Seven stages. One record. No gaps.

Media metrics, commerce events, experience telemetry, inventory movement, economic thresholds.

A qualified signal: what moved, where, by how much, with what evidence quality.

Continuous monitoring, threshold and anomaly detection, noise suppression, evidence assembly with provenance.

Define what "meaningful" means here: thresholds, scopes, and watchlists are business choices, not model defaults.

Detection thresholds are recorded configuration; this stage holds no action authority of any kind.

Blended CPA rises across two campaigns while creative engagement holds steady — a qualified signal opens with the evidence attached.

The qualified signal, operating context, and the organization's decision history.

Ranked candidate causes, each preserved with its supporting and disconfirming evidence.

Hypothesis generation across the whole stack, evidence testing, separating correlation from plausible cause.

Contribute ground truth the data cannot see — deploys, promotions, supplier issues — and challenge any diagnosis.

Eliminated hypotheses are preserved, not discarded; no conclusion is recorded without its disconfirming evidence.

Creative fatigue is ruled out by stable engagement; tracking failure by server-side verification; landing-page degradation survives the evidence.

The diagnosed cause, the business objectives, and the levers actually available.

A counterfactual comparison set: each candidate with its expected effect and stated assumptions.

Candidate generation, expected-effect estimation, surfacing the assumptions each estimate rests on.

Add levers the system cannot see — contract terms, launch timing — and set the weight each objective carries.

Every plan includes the no-action baseline; every expected effect is stated so it can be honestly scored later.

Rebalance, pause, and hold are compared against each other — and against doing nothing.

The candidate actions plus configured policy: financial floors, brand rules, risk boundaries, compliance constraints.

A pass-or-fail record per constraint; vetoes recorded with the exact check that failed.

Mechanical policy evaluation and constraint-conflict detection — the checks run the same way every time.

Own the policy content itself: finance sets the floors, brand sets the rules, risk sets the boundaries.

Margin floors, budget ceilings, brand and risk rules, compliance holds — every check recorded, no veto ever silent.

A candidate that would breach the margin floor is vetoed before anyone sees it as a recommendation — with the failing check quoted.

The validated recommendation and the organization's authority map.

Human approval, delegated approval, automatic approval, rejection, or escalation — with the authority recorded.

Routing, and packaging the complete record — evidence through policy — so the approver decides with full context.

The decision itself, everywhere a standing grant does not explicitly cover it.

Autonomy ceilings, named approvers, and escalation paths are configuration — and every invocation is recorded.

A budget shift crossing the configured ceiling routes to the media lead by name; nothing dispatches until they approve.

The authorized decision, its scope, and the executing system's contract.

A dispatched campaign update, budget change, alert, workflow, ticket, notification — or a recorded decision to hold.

Dispatch, execution monitoring, and rollback readiness.

Manual execution where systems require it, and rollback authority at all times.

Scope bounds enforced at dispatch; the rollback path is recorded before the action runs, not after it fails.

The approved, bounded budget change dispatches with its rollback path stored; "no action" would be recorded with the same rigor.

The prediction recorded at decision time and the realized outcome from the field.

Confidence adjustments, updated future weighting, and lessons written into outcome memory.

Outcome comparison, calibration, and drift detection across decision classes.

Review the lessons; approve any change to thresholds, policies, or models that the lessons suggest.

Predictions are recorded at decision time — never reconstructed afterward — and every learning change is versioned.

A rebalance recovers less than predicted; confidence for that action class is adjusted down, and the next recommendation says so.