Extensibility Model

Intelligence extends through explicit analytical policy and review contracts. A new ranking factor, scenario, interpretation, or recommendation is acceptable only when readers can inspect what evidence it uses, how it changes judgment, which uncertainty it preserves, and when it refuses to decide.

Extension routes

Need	Extend	Required evidence
New candidate property	candidates/	validation, provenance, lifecycle effect, and non-opaque ranking use
New ranking factor	candidates/ranking.py or candidate review policy	named factor, normalization, weight, contribution rows, tie behavior, sensitivity, and stability
New claim type	claims/, falsifiers.py, contradictions.py, and refusal.py	graph support rule, adverse-evidence rule, falsifier, refusal threshold, and belief audit
New analytical interpretation	interpretation/	typed upstream artifact, explicit limits, supporting IDs, cautions, and no new source curation
New decision scenario	judgment/scenarios.py	stable policy identifier, required metrics, action set, reasons, uncertainty, and escalation rule
New review artifact	reviews/	identifier alignment, provenance, alternatives, caveats, known exclusions, and validation
New learning behavior	learning/	attributable outcome, prospective effect, drift report, and immutable prior decision

Policy design

Policy must be data, not scattered conditionals. Give it a stable identifier, typed thresholds and weights, deterministic ordering, and an explanation for every factor. Emit the enforced policy with the result. Test boundary values, ties, missing evidence, contradictory evidence, and changes to one factor at a time.

A ranking extension is incomplete if it exposes only a total score. It must produce factor-level contributions, rejection or eligibility reasons, and robustness evidence. A scenario extension is incomplete if it cannot express hold, redesign, or escalation alongside favorable actions.

Analytical safeguards

Every strong claim needs graph-backed support, contradiction handling, a falsifier, and a refusal boundary. High-confidence claims must remain in belief audits rather than bypass them. Interpretation modules may project evidence into decision context, but may not manufacture curated facts or downgrade an upstream QC failure.

Benchmark new judgment against blinded cases, counterfactual changes, sensitivity, known decision traps, and a simpler baseline. Compare not only agreement but regret, confidence calibration, stability, refusals, and the quality of the explanation.

Extension smells

• a new “AI” or score abstraction with no owned decision semantics;
• free-text evidence not linked to governed identifiers;
• a factor whose scale or direction is implicit;
• confidence increased by missing evidence;
• adverse evidence removed during synthesis;
• recommendations that contain executable lab instructions; or
• adaptation that mutates historical policy or decision artifacts.

An extension is complete when policy lineage, reasoning, uncertainty, refusals, benchmark behavior, review output, and downstream authority remain inspectable as one decision path.