Validating Models and Claims

Decide what the available evidence actually supports. Treat a metric, chart, or successful run as evidence for a bounded claim, not as proof that the full system works.

Start with the claim

Write each material claim in a falsifiable form before selecting evidence:

• the decision or action the claim will influence;
• the population, unit of analysis, prediction horizon, and operating conditions;
• the comparator and practical improvement threshold;
• the cost of false positives, false negatives, abstention, and delay;
• the evidence that would refute or narrow the claim.

If these are unresolved and materially change the evaluation, surface the choice instead of silently selecting a convenient definition.

Build an evidence contract

For every claim, map the claim's scope to evidence of equal scope. A component test cannot establish end-to-end readiness; aggregate accuracy cannot establish subgroup safety; retrospective fit cannot establish future performance.

Separate at least these evidence sources when they exist:

1. development evidence used to choose features, thresholds, or models;
2. untouched or independently collected validation evidence;
3. human-reviewed ground truth and its adjudication record;
4. operational evidence produced under the real inference path.

Do not describe reused tuning data as independent validation. Record every contact with the evaluation set, including manual error review that influenced a subsequent model choice.

When label correctness depends on expertise or judgment, preserve the human ground-truth process: reviewer qualifications, instructions, blinding, disagreement rate, adjudication, unresolved cases, and sampling frame. Model agreement with noisy labels is not accuracy against reality.

Validate in layers

Use the smallest set of checks that covers the actual claim:

• Semantic validity: target, timestamps, joins, exclusions, and units mean what the claim assumes.
• Evaluation integrity: split boundaries, leakage controls, comparator, threshold selection, and metric implementation are valid.
• Performance: report uncertainty and practical effect size, not only a point estimate.
• Slices and errors: inspect decision-relevant subgroups, rare cases, temporal drift, missingness patterns, and high-cost errors.
• Operational fidelity: when the claim includes deployment or operational behavior, exercise the same preprocessing, feature availability, threshold, and output interpretation used in practice. Otherwise record why this layer is not applicable.
• Claim calibration: narrow the final language to what all completed evidence supports.

Predefine acceptance criteria when possible. If criteria are chosen after observing results, label them exploratory.

Preserve independence

Prefer a validator who did not construct the evaluated artifact. When that is not possible, create procedural independence: freeze the artifact and protocol, record hashes and versions, run the validation from a clean entry point, and keep exceptions visible.

Never overwrite failed or contradictory results. Explain discrepancies or carry them into the decision as unresolved risk.

Produce a decision record

Use assets/validation-report.md for a reviewable deliverable. Include:

• a claim-to-evidence matrix with status supported, partially supported, unsupported, or not tested;
• the provenance and independence of each evidence source;
• uncertainty, slice results, and representative error cases;
• human ground-truth quality and unresolved disagreements;
• limitations, rejected interpretations, and the exact supported wording;
• a decision recommendation with explicit conditions and owner.

For detailed protocol design, human-label review, and scope-matched proof rules, read references/validation-protocol.md.

Stop conditions

Do not declare validation complete while a material claim lacks scope-matched evidence, evaluation independence is unknown, the ground truth is unreviewed where it matters, or a claimed operational path has not been exercised. Report missing proof or a justified not-applicable layer rather than converting absence of evidence into a pass.