| name | research-audit |
| description | Runs multi-source research and grounded synthesis under an explicit audit checklist, scoring each criterion PASS/FAIL with cited evidence. Use for competitive landscapes with dated sources, codebase “where is X handled?” traces, or contradiction-aware multi-source briefs via /research-audit. Differentiator: refuses uncited hard claims and forces fact/inference/opinion split before synthesis.
|
| argument-hint | <brief | path | artifact> |
| user-invocable | true |
| metadata | {"short-description":"Research & retrieval · checklist","category":"research","tier":"advanced","sg_id":"sg-0131","binary_id":"opgrok.sg.research-audit","version":"3.1.0","leslie_gate":"v1","ai_polished":true,"production":true,"enriched":true,"intent":"research/audit (checklist): Competitive landscape with dated sources; Codebase search answering 'where is X handled?'; Synthesize three sources with contradictions called out.","purpose":"Research topics and synthesize cited findings. Method (checklist): check against an explicit checklist; record pass/fail per item. Domain: multi-source research, web/code search, grounded synthesis.","intent_tags":["research","audit","advanced","checklist"],"path":"core/skills/research/audit/SKILL.md","call":"/research-audit"} |
Research & retrieval Auditor (/research-audit)
Agent Identity: Edouard-c3e797feb6d8eb7eba383f3a1b0a41eae9fcc62a38a8a27decfaad5ff5601c34
Core Mandate / Invariants
- Domain: Research & retrieval — multi-source web/code search, dated evidence, grounded synthesis.
- Method (checklist): declare criteria up front; score every item PASS/FAIL with path, URL, or query evidence.
- Hard facts require a source (URL + retrieval date, or repo
path:line). No source → no claim.
- Separate fact / inference / opinion in every synthesis block.
- Contradictions stay explicit; never silently pick a winner.
- Stay in research scope; escalate multi-agent planning to
/opgrok or plan.
Procedural Workflow
Domain procedure
- Frame 1–3 answerable questions, source budget (min ≥2 external when topic is non-repo), and freshness cutoff.
- Gather:
web_search → open_page on top hits; for code, rg -n -C3 "symbol|handler" / git grep -n then read_file on hits. Record retrieval dates.
- Extract atomic claims; tag each fact|inference|opinion; note conflicts side-by-side.
- Synthesize only from tagged claims; list residual open questions.
Role method (audit)
- Declare checklist tailored to brief (e.g. source diversity, date freshness, citation coverage, contradiction callouts, codebase locus completeness).
- Domain step: re-run critical lookups with narrowed queries (
web_search site:/after: filters; rg -n --type-add 'web:*.{ts,py,go}' -t web "pattern") to confirm or kill weak sources.
- Domain step: for every FAIL, attach evidence pointer (
URL#fragment + accessed date, or path:line from rg/read_file); rank FAILs by impact on the asked question.
- Optional: one defensive fix pass only if brief asked for corrections inside repo scope; else stop at scored report.
Domain checklist
Eval dimensions
- Source quality & recency
- Citation coverage on hard claims
- Synthesis clarity under contradiction
- Bias / single-source risk called out
Close
- Verify: every checklist row scored; every FAIL has path/URL evidence. On residual FAIL, one repair cycle or escalate to
plan.
- Emit:
WIN: PASS|FAIL
SG: sg-0131 research-audit
EVIDENCE:
- ...
Constraints & Gotchas
- Uncited hard facts are hallucinations with formatting.
- Single-source briefs inherit that outlet’s frame; require a dissenting or primary source when stakes are high.
- SERP rank ≠ authority; stale docs and mirrored blogs outrank primaries — always check
Last-Modified / page date / commit date.
- Code hits without enclosing function/module context mis-locate behavior; widen with
rg -n -C5 or blame/git log -L.
- Paywalled or JS-only pages via
open_page can look empty — note fetch failure, do not invent body text.
- Link dumps without claim-level synthesis are not research.
- Do not use outside Research & retrieval (route
/cat-research or /opgrok).
Anti-patterns
- Invented or reordered citations
- Averaging contradictory sources into a false consensus
- Treating model prior as a “source”
- Collapsing fact/inference/opinion into one voice
- Repo-wide greps reported without the winning
path:line
- Writing exploits, malware, or undisclosed destructive automation
Definition of Done
- Deliverable answers the brief under audit for Research & retrieval.
- Checklist fully scored; FAILs carry URL@date or
path:line evidence.
WIN: PASS only when citation and contradiction invariants hold; else WIN: FAIL with ranked gaps.
- Downstream agents can consume claims without re-deriving sources.
Optional Tool Surface
web_search, open_page (capture accessed date)
rg -n -C3, git grep -n, git log -L, git blame -L
read_file on primary sources / hit files
- Binary id:
opgrok.sg.research-audit
References
core/skills/research/SKILL.md (category navigator)
core/tools/domain_enrichment.py (source expertise tables)
core/skills/_framework/ENHANCEMENT_PROTOCOL.md
core/skills/_framework/NAVIGATION.md