evidence-binder_skill

Overview

This skill binds addressable evidence IDs from a project-level evidence bank to each H3 subsection, producing subsection-scoped evidence plans. It creates outline/evidence_bindings.jsonl and a short report that surfaces gaps and tag mixes so writers and auditors can use only subsection-bound items. The process is strict: no invented evidence and deterministic selection rules to make writing executable.

How this skill works

The binder reads subsection briefs and the section->paper mapping, then selects evidence items from papers/evidence_bank.jsonl that best cover each subsection’s required axes and tags. It emits one JSONL record per H3 with mapped bibkeys, a curated bibkey subset, selected evidence_ids, counts by claim_type/tag/level, short binding rationale bullets, and binding_gaps. It also produces a compact report summarizing gaps and tag-mix heterogeneity.

When to use it

• You have papers/evidence_bank.jsonl and want subsection-first writing
• Preparing writer or auditor tasks that must cite only local evidence
• Enforcing evidence completeness before prose is drafted
• Running iterative research pipelines where section-specific evidence needs tracking

Best practices

• Treat binding_gaps as a routing signal: fix the evidence bank or mapping rather than inventing prose
• Aim for the A150++ width contract so each H3 has a usable citation pool (keep bibkeys and evidence_ids high enough)
• Keep bindings subsection-first; allow limited same-chapter reuse but require ≥3 subsection-specific citations
• Sanity-check tag mix across H3s to avoid mechanical uniformity before marking refined
• Do not overwrite bindings if outline/evidence_bindings.refined.ok exists

Example use cases

• Bind evidence after building the evidence bank to prepare a writer-focused task
• Audit subsection coverage and discover missing evidence types (benchmarks, failure modes, metrics)
• Tighten mapping.tsv and rebind when sections show repeated binding_gaps
• Generate deterministic evidence pools for automated reviewers or metrics extraction