writeup-sop

Write-up SOP

This is an agent-gated publication workflow, not a filesystem security boundary. Hooks block only a narrower set of obvious unsafe writes.

Inputs and final deliverables

Require the target audience/venue, output path and format, intended claims, linked hypothesis/proposition ids, metric pins, and whether each result is confirmatory or exploratory. Deliver the manuscript/report, a claim manifest, reviewer JSON, and a short list of unresolved limitations.

1. Build the claim manifest before drafting

List every meaningful claim with:

• exact intended wording;
• kind: result_metric, statistical_claim, context, or theorem;
• role: central, supporting, or background;
• mode: confirmatory, exploratory, or not applicable;
• linked hypothesis/proposition id;
• required evidence and current status.

Dates, versions, seed counts, baseline counts, model sizes, and timeouts are usually context. They must be accurate but are not automatic provenance gates.

2. Close the empirical evidence chain

For each publication-critical numeric or statistical claim:

1. Call mcp__verify__check_provenance. Missing provenance means rerun, remove the claim, or clearly downgrade it to exploratory.
2. Require a real pin_id for central metrics.
3. Call mcp__verify__refresh_claim. Any stale code, data, config, Git state, dependency lock, runtime, or tracked environment blocks the central claim. Legacy unchecked evidence must be disclosed and is insufficient as the only support for a headline result.
4. Require a current stable seed verdict for central experimental metrics, or narrow the wording to an unstable/exploratory observation.
5. For method-versus-baseline claims, require a fair baseline_fairness verdict or disclose the resource mismatch next to the comparison.
6. For confirmatory claims, require the matching preregistration to be met. Check its fixed family_id and family_size; an open or missed row blocks confirmatory wording.