Decision review workflow

Intelligence converts evidence into a challengeable recommendation. It does not authorize laboratory work or runtime execution. The workflow is designed to make a polished but weak recommendation lose to a grounded hold, refusal, or discriminating follow-up.

flowchart LR
    A[Versioned evidence bundle] --> B[Evidence posture]
    B --> C{Ready for recommendation?}
    C -->|refused| D[Hold and missing-evidence actions]
    C -->|degraded| E[Human review required]
    C -->|supported| F[Rank candidates]
    F --> G[Stability and skeptical challenge]
    G --> H[Evaluate progression, synthesis, scale-up, redesign]
    H --> I[Consensus, conflict, and unresolved questions]
    I --> J[Advisory decision envelope]
    J -->|explicit governance event| K[Enforced policy]

Build a reviewable recommendation

1. Freeze the candidate cohort and the exact evidence bundle. Record source provenance, freshness, contradictions, and decisive evidence.
2. Run assess_recommendation_readiness(). A blocking contradiction or thin grounding returns a refusal; stale or degraded evidence requires review.
3. Validate the metric catalog and factor weights, then rank candidates with a versioned RankingPolicy. Retain hard-filter rejections and reason codes.
4. Analyze ranking stability under plausible weight changes. A candidate that wins only in a narrow policy region is not robustly preferred.
5. Build the skeptical review and falsification routes. Keep safety, operational feasibility, missing evidence, and seductive weak-signal challenges visible.
6. Evaluate progression, synthesis, scale-up, and redesign independently. Preserve action conflicts, confidence spread, hold pressure, and unresolved questions before deriving the final recommendation.
7. Wrap the result as advisory decision support. Promotion requires a named policy, accountable actor, and rationale.

Route the outcome

Outcome	Next action
progress	Send an explicitly promoted decision and its evidence references to the owning executor.
hold	Preserve the candidate and record the evidence or review condition needed to release the hold.
redesign	Generate a discriminating next experiment tied to a specific blocker or uncertainty.
refused recommendation	Resolve contradiction, freshness, or grounding deficiencies before ranking again.
conflicting scenarios	Escalate the complete review packet to human review.

After observed outcomes return, use the planned-versus-observed learning loop. It records whether the recommendation was confirmed, contradicted, operationally blocked, or still missing an outcome. Learning may change the next recommendation, but it must not rewrite the earlier policy, evidence, or decision history.