# Scholar Verify

> Re-check a finished study against the corpus it cites — refs that moved, files that vanished, quotations that no longer match, numbers that changed

- Skill: `tony/scholar-verify` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add tony/scholar-verify`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tony/scholar-verify/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: tony (https://skillmd.com/u/tony)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/tony/scholar-verify

---


# this skill

A study is a set of claims about a corpus that keeps moving. Without this, it
rots into confident falsehood: every link still resolves, every number still
reads as measured, and none of it is true any more.

Read `references/citation.md` for what a citation was required to carry.

User arguments: $ARGUMENTS

## Procedure

### 1. Walk every citation

For each `cites` entry in `terms.jsonl`, confirm three things separately,
because they fail differently:

The ref still exists. A tag can be moved or deleted, and a deleted tag makes
every URL built on it a 404 rather than a wrong answer.

The file still exists at that path on that ref.

```console
$ curl -s -o /dev/null -w '%{http_code}\n' -L '<url>'
```

The quoted text still appears at the recorded locator. This is the one that
fails silently: a file that grew by ten lines leaves a line-anchor citation
resolving to unrelated code while still returning 200.

### 2. Re-run the measurements

With `--refresh-metrics`, run every command recorded in `evidence/metrics.md`
and diff the output against what was recorded. A number that changed is not
automatically wrong — the corpus moved — but it is no longer evidence for the
claim it was cited under.

### 3. Report by failure mode

Four categories, because the remedy differs:

- **Moved** — the ref advanced or the tag was retargeted. Re-pin.
- **Vanished** — the file or ref is gone. The claim needs a new source or it
  needs retracting.
- **Misquoted** — the locator resolves but the text there is not what was
  quoted. This is the dangerous one, and it is why quotations are recorded
  rather than just links.
- **Changed** — a measurement produced a different number.

## Rules

- Do not edit a claim in place. Anything that overturns a standing claim goes
  to the `scholar-revise` skill, which records what was believed and why it failed.
- Report a broken citation; do not repair it by finding a new source that
  happens to support the same claim. That is fitting evidence to a conclusion.
- A citation that cannot be checked — a source now behind a paywall, a deleted
  repository — is reported as uncheckable, not as passing.

## Output

Open with a one-line hero (`✓ <n> citations checked, <n> stale` or
`⚠ <n> misquoted`), then exactly these sections:

1. `## Moved` — each citation whose ref advanced, with the new ref.
2. `## Vanished` — each source that is gone.
3. `## Misquoted` — each locator whose text no longer matches, with both texts.
4. `## Changed` — each measurement that produced a different number, with both.

End with an `ask-user-choice` panel offering next steps (for example: revise
the overturned claims, re-pin the moved ones, stop here) — skip the panel only
in plan mode.


## Portability notes

- `ask-user-choice` — present the listed options and wait for the user to pick one. Hosts with a structured multiple-choice tool (Claude Code's `AskUserQuestion`) should use it; otherwise print a numbered list and wait for a numbered reply. Never proceed on an assumed answer.
- `$ARGUMENTS` — the text the user passed when invoking this skill. If your host does not substitute it, read it as the user's request in the current turn, and ask when there is none.
- Bundled files — every relative path in this skill points at a file shipped inside this skill directory. Read them from here, not from the host's plugin tree.