Evidence Self-loop (C3/C4 fix → rebind → redraft)

Triggers & routing

• Trigger: evidence self-loop, evidence loop, evidence gaps, binding gaps, blocking_missing, 证据自循环, 证据缺口回路.

Purpose: make the evidence-first pipeline converge without writing filler prose.

This skill reads the intermediate evidence artifacts (briefs/bindings/packs) and produces an actionable TODO list that answers:

• Which subsections are under-supported?
• Is the problem mapping/coverage (C2) or evidence extraction (C3) or binding/planning (C4)?
• Which skill(s) should be rerun, in what order, to unblock high-quality writing?

Inputs

• outline/subsection_briefs.jsonl
• outline/evidencebindings.jsonl (expects bindinggaps / binding_rationale if available)
• outline/evidencedrafts.jsonl (expects blockingmissing, comparisons, eval protocol, limitations)
• Optional (improves routing):

- outline/evidencebindingreport.md - outline/anchorsheet.jsonl - papers/papernotes.jsonl - papers/fulltext_index.jsonl - queries.md

Outputs

• output/EVIDENCESELFLOOPTODO.md (report-class; always written)

Self-loop contract (what “fixing evidence” means)

• Prefer fixing upstream evidence, not writing around gaps.
• If an evidence pack has blocking_missing, treat it as a STOP signal: strengthen notes/fulltext/mapping, then regenerate packs.
• If bindings show bindinggaps, treat it as a ROUTING signal: either enrich the evidence bank for the mapped papers, expand mapping coverage, or adjust requiredevidence_fields if unrealistic.

Recommended rerun chain (minimal):

• If C3 evidence is thin: pdf-text-extractor → paper-notes → evidence-binder → evidence-draft → anchor-sheet → writer-context-pack
• If C2 coverage is weak: section-mapper → outline-refiner → (then rerun C3/C4 evidence skills)

Workflow (analysis-only)

1. Read queries.md (if present)

• Use it only as a soft config hint (evidencemode / draftprofile); do not override the artifact contract.

1. Read outline/subsection_briefs.jsonl

• For each subid, capture axes + requiredevidence_fields (what evidence types this subsection expects).

1. Read outline/evidence_bindings.jsonl

• For each subid, surface bindingrationale and binding_gaps (what the binder could/could not cover from the evidence bank).

1. (Optional) Read outline/evidencebindingreport.md

• Use it as a human-readable summary; treat it as a view of outline/evidence_bindings.jsonl, not a separate truth source.

1. Read outline/evidence_drafts.jsonl

• Surface blocking_missing (STOP signals), and check for missing comparisons / eval protocol / limitations that would force hollow writing.

1. (Optional) Read outline/anchor_sheet.jsonl

• Check whether each subsection has at least a few citation-backed anchors (numbers / evaluation / limitations).

1. (Optional) Read papers/papernotes.jsonl and papers/fulltextindex.jsonl

• Use these to route fixes: if evidence is abstract-only and missing eval tokens, prefer enriching notes/fulltext before drafting prose.

What the report contains

• Summary counts: subsections with blockingmissing, with bindinggaps, and common failure reasons.
• Per-subsection TODO: the smallest upstream fix path (skills + artifacts) to make the subsection writeable.

Status semantics (unblock rules)

This skill is the prewrite router for evidence quality. Treat its Status: line as the unblock contract:

• PASS: no blockingmissing and no bindinggaps -> proceed to C5 writing (but still scan non-blocking writability smells: low comparisons/eval/anchors often predict hollow prose).
• OK: no blockingmissing, but some bindinggaps -> you may draft, but expect weaker specificity; prefer fixing gaps first.
• FAIL: missing inputs OR any blocking_missing -> do not write filler prose; fix upstream and rerun C3/C4.

Routing matrix (symptom -> root cause -> upstream fix)

Use this as a semantic routing table (not a script checklist). The goal is to fix the earliest broken intermediate artifact.

Interface with the writer self-loop (avoid writing around evidence)

• If writer-selfloop is FAIL due to missing anchors/comparisons and the corresponding writer pack has pack_warnings, stop and run this evidence self-loop: the section is telling you the pack is not writeable.
• Prefer fixing evidence gaps once, upstream, rather than patching every H3 with generic filler.

What this skill does NOT do

• It does not edit papers/, outline/, or sections/*.
• It does not invent new facts/citations.
• It does not "relax" quality by changing thresholds; it routes you to the earliest artifact to fix.

Script

Quick Start

• uv run python .codex/skills/evidence-selfloop/scripts/run.py --workspace <workspace>

All Options

• --workspace <dir>
• --unit-id <U###> (optional)
• --inputs <semicolon-separated> (optional override)
• --outputs <semicolon-separated> (optional override; default writes output/EVIDENCESELFLOOPTODO.md)
• --checkpoint <C#> (optional)

Examples

• Generate an evidence TODO list after C4 packs are generated:

- uv run python .codex/skills/evidence-selfloop/scripts/run.py --workspace <workspace>