Error Model

Intelligence outcomes distinguish invalid evaluation, governed refusal, uncertainty-driven hold, and successful recommendation. Only the first is necessarily a software or contract error.

Outcome	Meaning	Required evidence
Invalid input	Candidate, policy, scenario, or evidence contract cannot be evaluated	Validation issues naming the affected field or record
Refused claim	A strong claim fails design, QC, peptide-support, or localization thresholds	Stable refusal reason and minimum missing evidence
Hold	Evaluation is valid, but contradiction, uncertainty, or evidence readiness blocks action	Hold pressure, downgrade chain, unresolved questions, and gate result
Human escalation	Scenarios conflict or confidence spread exceeds policy	Escalation flags and the competing scenario outcomes
Degraded recommendation	A recommendation exists with material caveats	Degradation reasons and required review
Recommendation	Policy supports one action	Evidence snapshot, candidate cohort, policy, sensitivity, and rationale

flowchart TD
    E[Evaluation request] --> V{Contracts valid?}
    V -->|no| X[Validation failure]
    V -->|yes| G{Evidence gate}
    G -->|refused| R[Refusal or hold]
    G -->|degraded| D[Conditional recommendation and human review]
    G -->|supported| S{Scenario agreement}
    S -->|conflict| H[Escalate]
    S -->|agreement| C[Recommendation]

The absence of a recommendation is not a crash, and a completed calculation is not necessarily decision success. Errors must not be “handled” by substituting neutral scores, dropping candidates, resolving ties arbitrarily, or suppressing missing evidence. Those behaviors would turn uncertainty into false precision.

Route the outcome to its authority

Outcome	Accountable response	Closure evidence
invalid candidate, policy, scenario, or evidence contract	data or policy owner corrects the named contract	corrected identity plus validation result; unchanged retry is not closure
evidence refusal	Knowledge or scientific owner supplies the missing burden or accepts a narrower claim	new evidence-bundle identity, sufficiency decision, or explicit refusal acceptance
hold	decision owner waits for the named resolving event	hold condition, expiry or review trigger, and resulting disposition
human escalation	named domain, safety, cost, or operational authority decides the contested dimension	competing outcomes, owner, rationale, conditions, and decision time
degraded recommendation	consumer preserves caveats and satisfies required review before use	accepted limitations, allowed use, reviewer, and downstream restrictions
recommendation	accountable human or governed promotion policy decides whether action may proceed	recommendation identity, authority decision, conditions, and handoff target

No row authorizes Intelligence to manufacture missing evidence or operational approval. If the named authority does not respond, the durable state remains a hold, escalation, or refusal rather than silently aging into approval.