Evaluation Anchor Checker (make numbers reviewer-safe)

Purpose: fix a reviewer-magnet failure mode in agent surveys:

• strong numeric/performance statements appear
• but the minimal evaluation context is missing

This skill treats numeric claims as contracts:

• if a number stays, the same sentence must contain enough protocol context to interpret it
• if that context is not in evidence, the claim must be downgraded (no guessing)

Inputs

Preferred (pre-merge, keeps anchoring intact):

• the affected sections/*.md files

Optional context (read-only; helps you avoid guessing):

• outline/writer_context_packs.jsonl (look for evaluation_anchor_minimal, evaluation_protocol, anchor_facts)
• outline/evidence_drafts.jsonl / outline/anchor_sheet.jsonl
• citations/ref.bib

Outputs

• Updated sections/*.md (or output/DRAFT.md if you are post-merge), with safer evaluation anchoring
• Optional completion marker: output/eval_anchors_checked.refined.ok

Read Order

Always read:

• references/numeric_hygiene.md

Machine-readable asset:

• assets/numeric_hygiene.json

The asset defines the keyword families and qualitative fallback templates. Keep the script deterministic and let the policy live in the asset/reference pair.

Role prompt: Reviewer-minded Editor (evaluation hygiene)

You are a reviewer-minded editor for evaluation claims in a technical survey.

Goal:
- make every numeric/performance claim interpretable and reviewer-safe

Hard constraints:
- do not invent numbers
- do not add/remove/move citation keys
- if protocol context is missing, weaken or remove the numeric claim

Minimum context to include when keeping a number:
- task / setting (what kind of task)
- metric (what is being measured)
- constraint (budget/cost/tool access/horizon/seed/logging) when relevant

Avoid:
- ambiguous model naming that looks hallucinated (e.g., “GPT-5”) unless the cited paper uses it verbatim

Workflow (explicit inputs)

• Use outline/writer_context_packs.jsonl to locate the subsection's allowed citations and any extracted evaluation_protocol/anchor_facts.
• Cross-check outline/evidence_drafts.jsonl and outline/anchor_sheet.jsonl for task/metric/constraint context before touching numbers.
• Validate every cited key against citations/ref.bib (do not introduce new keys).

What to enforce (the “minimum protocol trio”)

When a sentence contains digits (%, x, or numbers):

• Keep the number only if you can attach at least 2 of the following in the same sentence without guessing:
  • task family / benchmark name
  • metric definition
  • constraint (budget, tool access, cost model, retries, horizon)

If you cannot, downgrade:

• remove the number and rewrite as qualitative (“often”, “can”, “may”) with the same citation
• or move the specificity into a verification target (“evaluations need to report …”) without adding new facts

Mini examples (paraphrase; do not copy)

Bad (underspecified):

• Model X achieves 75% exact performance [@SomeBench].

Better (minimal context):

• On <task/benchmark>, Model X reaches ~75% <metric>, under <constraint/budget/tool access> [@SomeBench].

Better (downgrade when context is missing):

• Reported gains vary, but comparisons remain fragile when budgets and retry policies are not reported [@SomeBench].

Done checklist

• No numeric claim remains without minimal protocol context.
• No ambiguous model naming remains unless explicitly supported by citations.
• Citation keys are unchanged.
• If you removed/downgraded numbers, the paragraph still makes a defensible, evidence-bounded point.

Script

Quick Start

• python .codex/skills/evaluation-anchor-checker/scripts/run.py --workspace workspaces/<ws>