| name | ev-adversarial-audit |
| description | Run this skill for Phase 3 of the evidence verification pipeline — the adversarial audit step. Use when you need to check the evidence data lake for bad data before ingestion: dead links, wrong URL formats (tree/ vs blob/), subjective wording ("elite", "high-quality"), stale migration notes, or skills whose star tier doesn't match their classified evidence level. Triggers on phrases like: "audit the data lake", "adversarial check", "ev-adversarial-audit", "check for noise in evidence", "flag bad evidence", "run the audit phase", "quality check the tier files", or any reference to Phase 3 of the pipeline. Deploys 4 parallel adversarial reviewer subagents across the tier files, then a 5th synthesis subagent to merge findings, and appends results to the daily source report.
|
Adversarial Evidence Audit (ev-adversarial-audit)
Phase 3 of the evidence verification pipeline. After ev-star-verification has partitioned the data lake into tier files, this skill deploys 4 parallel adversarial reviewer subagents to scan those files from a "Devil's Advocate" perspective, then a 5th synthesis subagent merges their findings into the master source report.
The adversarial approach matters because a single sequential pass over a large tier file tends to miss subtle issues — evaluative adjectives slipping through, a tree/ URL that looks plausible at a glance, or a star threshold mismatch that only becomes visible when cross-referencing the tier. Parallelising reviewers and prompting them to actively look for problems (rather than rubber-stamp entries) catches the class of errors that confident single-pass scanning misses.
What Each Reviewer Looks For
Reviewers check for four failure modes:
Evaluative noise — subjective wording like "elite", "high-quality", "best-in-class", or procedural rank annotations and stale database migration comments that crept into evidence descriptions. These are not evidence; they inflate perceived credibility.
URL format errors — bare repository URLs (https://github.com/owner/repo) used where a blob/branch/subpath pointing to the actual skill directory is required; tree/ folder references that must be blob/; case-sensitivity mismatches; missing /SKILL.md suffix on installable skills.
Proxy mismatches — evidence typed as github-stars-own pointing to a URL that is already covered by a repo-own entry at the same path (same-source dedup means only the higher-scoring entry counts — the duplicate is silent waste).
Tier/classification drift — skills whose evidence star tier doesn't align with their classified rank in registry/named-skills.json, indicating a stale tier file or a promotion that wasn't reflected downstream.
Workflow
Split the tier files across 4 parallel adversarial reviewer subagents to stay within context limits for large tiers (tier_2.md is the biggest and must be split):
- Agent 1:
tier_1.md, tier_5.md, tier_6.md
- Agent 2:
tier_2.md lines 1–1382
- Agent 3:
tier_2.md lines 1383–2767
- Agent 4:
tier_3.md, tier_4.md
Each agent receives the same adversarial reviewer prompt: act as a Devil's Advocate, assume every entry has at least one issue, and produce a structured findings list (entry name, tier file, line range, issue category, recommended fix).
After all 4 complete, run a 5th synthesis subagent (type self) to:
- Deduplicate overlapping findings across agents.
- Group by issue category.
- Produce a final ranked summary ordered by severity.
Append the merged findings to the active source report under a new section:
## 6. Adversarial Data Lake Audit Findings (YYYY-MM-DD)
The source report lives at evidence/source_report_YYYY_MM_DD.md (use today's date).
Invocation
/ev-adversarial-audit
This skill is normally invoked as part of the full pipeline via /ev-pipeline. Run it standalone when you want to re-audit the tier files after making corrections — for example, after fixing URL errors flagged in a previous run, re-run this skill to confirm the fixes are clean before proceeding to /ev-link-validation.