocation
- Constraint
- Finance baseline + SKU eligibility
- Evidence
- Contribution economics
- Outcome
- Incremental contribution margin
J-04
Learning stability
- Signal
- Platform feedback becomes noisy
- Decision
- Stage or withhold a change
- Constraint
- Experiment integrity + rollback
- Evidence
- Calibration and drift record
- Outcome
- Reliable learning velocity
J-05
Executive defensibility
- Signal
- Competing reports disagree
- Decision
- Preserve one governed decision trace
- Constraint
- Evidence provenance + authority
- Evidence
- ValidationPassport + ledger
- Outcome
- Auditable investment rationale
J-06
Team leverage
- Signal
- Recurring analysis consumes operators
- Decision
- Automate evidence assembly, not authority
- Constraint
- Human approval + scoped capability
- Evidence
- Reusable graph and outcome memory
- Outcome
- Faster governed decisions
07 / COMPETITIVE BOUNDARY
More than measurement.
Less than reckless automation.
Signal earns authority from evidence, not from a model score alone.
Beyond attribution alone
Reporting describes credited performance. Signal also tests causal contribution and preserves the supporting evidence.
Beyond isolated bidding alone
Signal reasons across media, site, inventory, CRM, margin, and policy before an authorized platform action.
Beyond profile unification alone
Identity and activation plumbing do not replace a governed cross-stack decision record or outcome memory.
Beyond personalization alone
Message selection becomes one bounded input to incremental economics—not a psychological profile or direct actuator.
The defensible distinction: cross-stack causal diagnosis + incrementality evidence + policy-gated operational action + outcome memory.
08 / CAUSAL CREDIT
Forecasting is not incrementality.
Signal keeps three questions separate: what outcome may occur, who is likely to receive a treatment, and what the treatment changes. Only qualified treatment-effect evidence can support incremental-value claims.
SESSION OUTCOME FORECASTAdvisory probabilityCalibration · drift · abstention
TREATMENT ASSIGNMENTExperiment eligibilityConsent · exclusion · support
TREATMENT EFFECTQualified incrementalityEstimator selection · refutation · holdout
DECISION SETNo-action remains eligible.Risk · budget · policy · authority
Estimate one defined intervention.
Use as an anchor only when assignment, power, contamination, and outcome integrity support the population and test window.
Estimate where effects may vary.
Select S-, T-, X-, or doubly robust approaches by assignment, overlap, imbalance, sample support, and diagnostics. No learner is uniformly best.
Inform the longer allocation horizon.
Calibrate aggregate channel estimates against qualified experiments where definitions align; never present MMM as user-level causal truth.
Triangulation rule: align the treatment, outcome, population, and time horizon before reconciling results. Never average incompatible estimands. When qualified evidence conflicts, abstain and investigate.
08B / NET YIELD · IN BUILD
Revenue is not contribution.
Preview · in development. The economics layer is designed to preserve missingness instead of inventing profitability.
Observed revenue→Source transaction evidence with order and SKU lineage
Refunds + observed returns→Actual-only until at least one observed return cycle exists per SKU
Fulfillment + required costs→Missing required costs leave a row incomplete and excluded; costs are never invented
Net contribution evidence→Recommendation input only · NET_YIELD_WRITEBACK remains OFF
Control boundary: no claim that the platform currently bids on net contribution. Writeback requires its own verified pilot and separate approval.
09 / EVIDENCE MATURITY
Demonstrate the logic.
Validate the outcome.
A polished scenario is not a customer result. Authority expands only when customer-specific evidence supports it.
DEMONSTRABLE LOGIC
Enter with synthetic or replayable scenarios. Exit only when schemas, deterministic traces, vetoes, abstention, and rollback behavior are reproducible.
PILOT VALIDATION
Enter observe-only with customer-specific data. Exit only when data quality, causal design, calibration, contribution economics, and policy checks qualify.
GOVERNED EXPANSION
Enter with qualified pilot evidence. Advance an account × action class only after security, authority, limits, approvals, monitoring, and rollback are proven.
Illustrative figures and scenarios demonstrate intended product logic—not production telemetry or customer results—and require customer-specific validation during implementation.
10 / ADVANCED CONTROL FRONTIERS · IN BUILD
Five expansions.
Five different gates.
These frontiers have advanced from roadmap to governed implementation. They are not live capabilities, customer results, permission to act, or evidence that an activation gate has cleared.
Creative component effects
Embeddings describe promotion framing, layout, copy length, and imagery. Only randomized or otherwise qualified variation may identify causal effects.
Every estimate remains provisional until configured volume is met. Generated creative always requires human approval.Multi-quarter value
Balance immediate conversion against discounted contribution margin and repeat purchase across declared horizons.
Publish only with at least two observed quarters per cohort; otherwise return DATA_INSUFFICIENT.Inventory + fulfillment
Stock, safety thresholds, holding cost, and regional capacity may enter the graph as recommendation evidence.
No F3-originated action can reach a platform adapter, regardless of flags.Micro-geo calibration
Owner-configured aggregate geographies, exclusions, donors, and caps may support a bounded reservation proposal.
Every reservation halts for Level 2 approval. Two clean cycles plus an owner-approved config change are required before bounded-authority eligibility.Treasury-gated spend
Validated liquidity floors, covenant limits, and tightening curves may become named hard constraints in Validate and Decide.
Missing or invalid configuration means NOT_CONFIGURED—never an invented constraint or silent default.Expansion workflow: merged artifact → registry verification → synthetic replay → qualified pilot → policy, security, and rollback review → separately approved account × action-class authority. MEASUREMENT_WRITEBACK and NET_YIELD_WRITEBACK remain OFF.
11 / COMMERCIAL ENTRY
A 90-day Causal Growth Control Pilot.
One brand. Limited, agreed source systems. Two consequential decision classes. Media, site, inventory, CRM, and margin inputs as available. Customer-specific evidence and economic baselines.
Begin observe-only. Advance to human-approved actions only after evidence quality, economics, and policy boundaries are validated. Make a separate expansion decision at completion.
Verified pilot numbers—not marketing copy—are what may later flip Preview labels. Real tenant selection, costs, treasury constraints, geo candidates and caps, creative volume, and holdout registration are owner inputs.
Design a 90-day growth pilot ↗Observe
Connect agreed sources and define two decision classes. Exit with provenance, consent, completeness, latency, and baseline economics documented—without mutation.
Validate
Pre-register qualified tests, select estimators from assignment and support, calibrate forecasts, challenge policies, and reconcile contribution economics.
Recommend
Surface human-approved proposals with ValidationPassport records, verified rollback paths, independently observed outcome comparisons, and an explicit expand, hold, or stop decision.
BUILT ON MIZ OKI
MIZ OKI Signal is powered by the MIZ OKI Decision Graph and governed runtime.
Signal is the first deeply productized commercial application. The platform operates above systems of record and shares temporal evidence, causal reasoning, policy authorization, action traces, and outcome memory across domains.
MIZ OKI SIGNAL
Build qualified evidence for why performance moved.
Govern the next move.
Run one complete decision trace, define the evidence required, and start in observe-only mode.