agentleFS
bm629bm629/agent-skills/skills/reviewing-ml-prior-art-survey/SKILL.md

reviewing-ml-prior-art-survey

Skill2 starsChanged 3 months agoMITView on GitHub

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.

The file

We split this skill into the rules it gives an agent

2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    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 …

    line 34
  2. Distinctive:

    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 — …

    line 103

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

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

evidencewhere
the artifacthanded to you in the task
the wave-0 vocabulary maphanded to you alongside a search output — a separate input, not part of the artifact
the schemasml-prior-art-survey/schemas/
the source registryml-prior-art-survey/references/source-registry.yaml
the angle referenceml-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.

outcomewhat is owedwhat the coverage conditions expect
rancells, and candidates if anything was admittedevery owed (group × source) pair has a cell
not_runNOTHING — the map's own verdict ruled this angle outno cells and no candidates is CORRECT
vacatedcells and causes; no candidatescells 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.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.