Claim Evidence Record
A claim evidence record links a public claim to the evidence that supports it. It is the bridge between Experiment lifecycle, Run artifact, Result interpretation, and paper-style Evidence table writing.
A claim should be represented as:
where the statement is the narrow claim, scope states when it applies, evidence lists the supporting runs or sources, and limitations state what the evidence does not prove.
Evidence Chain
For an experimental claim:
For a literature claim:
In both cases, the claim should become weaker when split details, baselines, uncertainty, or artifact availability are missing.
Record Fields
- Claim: one narrow sentence.
- Scope: dataset, task, model family, metric, split, and deployment assumption.
- Evidence source: run identifier, public artifact, paper figure, table, or proof.
- Data: dataset version, lineage, split, and preprocessing contract.
- Metric: definition, aggregation rule, confidence interval or seed variance when available.
- Baseline: comparison that makes the claim meaningful.
- Failure modes: known errors, hard slices, and negative results.
- Limitations: what the evidence does not establish.
- Public boundary: what was omitted because it is private, unpublished, or unsafe to disclose.
Claim Strength
A lightweight evidence strength checklist is:
This is not a real score. It is a reminder that unsupported claims should stay narrow.
Checks
- Is the claim narrower than the evidence?
- Is the split and data lineage clear enough to interpret generalization?
- Is the metric aligned with the task and decision?
- Is there a baseline, ablation, or previous run for comparison?
- Is uncertainty, seed variance, or confidence interval needed?
- Are missing artifacts marked
not released,to verify, ornot applicable? - Does the public record avoid private paths, internal task names, collaborator details, and unpublished results?