Invariants

Intelligence invariants make every decision reproducible, challengeable, and bounded. They apply whether the output is a ranking, recommendation, downgrade, hold, escalation, or refusal.

Decision invariants

Invariant	What remains reviewable	Observable violation
evidence is referenced, not rewritten	upstream artifact identity and revision	decision record contains altered evidence with no lineage
candidate scope is recoverable	full considered set, exclusions, duplicates, and fingerprints	only the winner is retained
policy is identified	normalized configuration, components, weights, thresholds, constraints, and version	same inputs produce a changed decision through hidden defaults
component meaning is explicit	orientation, scale, missing-data behavior, and contribution	a high score has no stable interpretation
ordering is deterministic	tie-breaking and stable candidate order	equal candidates reverse by input or hash order
adverse evidence can weaken posture	contradictions, falsifiers, instability, and uncertainty reach downgrade, hold, or refusal	every valid input produces a winner
confidence stays inside calibration	corpus, method, interval or category meaning, and drift limits	uncalibrated score is presented as probability
alternatives and regret remain visible	plausible actions, costs, and consequence assumptions	recommendation hides a materially safer alternative
learning is append-only in meaning	prior policy, decision, outcome lineage, new policy, and comparison	adaptation rewrites the rationale for historical decisions
authority remains advisory	approving human or Lab authority and Runtime execution stay separate	recommendation becomes automatic authorization

flowchart LR
    E["evidence revision"] --> C["candidate universe"]
    C --> P["named policy"]
    P --> D["decision and alternatives"]
    D --> H["challenge and calibration"]
    H --> A["bounded advisory posture"]

Decision equivalence

Equal final ranks do not prove equivalent reasoning. Two policies can produce the same winner with different component scores, exclusions, sensitivity, or regret. Review invariants over the entire decision record, not only the selected candidate.

Failure response

When an invariant fails, preserve the inputs and intermediate decision record, then downgrade, hold, or refuse. Do not fill missing evidence with a neutral score, discard contradictory candidates, or manufacture confidence so a positive recommendation remains available.