Qc Review

Skill: qc-review

Assess whether the dataset is in good shape to build on and release: is it BIDS-valid (Self-containment), and how does the project score across STAMPED? This is a read-only review that turns the current state into a short report with concrete next actions. You delegate validation to the bids doer and state inspection to the datalad doer; you do not modify anything.

When to use

• Before grouping/releasing a product, at the end of curation, or whenever the user wants a quality

or reproducibility check.

• Do NOT use to fix issues — this skill only diagnoses. It routes each gap to the skill that fixes

it (curate/annotate, curate/raw-to-bids, analyze/checkpoint, disseminate/publish, …).

Steps

1. BIDS validation (bids doer) — delegate:

> "validate this dataset against BIDS and report result (valid/invalid/unverified), error and > warning counts, and any dataset_description/participants/sidecar gaps." Relay the summary. If the validator is absent, report unverified with the install hint — never claim validity that was not checked.

1. State inspection (datalad doer) — delegate a read-only check:

> "status + log: is the working tree clean, what is the current branch, is a sibling configured?"

1. STAMPED self-assessment — from the bids result, the datalad state, and the ledger

(project.yaml), score each letter and name the gap-closing skill:

• S self-contained → BIDS valid + dataset_description.json/README present (curate/*)
• T tracked → working tree clean, runs recorded (analyze/checkpoint, run-comparison)
• A actionable → analyses are provenanced datalad runs, not ad-hoc shell
• M metadata → data dictionary / sidecars / controlled terms present (curate/annotate)
• P/E portable/ephemeral → a container recipe exists and analyses run in it

(process/run-pipeline, run-comparison)

• D distributable → a sibling exists / a product is released (disseminate/publish,

dataset-release)

1. Log it — append { ts, op: qc-review, stage: qc, note: "bids ; STAMPED gaps: ", branch: }, then delegate the save to the datalad doer.
2. Report — a compact scorecard: BIDS result (errors/warnings), the STAMPED letters that are

satisfied vs. the gaps, and for each gap the single skill that closes it. Recommend the highest priority next action.

Constraints

• Read-only: never modify, rename, or "fix" dataset files here — diagnose and route to the fixing

skill. Validation and state inspection go through the bids and datalad doers respectively.

• Never assert BIDS validity the bids doer did not verify (an absent validator is unverified, not

a pass).

• The STAMPED assessment is a snapshot, not a gate — report it plainly with gaps; do not block other

work. Keep log: append-only and the ledger schema-valid.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: neuroanalytics
• Source: neuroanalytics/data-science-harness
• License: MIT

Install and usage instructions live in the source repository linked above.