evaluation-anchor-checker_skill

Overview

This skill audits and rewrites numeric and evaluation claims to make them reviewer-safe by ensuring each retained number carries minimal protocol context (task, metric, constraint). It prevents underspecified model naming and weakly supported numeric assertions without inventing facts or changing citation keys. Use it to convert fragile percentages and counts into interpretable, evidence-bounded statements or to downgrade them into qualitative claims when context is missing.

How this skill works

The checker scans target section files for digits, percents, and explicit numeric claims. For each claim it attempts to attach at least two pieces of protocol context (task/benchmark, metric, constraint) within the same sentence using available evidence files; if it cannot, it weakens or removes the numeric claim rather than guessing. It preserves citation keys, refuses to invent numbers, and flags ambiguous model names unless the citation explicitly supports them.

When to use it

• Pre-merge review when numeric statements may be reviewer magnets.
• When reviewers commonly flag numbers lacking task/metric/budget context.
• After a pipeline-auditor warns about underspecified model naming or suspicious claims.
• When polishing survey sections that report performance comparisons.
• Before finalizing drafts that will be submitted or circulated externally.

Best practices

• Always cross-check outline/writer_context_packs.jsonl and evidence drafts before editing numbers.
• Keep every retained number in the same sentence as at least two protocol elements: benchmark/task, metric, or constraint.
• If protocol context cannot be found, downgrade to qualitative wording (e.g., often, may, tends to) with the same citation.
• Do not invent, infer, or extrapolate numeric values; route thin evidence upstream when necessary.
• Never add, remove, or relocate citation keys; edit only wording and context placement.

Example use cases

• Convert 'Model X achieves 78% on XYZ' into 'On the XYZ benchmark, Model X reaches ~78% accuracy under the reported budget [@key]' when budget/metric are available.
• Downgrade a lone percentage to 'reported gains vary and depend on budget/retries [@key]' when constraint context is missing.
• Replace ambiguous model mentions like 'GPT-5' with the exact cited model name or a cautious phrasing if the citation does not support the label.
• Audit survey sections to ensure every numeric claim remains interpretable in isolation for reviewers.
• Produce a refined draft and optional completion marker after all evaluation anchors are reviewed.