Decision workflows

Intelligence turns governed evidence into an advisory action while preserving the reasons to disagree. A normal workflow produces more than a rank: it produces policy lineage, challenges, uncertainty, refusal state, and a review packet.

Candidate review

Use candidate selection when multiple protein or assay candidates must be compared under shared metrics.

1. Validate identifiers, sequence, provenance, metrics, and confidence inputs.
2. Apply hard constraints before scoring; retain exclusion flags.
3. Declare ranking factors, directions, weights, and tie-breaking policy.
4. Compute ranked scores and the Pareto frontier.
5. Freeze a proposed shortlist with human_required=true.
6. Build a comparative review packet before progression.

flowchart LR
    candidates["validated candidates"]
    filters["hard constraints"]
    scores["policy-bound scores"]
    pareto["Pareto frontier"]
    packet["comparative review packet"]
    candidates --> filters --> scores --> pareto --> packet

Review factor-level reasons and provenance, not only final ordering. A candidate with high structural confidence but weak empirical evidence should remain distinguishable from one supported by orthogonal observations.

Claim challenge

Before generating a recommendation, evaluate each consequential claim through four independent routes:

Route	Result
support validation	whether cited evidence satisfies declared claim support
contradiction detection	conflicting claim pairs, relationship, and severity
falsifier generation	evidence that would overturn the claim
strong-claim refusal	invalid design, failed QC, weak peptide support, or low localization

Refusal does not delete the claim. It blocks a stronger use and names the minimum evidence required for reconsideration.

Scenario recommendation

Evaluate plausible scenarios separately, preserving action, confidence, hypothesis status, and unresolved questions. Then derive:

• consensus or conflicting actions;
• hold pressure and confidence spread;
• escalation flags and required human arbitration;
• the ordered downgrade chain;
• the final advisory action or refusal.

Counterfactual review should remove or perturb important support—comparators, literature, lab capacity, thresholds, or weights—and record whether the action changes. Sensitivity is part of the result, not a private diagnostic.

Review board

A board packet should contain candidate comparisons, evidence lines, claim challenges, scenario disagreement, unresolved questions, and next-experiment options. Record agenda, votes, abstentions, decision, and evidence freshness. Do not replace the packet with meeting notes.

When advice is promoted to enforced policy, wrap it in an enforced decision support envelope with policy identity, promoting actor, and rationale. Without that promotion record, the output remains advisory.

Feedback and learning

Observed lab outcomes may update confidence, regret, calibration, and future priorities. Append a new decision record linked to the prior recommendation and the new evidence. Preserve what was known, what action was proposed, and what changed afterward.

Completion criteria

A decision workflow is complete when the candidate and evidence inputs are identifiable, ranking policy is recoverable, contradictions and falsifiers are visible, refusal gates were applied, uncertainty and scenario disagreement are preserved, human authority is explicit, and the output can be reconstructed from its typed records.

See decision data contracts and decision artifact contracts for the machine-readable boundary.