Evidence Binder (NO PROSE)

Triggers & routing

• Trigger: evidence binder, evidence plan, section->evidence mapping, 证据绑定, evidence_id.
• Use when: papers/evidence_bank.jsonl exists and you want writer/auditor to use section-scoped evidence items (WebWeaver-style memory bank).

Goal: convert a paper-level pool into a subsection-addressable evidence plan.

This skill is the bridge from “Evidence Bank” → “Writer”: the writer should only use evidence IDs bound to the current subsection.

Why this matters for writing quality:

• Weak/undifferentiated bindings force the writer to either pad prose or cite out-of-scope.
• Treat binding_gaps as a routing signal: fix upstream evidence/mapping instead of "writing around" missing evidence.

Inputs

• outline/subsection_briefs.jsonl
• outline/mapping.tsv
• papers/evidence_bank.jsonl
• Optional:

- citations/ref.bib (to validate cite keys when evidence items carry citations)

Outputs

• outline/evidence_bindings.jsonl (1 JSONL record per subsection)
• outline/evidencebindingreport.md (summary; bullets + small tables)

- Includes gaps (missing required evidence fields) and tag mix (selected evidence tags) so subsection-specific evidence needs are visible.

Output format (outline/evidence_bindings.jsonl)

JSONL (one object per H3 subsection). Best-effort fields (keep deterministic):

• sub_id, title
• paper_ids (papers in-scope for this subsection, from mapping.tsv)
• mapped_bibkeys (bibkeys mapped to this subsection)
• bibkeys (a selected subset to encourage subsection-first citations)
• evidenceids (selected evidence items from papers/evidencebank.jsonl)
• evidencecounts (small summary by claimtype / tag / evidence_level)
• binding_rationale (short bullets; why the selected evidence covers this subsection’s axes / desired tags)
• binding_gaps (list[str]; required evidence fields not covered by selected evidence; drives the evidence self-loop upstream)

A150++ density contract (default)

• Use queries.md:per_subsection as the width contract (A150++ default: 28).
• Bind enough evidence to make writing concretely executable without out-of-scope pressure:

- mappedbibkeys: >= persubsection - evidenceids: >= persubsection - 4 (A150++: >=24) - bibkeys (selected): >= 20 (so each H3 has a usable citation pool, not just a long mapped list)

For draftprofile=coursepaper, retain the same traceability shape at compact density: 8 mapped bibkeys, at least 6 selected bibkeys, and at least 6 evidence IDs from at least 4 papers per H3.

Binding policy (how strict to be)

• Subsection-first by default: the writer should primarily cite bibkeys and use evidenceids bound to this subid.
• Allow limited within-chapter reuse: citations from sibling H3s within the same H2 chapter may be reused for background/evaluation framing, but:

- keep >=3 subsection-specific citations per H3 (avoid “free cite drift”) - avoid cross-chapter reuse unless the outline explicitly calls for it

Workflow (NO PROSE)

1. Read outline/subsection_briefs.jsonl to understand each H3’s scope/rq/axes.
2. Read outline/mapping.tsv to know which papers belong to each subsection.
3. Read papers/evidencebank.jsonl and select a subsection-scoped set of evidenceid items per H3.
4. If citations/ref.bib exists, sanity-check that any cite keys referenced by selected evidence items are defined.
5. Write outline/evidencebindings.jsonl and outline/evidencebinding_report.md.

Freeze policy

• If outline/evidencebindings.refined.ok exists, the script will not overwrite outline/evidencebindings.jsonl.
• Treat this marker as an explicit refinement/completion signal (especially in strict runs): only create it after you have checked binding_gaps and tag mix look subsection-specific.

Heterogeneity sanity check (avoid recipe-like bindings)

A common hidden failure mode is mechanical uniformity: every H3 ends up with the same claim_type/tag mix, which hides what each subsection is actually missing and pushes the writer toward generic prose.

Before you mark bindings as refined:

• Scan outline/evidencebindingreport.md: different H3 should show meaningfully different tag mix / claim_type balance.
• If most H3 look identical, treat it as a binder smell: tighten requiredevidencefields, adjust selection rationale, or enrich the evidence bank, then rerun.

Script

Quick Start

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

All Options

• --workspace <dir>: workspace root
• --unit-id <U###>: unit id (optional; for logs)
• --inputs <semicolon-separated>: override inputs (rare; prefer defaults)
• --outputs <semicolon-separated>: override outputs (rare; prefer defaults)
• --checkpoint <C#>: checkpoint id (optional; for logs)

Examples

• Bind evidence IDs after building the evidence bank:

- Ensure papers/evidence_bank.jsonl exists. - Run: uv run python .codex/skills/evidence-binder/scripts/run.py --workspace <workspace>

Troubleshooting

Issue: some subsections have too few evidence IDs

Fix:

• Strengthen papers/evidence_bank.jsonl via paper-notes (more extractable evidence items).
• Or broaden the mapped paper set for the subsection via section-mapper, then rerun binder.

Issue: binding_gaps is non-empty (missing evidence types)

What it means:

• The subsection brief requires certain evidence fields (e.g., benchmarks/metrics/security/tooling), but the bound evidence items do not cover them.

Fix (self-loop upstream):

• Prefer enriching papers/evidencebank.jsonl / papers/papernotes.jsonl for mapped papers (extract benchmark/metric/failure-mode details).
• If the mapping is weak for that evidence type, expand outline/mapping.tsv for the subsection and rerun binder.
• If the requirement is unrealistic for the subsection’s scope, revise outline/subsectionbriefs.jsonl:requiredevidence_fields and rerun binder.