evidence-selfloop_skill

Overview

This skill analyzes intermediate evidence artifacts and produces an actionable upstream TODO report to fix evidence gaps before drafting. It reads subsection briefs, evidence bindings, and evidence packs, then outputs a targeted plan (output/EVIDENCE_SELFLOOP_TODO.md) describing which upstream skill(s) to rerun and in what order. The goal is to prevent hollow prose by routing fixes to the earliest broken artifact rather than patching in the writer stage.

How this skill works

The skill scans outline/subsection_briefs.jsonl, outline/evidence_bindings.jsonl, and outline/evidence_drafts.jsonl to surface blocking_missing flags and binding_gaps. It summarizes counts and failure modes, then emits per-subsection minimal upstream fix chains (C1–C4 skills) such as rerunning paper-notes, evidence-binder, or section-mapper. It treats blocking_missing as a STOP signal and recommends the smallest, earliest rerun needed to produce writeable evidence packs.

When to use it

• When C4 evidence packs exist but writing reads hollow or thinly supported
• If evidence_drafts.jsonl contains blocking_missing entries
• When evidence_bindings.jsonl shows binding_gaps or weak binding_rationale
• Before attempting C5 authoring when anchor coverage, comparisons, or eval tokens look sparse
• When writer self-loop reports pack_warnings or FAIL status indicating missing anchors

Best practices

• Prefer upstream fixes over adding filler prose in C5; fix evidence once and rerun dependent skills
• Treat blocking_missing as a hard stop: enrich notes/fulltext or mapping before drafting
• Follow the minimal rerun chain: fix earliest broken artifact (C1/C2/C3) then rerun C4 and evidence-draft
• Use queries.md as a soft hint only; do not alter artifact contracts or thresholds
• Document the rerun order and the specific missing evidence type (metrics, snippets, bibkeys) in the TODO report

Example use cases

• Evidence packs show blocking_missing: no evaluation tokens — TODO recommends rerun chain: pdf-text-extractor → paper-notes → evidence-binder → evidence-draft
• Bindings flag binding_gaps for benchmarks — TODO routes to C3 paper-notes to extract metrics, then rerun anchor-sheet and evidence-draft
• Subsection mapping lacks coverage — TODO routes to C2 section-mapper → outline-refiner, then C3/C4 loop
• Citation keys missing across packs — TODO indicates C3 paper-notes to ensure bibkeys, then citation-verifier and rerun binder