Evaluation Anchor Checker

What the source does

• FilesystemNone observedNo reads or writes in the source
• NetworkNone observedNo outbound calls in the source
• ShellNone observedNo commands run in the source
• DependenciesPinnedNothing floats
• SecretsNone foundNo keys in source

Five audit layers

Gates applied: no behavioural pass, no license.

.claude/skills

digest = hashlib.sha1(str(seed or "").encode("utf-8", errors="ignore")).hexdigest()

Audit 0.4.1 · source sha 81994516a430 · 19 Oct 2026Report an issue or request a re-scan

• PublisherWILLOSCAR GitHub userOwner of the repository this skill is read from
• Source folderWILLOSCAR/research-units-pipeline-skills/.codex/skills/evaluation-anchor-checkerThe exact folder we audited
• LicenceNot declaredAs the repository states it
• Commit pinned for this verdict81994516a430A new audit runs when the source changes
• Repository statusNot archivedLast push 14 days ago
• Sandboxed runNot runSkills are read, not executed; layer L4 is skipped

Its description: Audit and rewrite evaluation/numeric claims to ensure they carry minimal protocol context (task + metric + constraint) and avoid underspecified model naming.

---
name: evaluation-anchor-checker
description: "Audit and rewrite evaluation/numeric claims to ensure they carry minimal protocol context (task + metric + constraint) and avoid underspecified model naming."
---

# Evaluation Anchor Checker (make numbers reviewer-safe)

## Triggers & routing

- **Trigger**: evaluation anchor checker, numeric claim hygiene, underspecified numbers, protocol context, 评测锚点检查, 数字断言, 指标上下文.


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
- `output/EVAL_ANCHOR_REPORT.md` (always; short report with files checked / changed / weakened sentences)
- Optional completion marker: `output/eval_anchors_checked.refined.ok`

## Recommended slot in the survey pipeline

Use this as the **last section-level numeric hygiene sweep before merge**:
- after `style-harmonizer`, `opener-variator`, `section-logic-polisher`, and
  `paragraph-curator`
- immediately before the final `argument-selfloop` snapshot and merge

Reason:
- earlier section-level rewrite passes can legitimately rephrase or fuse numeric sentences
- if you only wait for `pipeline-auditor`, numeric-context issues are discovered too late in the merged draft
- section-scoped fixes are cheaper and preserve citation anchoring better than post-merge patching

## 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)

```text
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).
- Write `output/EVAL_ANCHOR_REPORT.md` so the pipeline has an auditable completion artifact for this sweep.

## 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

- [ ] `output/EVAL_ANCHOR_REPORT.md` exists and reports a non-zero file count.
- [ ] 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

- `uv run python .codex/skills/evaluation-anchor-checker/scripts/run.py --workspace <workspace>`

### All Options

- `--workspace <dir>`: workspace containing `sections/*.md` or merged draft artifacts
- `--unit-id <id>`: optional harness metadata
- `--inputs <semicolon-separated>`: optional override from `UNITS.csv`
- `--outputs <semicolon-separated>`: optional output override; default includes `output/EVAL_ANCHOR_REPORT.md`
- `--checkpoint <C*>`: optional harness metadata

### Examples

- Run the numeric hygiene sweep before merge:
  - `uv run python .codex/skills/evaluation-anchor-checker/scripts/run.py --workspace <workspace> --inputs 'sections/*.md;outline/writer_context_packs.jsonl;citations/ref.bib' --outputs 'sections/*.md;output/EVAL_ANCHOR_REPORT.md;output/eval_anchors_checked.refined.ok'`

Excerpt, first 5,750 characters of SKILL.md, verbatim from the audited commit. It is text from someone else’s repository, quoted so you can judge it before your agent reads it.

1. 2026-10-19
   ✓SAFE B 89 / 100

   first audit · source 81994516a430

Questions about Evaluation Anchor Checker

• github.com/WILLOSCAR/research-units-pipeline-skills/tree/main/.c — the audited source at commit 81994516a430 and every finding aboveObserved · 19 Oct 2026

Spotted something wrong? Report a correction · Methodology