# Reviewing Ml Prior Art Survey

> Use when reviewing an artifact produced by ml-prior-art-survey — an ML task vocabulary map, one angle's search output, one artifact's extract record, or the option register and its report — and deciding whether it can be built on. Judges against numbered conditions covering canonical terms the corpus actually uses, per-angle applicability verdicts in both directions, verbatim query recording, the recorded zero, cause evidence on every unreached source, the evaluation frame behind any leaderboard result, authority as a ranking rather than a cut, and the absence-as-finding rule. Also judges whether each adoption-ladder descent's reason is true of the record it names, and whether any figure was recomputed on its way into the register. Emits exactly one VERDICT approve or revise, with findings naming their condition, and does not revise a thin-but-honest result. Keywords: ML prior-art review, model survey review, survey quality gate, coverage review.

- Skill: `bm629/reviewing-ml-prior-art-survey` (Agent Skill, multi-file: 13 files)
- Install (CLI): `npx skillmds@latest add bm629/reviewing-ml-prior-art-survey`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bm629/reviewing-ml-prior-art-survey/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: bm629 (https://skillmd.com/u/bm629)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/bm629/reviewing-ml-prior-art-survey

---

# Reviewing an ML prior-art artifact (wave 1)

You are the judgment half of a two-part gate. The deterministic half has already run: shape, enums,
ranges, arithmetic and reconciliation are checked by `validate_ml_prior_art.py`, and an artifact
reaching you has passed it at exit 0.

**So never report what the validator already checks.** If your finding could have been produced by
the script, it is not a review finding — and repeating it costs a revise round on work that was
correct.

## What you judge

`references/conditions.md` is the single source of the bar. Read it and judge against it. The
producer skill assigns every duty; these conditions say how each is JUDGED, and add none of their
own.

## Your evidence, and where it lives

| evidence | where |
| --- | --- |
| the artifact | handed to you in the task |
| **the wave-0 vocabulary map** | handed to you alongside a search output — a separate input, not part of the artifact |
| the schemas | `ml-prior-art-survey/schemas/` |
| the source registry | `ml-prior-art-survey/references/source-registry.yaml` |
| the angle reference | `ml-prior-art-survey/references/angles/<angle_id>.md` |

**Three of those five live in the PRODUCER package.** C2 needs the map — its whole test is a
candidate's evidence against the group its `found_by` names. C10 and C15 need the angle reference
and the registry, and deriving an angle's source list from the artifact is circular: an artifact
that omits a source omits it from any list you read off it.

**If you cannot reach what a condition names, say so and emit no verdict on that condition**
rather than downgrading it silently.

**Anything you cannot ground in those five is an OBSERVATION, not a finding** — say it as one,
plainly, and do not attach a condition to it. An ungrounded finding costs a revise round on correct
work, and at the revision cap it parks the ticket. That is the price of guessing.

## Read `outcome` first

It decides which conditions apply at all.

| `outcome` | what is owed | what the coverage conditions expect |
| --- | --- | --- |
| `ran` | cells, and candidates if anything was admitted | every owed (group × source) pair has a cell |
| `not_run` | NOTHING — the map's own verdict ruled this angle out | no cells and no candidates is CORRECT |
| `vacated` | cells and causes; no candidates | cells as for `ran`; an empty candidate list is not a gap |

A `not_run` artifact has no cell for any pair, and the deterministic gate REQUIRES that. Reading a
coverage condition against it would revise work the other half of the gate certified.

## Proportionality

A thin result is not a failed result. A task with three published models yields three, and a survey
that says so honestly is complete. **Revise only on a named gap against a numbered condition** —
never because the artifact could have been longer.

## Output

Findings, each naming its condition, then exactly one terminal line:

```
VERDICT: approve
```

or

```
VERDICT: revise
```

Nothing after it.

## Upstream remedies

If a finding's remedy lies OUTSIDE the file this artifact's author was asked to write — the map, a
registry row, another angle's output — they cannot perform it, and ordering the revision anyway
burns cycles and ends somewhere naming neither the artifact nor the inconsistency.

Label it `UPSTREAM:` and **name the exact file and field that must change**. One that names no
target is a finding you could not localise, and it belongs in Observations.

**A file whose every finding is UPSTREAM gets `approve`.** You judge THIS artifact against its own
contract; if it satisfies that contract and the defect is elsewhere, it is not the defective thing.
Reserve `revise` for a remedy the author can actually perform. There is no third verdict —
escalation is for whoever dispatched you, not for you.