# Evidence Review

> Read a gathered source and record what it actually establishes, which is usually narrower than what it gets cited for.

- Skill: `tinyhumansai/evidence-review` (Agent Skill)
- Install (CLI): `npx skillmds@latest add tinyhumansai/evidence-review`
- Raw SKILL.md: https://api.skillmd.com/api/skills/tinyhumansai/evidence-review/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: tinyhumansai (https://skillmd.com/u/tinyhumansai)
- Updated: 2026-09-22
- Page: https://skillmd.com/skills/tinyhumansai/evidence-review

---


# Evidence Review

Read what was gathered. Record what it supports — not what the line of inquiry
would like it to support.

## When to use

- A `sources` row is in `gathered`.
- A finding cites something nobody has read end to end.

## Steps

1. **Read the thing itself**, not its abstract, not its summary, not the article
   about it. Abstracts overstate; that is what they are for.
2. **Find the actual claim** and the actual support for it — the method, the
   sample, the data. A conclusion with no method behind it is commentary
   whatever the venue.
3. **Write `establishes` narrowly.** "In one cohort of 40 over six months" is
   the finding; "X causes Y" is what somebody will cite it for.
4. **Note what it does not establish**, particularly where it is likely to be
   over-read.
5. **Flag contradictions rather than resolving them.** Two sources that disagree
   are the most informative thing the lab has today; smoothing them is the
   failure this step exists to prevent.
6. **Move the row to `read`,** and record or revise the `findings` row it bears
   on, with the source id in `refs`.

## Output

The `sources` row moved to `read` with an honest `establishes`, plus any
`findings` row it creates or changes. Sources that turn out not to say what they
were fetched for are `discarded` with that as the reason — expensive knowledge,
and expensive to learn twice.