Change validation

Validate Core changes against the scientific sentence they alter. The proof must widen when a change moves from implementation detail to public model, algorithm, artifact, workflow-family behavior, or benchmark posture.

Change-to-proof map

Change	Required proof	Claim review
parser or adapter	accepted and rejected fixtures, mutation cases, producer/version provenance, stable export	native versus imported language
domain model or lifecycle	invariants, forbidden combinations and transitions, schema and consumer compatibility	public contract meaning
score, FDR, threshold, or inference	reference values, orientation, boundaries, ties, missingness, decoys, sensitivity	acceptance and confidence language
quantification or normalization	design, scale, zero/missing distinctions, batch, ordering, reproducibility	transfer and quantitative truth
PTM, DIA, or targeted rule	ambiguity, library or panel context, calibration, interference, acceptance	family-specific ceiling
benchmark asset or acceptance bar	license, lineage, freshness, holdout role, expected failure, generated artifacts	release-readiness matrix
performance implementation	serial equivalence, determinism, exhaustion, cancellation, partial output	no scientific claim widens from speed alone
public CLI, API, or artifact	parity, schema, round trip, errors, lineage, downstream reader	public examples and compatibility

Validation route

flowchart TD
    E["Core edit"] --> S["name scientific sentence"]
    S --> I["identify invariant and failure modes"]
    I --> F["focused positive and negative proof"]
    F --> B{"benchmark or transfer claim?"}
    B -->|yes| H["primary, companion, holdout, acceptance"]
    B -->|no| X["boundary and consumer review"]
    H --> X
    X --> C{"evidence supports sentence?"}
    C -->|yes| V["validated"]
    C -->|no| N["narrow, refuse, or add evidence"]

Inspect accepted, rejected, failed, and missing counts—not only the headline result. Compare policies and provenance as well as numeric output. Run affected Runtime, Knowledge, Intelligence, or Lab boundary tests when the artifact crosses package ownership.

Validation record

State input level and provenance, scientific policy, expected result and negative disposition, reference or benchmark independence, tolerances, workflow family, transfer envelope, exact checks, and resulting public claim. If a relevant corpus or consumer was not run, preserve that gap explicitly.