Evidence Selfloop Skill for Codex

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

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/evidence_bindings.jsonl (expects binding_gaps / binding_rationale if available)
• outline/evidence_drafts.jsonl (expects blocking_missing, comparisons, eval protocol, limitations)
• Optional (improves routing):
• outline/evidence_binding_report.md
• outline/anchor_sheet.jsonl
• papers/paper_notes.jsonl
• papers/fulltext_index.jsonl
• queries.md

Outputs

• output/EVIDENCE_SELFLOOP_TODO.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 binding_gaps, treat it as a ROUTING signal: either enrich the evidence bank for the mapped papers, expand mapping coverage, or adjust required_evidence_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 (evidence_mode / draft_profile); do not override the artifact contract.

2) Read outline/subsection_briefs.jsonl

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

3) Read outline/evidence_bindings.jsonl

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

4) (Optional) Read outline/evidence_binding_report.md

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

5) Read outline/evidence_drafts.jsonl

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

6) (Optional) Read outline/anchor_sheet.jsonl

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

7) (Optional) Read papers/paper_notes.jsonl and papers/fulltext_index.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 blocking_missing, with binding_gaps, 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 blocking_missing and no binding_gaps -> proceed to C5 writing (but still scan non-blocking writability smells: low comparisons/eval/anchors often predict hollow prose).
• OK: no blocking_missing, but some binding_gaps -> 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.

| Symptom (where you see it) | Likely root cause | Inspect first | Smallest upstream fix chain |

|---|---|---|---|

| evidence_drafts.blocking_missing: no usable citation keys | mapped papers lack bibkey / bibkeys not in citations/ref.bib | papers/paper_notes.jsonl (bibkey fields), citations/ref.bib | C3 paper-notes (ensure bibkeys) -> C4 citation-verifier -> rerun evidence-binder -> rerun evidence-draft |

| blocking_missing: title-only evidence | retrieval/metadata lacks abstracts (or aggressive filtering) | papers/papers_raw.jsonl abstracts, papers/paper_notes.jsonl evidence_level | C1 literature-engineer (enrich metadata) OR C3 pdf-text-extractor (fulltext) -> rerun paper-notes |

| blocking_missing: no evidence snippets extractable | notes are too thin / evidence bank empty for mapped papers | papers/evidence_bank.jsonl (counts), papers/paper_notes.jsonl | C3 paper-notes (richer extraction; prefer fulltext when possible) -> rerun C4 packs |

| blocking_missing: no concrete evaluation tokens | notes/bank did not extract benchmarks/metrics/budgets | papers/paper_notes.jsonl (metrics/benchmarks fields), outline/anchor_sheet.jsonl | C3 paper-notes (extract eval anchors) -> rerun anchor-sheet + evidence-draft |

| evidence pack comparisons are sparse (signals: comparisons low) | clusters are not contrastable OR mapping coverage too weak | outline/subsection_briefs.jsonl (clusters), outline/mapping.tsv | C2 section-mapper (coverage) OR C3 subsection-briefs (better clusters) -> rerun evidence-draft |

| bindings.binding_gaps mentions benchmarks/metrics/protocol | binder cannot find evaluation-tagged evidence for this subsection | outline/evidence_binding_report.md (tag mix), papers/evidence_bank.jsonl tags | C3 paper-notes (tag/evidence extraction) OR C2 expand mapping for that subsection -> rerun evidence-binder |

| binding_gaps mentions security/threat model/attacks | mapped set lacks security-focused works or notes lack threat-model detail | outline/mapping.tsv, papers/paper_notes.jsonl | C2 expand mapping (+ C1 queries if needed) OR C3 enrich notes -> rerun binder/packs |

| binding report looks mechanically uniform across H3 (same mix, low tag variance) | binder selection too recipe-like OR evidence bank tags too coarse | outline/evidence_binding_report.md (tag mix), evidence bank tags | tighten required_evidence_fields + improve evidence bank tags, then rerun binder; avoid writing around non-specific bindings |

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/EVIDENCE_SELFLOOP_TODO.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>