Risk Register

Intelligence can remain operational while becoming untrustworthy: rankings still sort, recommendations still serialize, and review packets still render. The critical risks are therefore silent changes to evidence use, policy, calibration, explanation, refusal, and downstream authority.

flowchart LR
    E["evidence and results"] --> P["declared policy"]
    P --> J["judgment"]
    J --> C["confidence and rationale"]
    C --> D{"recommend or refuse"}
    D --> Q["consequence review"]
    F["feedback"] --> P

Active And Structural Risks

Risk	Observable failure	Required control
duplicate semantic ownership	the same belief-audit model exists in Core and Intelligence	resolve the active ownership blocker or govern the exact shared contract
evidence laundering	a score appears authoritative without source claim and lineage	require Knowledge references and expose missing support
policy drift	threshold or default changes move recommendations silently	version policy and compare decision outcomes
score orientation error	larger/smaller or sign meaning reverses a rank	type and test orientation at the boundary
missingness coercion	absence becomes zero, neutral, or low confidence implicitly	declare missingness semantics and refusal behavior
unstable ranking	ties or floating-point differences reorder candidates	deterministic tie policy, tolerances, and retained comparison
confidence miscalibration	high confidence persists under weak or shifted evidence	calibration and counterfactual challenge corpus
circular benchmark	the same evidence tunes and validates policy	separated challenge evidence and provenance
rationale decay	output remains usable but no longer explains decisive factors	structured rationale and review packet checks
refusal erosion	downstream demand converts blockers into warnings	refusal tests and release-language guards
consequence bypass	recommendation is treated as a laboratory instruction	mandatory Lab feasibility and consequence handoff
feedback contamination	outcomes adapt policy without eligibility or provenance checks	governed learning inputs and adaptation audit

Review Priority

Prioritize risks that can change a recommendation without changing its schema. Those failures are hardest to detect through ordinary compatibility tests. Challenge policy changes with unchanged inputs, missing evidence, ties, contradictions, adversarial counterfactuals, and downstream infeasibility.

A recommendation is trustworthy only when a reviewer can reconstruct its inputs, policy, alternatives, confidence, rationale, and refusal boundary.