---

name: evidence-draft description: | Create per-subsection evidence packs (NO PROSE): claim candidates, concrete comparisons, evaluation protocol, limitations, plus citation-backed evidence snippets with provenance. Trigger: evidence draft, evidence pack, claim candidates, concrete comparisons, evidence snippets, provenance, 证据草稿, 证据包, 可引用事实. Use when: outline/subsection_briefs.jsonl exists and you want evidence-first section drafting where every paragraph can be backed by traceable citations/snippets. Skip if: outline/evidence_drafts.jsonl already exists and is refined (no placeholders; >=4 comparisons per subsection; blocking_missing empty). Network: none (richer evidence improves with abstracts/fulltext). Guardrail: NO PROSE; do not invent facts; only use citation keys that exist in citations/ref.bib.

Evidence Draft (NO PROSE)

Purpose: turn papers/paper_notes.jsonl + subsection mapping into writeable evidence packs so the writer never has to guess (and never copies outline placeholders).

Key design: every pack should contain evidence snippets (1–2 sentences) with provenance (abstract/fulltext/notes pointer). Even abstract-level snippets are better than template prose.

Why this matters for writing quality:

• Packs are the writer substrate; if packs are thin, C5 will either pad prose or guess.
• Treat blocking_missing as a stop signal: route upstream (notes/bindings) instead of writing around gaps.

Role cards (prompt-level guidance)

• Evidence Curator

  • Mission: turn paper notes into contrastable, citeable evidence (not summaries).
  • Do: extract snippet-backed comparisons; surface protocol details and failure modes.
  • Avoid: axis-driven "hypotheses" that are not supported by snippets.

• Provenance Accountant

  • Mission: keep every snippet auditable.
  • Do: attach provenance pointers (abstract/fulltext/notes location) and keep excerpts sentence-level.
  • Avoid: untraceable paraphrases that cannot be verified.

• Skeptic

  • Mission: prevent evidence inflation.
  • Do: downgrade claims when evidence is abstract/title-only; convert unknowns into verify_fields (not repeated boilerplate in prose).
  • Avoid: strong conclusions without protocol/metric context.

Non-negotiables

• NO PROSE: packs are bullets-only evidence, not narrative paragraphs.
• No fabrication: do not invent datasets/metrics/numbers.
• Citation hygiene: every cite key must exist in citations/ref.bib.
• Claim candidates must be snippet-derived (no axis-driven “Hypothesis: …” items); put questions into verify_fields instead.
• Avoid silent truncation: keep claim_candidates[].claim long enough to carry concrete detail (<= ~400 chars) and keep highlight excerpt sentence-level (<= ~280 chars).
• Numeric-claim hygiene (evidence substrate):
  • If a snippet/claim includes digits or %, also include minimal context in the same bullet (at least 2 of: task/setting, metric definition, constraint/budget/tool access).
  • If the context is not present in notes/fulltext, do not keep the number; downgrade to qualitative wording and add a verify_fields item instead.
• Evidence-aware language:
  • fulltext-backed → can summarize comparisons
  • abstract-only/title-only → must be provisional + list verify-fields (no strong “dominant trade-offs” language)

Inputs

Required:

• outline/subsection_briefs.jsonl
• papers/paper_notes.jsonl
• citations/ref.bib

Optional (recommended for addressable evidence):

• papers/evidence_bank.jsonl
• outline/evidence_bindings.jsonl

Outputs

• outline/evidence_drafts.jsonl

Optional (human-readable):

• outline/evidence_drafts/ (folder of per-subsection Markdown packs)

Output format (outline/evidence_drafts.jsonl)

JSONL (one JSON object per line). Required fields per record:

• sub_id, title
• evidence_ids (list[str]; if outline/evidence_bindings.jsonl exists, this should match the binding for the subsection)
• evidence_level_summary (counts: fulltext|abstract|title)
• evidence_snippets (list; each has text, paper_id, citations, provenance)
• definitions_setup (list of cited bullets)
• claim_candidates (3–5 items; each has claim, citations, evidence_field)
• concrete_comparisons (>=4 items; each has axis, A_papers, B_papers, citations, evidence_field; may also include A_highlights/B_highlights with snippet-backed contrast anchors)
• evaluation_protocol (list of concrete protocol bullets + citations)
• failures_limitations (2–4 cited bullets)
• blocking_missing (list[str]; if non-empty, drafting must stop)
• verify_fields (list[str]; non-blocking: fields to verify before making strong claims)

Provenance schema (per snippet)

Example:

Copy{"source":"abstract","pointer":"papers/paper_notes.jsonl:paper_id=P0012#abstract","evidence_level":"abstract"}

Allowed source: fulltext|abstract|paper_notes|title.

Workflow

1. Load outline/subsection_briefs.jsonl and read each subsection’s rq, axes, clusters, and evidence-level policy.
2. Load papers/paper_notes.jsonl and build a per-paper evidence index (bibkey, evidence_level, abstract, fulltext_path, limitations).
3. For each subsection:
   • Build evidence_snippets from mapped papers (prefer fulltext, else abstract), and record provenance.
   • Definitions/setup: 1–2 bullets that define setup + scope boundary (with citations).
   • Claim candidates: 3–5 checkable candidates (prefer snippet-derived; tag with evidence_field).
   • Concrete comparisons: >=4 A-vs-B comparisons (cluster-vs-cluster) along explicit axes.
   • Evaluation protocol: list concrete benchmark/metric tokens if extractable; otherwise treat as a blocking gap.
   • Failures/limitations: 2–4 concrete limitations/failure modes with citations.
   • Set blocking_missing for hard blockers (e.g., no usable citations; title-only evidence; no eval tokens for an eval-heavy subsection).
4. Write outline/evidence_drafts.jsonl and per-subsection Markdown copies.

Quality checklist

• Every subsection has >=4 concrete comparisons.
• evidence_snippets is non-empty and includes provenance.
• Any bullet containing numbers/% also carries minimal protocol context (task/metric/constraint), or the number is removed and moved to verify_fields.
• claim_candidates has >=3 snippet-derived items (no axis-driven hypotheses).
• blocking_missing is empty.
• No TODO / (placeholder) / <!-- SCAFFOLD --> / unicode ellipsis (…) remains.

Script

Refinement marker (recommended; prevents churn)

When you are satisfied with evidence packs (and blocking_missing is empty), create:

• outline/evidence_drafts.refined.ok

This is an explicit "I reviewed/refined this" signal:

• prevents scripts from regenerating and undoing your work
• (in strict runs) can be used as a completion signal before writing

Quick Start

• python .codex/skills/evidence-draft/scripts/run.py --help
• python .codex/skills/evidence-draft/scripts/run.py --workspace <ws>

All Options

• See --help.
• Inputs (required): outline/subsection_briefs.jsonl, papers/paper_notes.jsonl, citations/ref.bib.
• Inputs (optional): papers/evidence_bank.jsonl, outline/evidence_bindings.jsonl.

Examples

• Generate evidence packs after citations:
  • Ensure citations/ref.bib exists.
  • Ensure outline/subsection_briefs.jsonl exists (axes/clusters/plan filled).
  • Ensure papers/paper_notes.jsonl has usable evidence (abstract/fulltext/limitations).
  • Run: python .codex/skills/evidence-draft/scripts/run.py --workspace workspaces/<ws>

Troubleshooting

Issue: evidence packs have blocking_missing entries

Fix:

• Treat blocking_missing as a stop signal: enrich papers/paper_notes.jsonl / papers/evidence_bank.jsonl (or adjust scope) before drafting prose.

Issue: evidence snippets are empty or untraceable

Fix:

• Ensure each snippet includes provenance (paper id + location) and that outline/evidence_bindings.jsonl is non-empty for the subsection.