Change Principles

An Intelligence change is safe only when a reviewer can determine whether it changed evidence use, candidate eligibility, decision policy, uncertainty, refusal, explanation, or downstream authority. Schema compatibility alone is not enough: a recommendation can reverse while the payload shape stays fixed.

Classify the decision change

Changed surface	Questions to answer	Required evidence
candidate or filter	Who entered, who was excluded, and can ordering or selection change?	fixed cohort, rejection cases, and before/after ranking
score or orientation	What does the value mean, in which direction, on which scale, and under missingness?	boundary values, monotonicity, ties, and missing inputs
weight or threshold	Which decisions move under the new policy?	policy identity, sensitivity envelope, and decision diff
confidence or calibration	Which corpus and outcome definition support the value?	calibration, drift, subgroup, and abstention evidence
contradiction or falsifier	Can favorable evidence bypass challenge?	adverse, ambiguous, and unresolved cases
scenario or regret	Which alternatives, costs, and uncertainties are modeled or omitted?	counterfactual and consequence fixtures
refusal or hold	Which blocker stops progression, and can a caller flatten it?	refusal reason, consumer branch, and rendered packet
promotion	Who grants authority, to what target, under which policy?	attributable promotion and unauthorized-path rejection
learning	Which outcomes are eligible to alter future policy?	lineage, eligibility, audit, and preserved prior decision

flowchart LR
    change["decision change"] --> fixed["fixed evidence and candidates"]
    fixed --> compare["compare policy outcomes"]
    compare --> challenge["adverse and boundary cases"]
    challenge --> authority["verify advisory and enforced separation"]
    authority --> record["retain decision diff and limitations"]

Invariants

• evidence identity and candidate scope are explicit before ranking;
• score orientation, missingness, ties, and thresholds never rely on caller inference;
• challenge evidence can downgrade, hold, or refuse a favorable result;
• confidence is not described as probability beyond its calibration evidence;
• rejected candidates, alternatives, and unresolved questions remain visible;
• advisory output cannot become enforced through serialization or transport;
• new evidence or policy supersedes rather than rewrites history; and
• feedback changes future policy only through an attributable review record.

Review the same evidence and candidate set under both policies whenever a change can affect judgment. If the recommendation moves, publish the decisive factor and downstream consequence. If the proof does not cover the new policy envelope, narrow the posture or refuse it.