Evidence Draft Skill for Codex

Evidence Draft

Build deterministic outline/evidence_drafts.jsonl packs from briefs + notes + optional evidence bindings.

Compatibility mode is active: this migration preserves the existing JSONL contract while moving evidence-quality policy, sparse-evidence routing, and evaluation-anchor rules into references/ and assets/.

Load Order

Always read:

• references/overview.md
• references/evidence_quality_policy.md

Read by task:

• references/block_vs_downgrade.md when deciding whether thin evidence should block drafting or only downgrade claim strength
• references/evaluation_anchor_rules.md when evaluation tokens, protocol context, or numeric claims are weak
• references/examples_sparse_evidence.md for evidence-thin pack calibration
• references/source_text_hygiene.md when paper self-narration or generic result wrappers are leaking into pack snippets / claim candidates

Machine-readable assets:

• assets/evidence_pack_schema.json
• assets/evidence_policy.json
• assets/source_text_hygiene.json
• repo-wide assets/limitation-signals.json — shared polarity rules that keep

resolved failures and positive improvements out of limitation slots

Inputs

Required:

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

Optional but recommended:

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

Outputs

Keep the current output contract:

• outline/evidence_drafts.jsonl
• optional human-readable mirrors under outline/evidence_drafts/

Script Boundary

Use scripts/run.py only for:

• deterministic joins across briefs / notes / evidence bank / bindings
• snippet extraction and provenance assembly
• policy-driven blocking_missing / downgrade_signals / verify_fields materialization
• pack validation and Markdown mirror generation

Do not treat run.py as the place for:

• filler bullets that make thin evidence look complete
• hidden sparse-evidence judgment that is not inspectable from references/ / assets/
• reader-facing narrative prose

Output Shape Rules

Keep these stable:

• preserve the existing top-level pack fields already used by downstream survey pipelines
• claim_candidates must remain snippet-derived
• concrete_comparisons must remain genuinely two-sided; if one cluster has no usable highlight, drop the card and surface thin evidence upstream instead of fabricating an A-vs-B contrast
• snippet sampling should stay cluster-aware: when a subsection has explicit clusters, evidence selection should avoid collapsing onto one route just because its abstracts contain louder result sentences
• sparse evidence should surface as explicit blockers / downgrade signals / verify fields, not filler bullets
• citation keys must remain constrained to citations/ref.bib

Compatibility Notes

Current mode is reference-first with deterministic compatibility:

• assets/evidence_policy.json defines pack thresholds and sparse-evidence routing
• assets/evidence_pack_schema.json documents/validates the stable pack shape
• assets/source_text_hygiene.json owns this Skill's wrapper cleanup, while the

repo-wide assets/limitation-signals.json owns limitation polarity across

paper-notes, evidence-draft, and writer-context-pack

• scripts/run.py still materializes the existing JSONL + Markdown outputs, but no longer pads sparse sections with generic caution prose

Quick Start

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

Execution Notes

When running in compatibility mode, scripts/run.py currently reads:

• outline/subsection_briefs.jsonl
• papers/paper_notes.jsonl
• citations/ref.bib
• optionally papers/evidence_bank.jsonl and outline/evidence_bindings.jsonl
• assets/evidence_policy.json and assets/evidence_pack_schema.json

Script

Quick Start

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

All Options

• --workspace <dir>
• --unit-id <id>
• --inputs <path1;path2>
• --outputs <path1;path2>
• --checkpoint <C*>

Examples

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

Troubleshooting

• If packs look complete despite thin evidence, inspect assets/evidence_policy.json and references/block_vs_downgrade.md before changing Python.
• If evaluation bullets are generic, inspect references/evaluation_anchor_rules.md and the policy asset.
• If claims are strong but evidence is abstract/title-only, downgrade via downgrade_signals and verify_fields rather than adding narrative caveats.