bm629 reviewing-user-research-prior-art-survey — SKILL.md
Overview
Judge one produced artifact against the twenty-seven-condition bar in
references/conditions.md. That file is the authoritative bar — this body describes how to
apply it and never restates a condition normatively.
The arrangement exists because a bar duplicated across a producer and its reviewer drifts: one
document says kept counts results, the other says it counts rows, both look right in isolation,
and the disagreement only surfaces when an artifact is graded by the half that did not write it.
When to activate
- Judging a research vocabulary map, or one angle's search output, before it is accepted.
Do NOT activate for: authoring either artifact (the producing twin); judging an extract record or a findings register (wave 2, not yet shipped); assessing whether the underlying research is any good as science — the bar is whether the ARTIFACT is honest about what was searched and what was found.
What you are handed
The artifact, and — for a search output — the vocabulary map it was produced from. Stop the review and say so if a search output arrives without its map: the applicable set, and therefore coverage completeness, cannot be derived without it, and a review that proceeds anyway silently drops the condition that matters most.
Workflow
- Run the deterministic gate if it has not run:
validate_user_research_prior_art.py keyword-map <file>or... search <file> --keyword-map <map>. Its failures are the producer's to fix before review. Conditions marked (gated) are discharged by that run and are not restated as findings — duplicating the validator buries the judgment half in noise. - Walk the judgment conditions in order. Read the artifact against each; do not skim for problems.
- Ground every finding. Quote the artifact text that fails, and name the condition number. Anything you cannot ground in the artifact, its schemas, the registry or the map is an OBSERVATION, not a finding.
- Emit one verdict line, then the findings.
Where reviews of this artifact go wrong
- Grading the research instead of the record. Whether a study is well designed is not this bar. Whether the artifact honestly says what was searched, what was found, and what was not retrievable, is.
- Using your own knowledge of the literature as evidence. If a landmark paper seems missing, the finding is that the coverage does not account for it — not that you know its title. The producer queried from a map; the map is what a finding must point at.
- Reading thinness as failure. Most product scopes have very little research addressing them directly. C27 is a numbered condition precisely because the reflex is strong.
- Re-probing a source and treating your success as proof. One source here runs a globally shared pool whose throttling varies minute to minute. That you reached it now says nothing about whether the producer could.
- Reporting findings one at a time. Each round costs a revise cycle, and the loop caps out on work that was nearly right. One pass, every finding.
The two conditions this type turns on
Stated here because they are what makes this survey different from its siblings, not because they
are restated normatively — the wording that binds is in conditions.md.
Admission has two conjuncts (C18, C19). Retrievable full text alone admits practitioner argument that reports no study. A stated method alone admits a record built from an abstract, which reads exactly like one grounded in the method section. A reviewer who checks one and not the other passes half the artifact.
Certainty and transferability are not wave-1 judgments (C26). Both turn on the full read, and admission is decided before it. A wave-1 artifact that grades its evidence has invented a precise-looking value nothing can check.
Output
Exactly one verdict line — VERDICT: approve or VERDICT: revise — followed by findings, each
naming its condition number and quoting the artifact text that fails it. Nothing after the
verdict line, unless the caller's brief names an equivalent form, in which case that replaces it.
Emitting both forms is the violation; the cardinality is what is fixed, not the wording.
Related
user-research-prior-art-survey— the producing half. Its validator discharges the (gated) conditions; itsreferences/source-registry.yamlis part of your evidence.references/conditions.md— the authoritative bar.