| name | adversarial-review |
| description | Adversarial edition review — actively tries to break the compiled edition before publish. Probes contradictions, unsourced claims, narrative gaps, and framing that only holds under one reading. Complements reward-hacking scans (39.8) which watch for evaluator gaming. This watches for content problems that normal review misses. |
| version | 1.0 |
| updated | "2026-04-17T00:00:00.000Z" |
| tags | ["media","active"] |
| effort | medium |
| disable-model-invocation | true |
| argument-hint | [cycle-number] |
/adversarial-review — Break the Edition
Purpose
Normal review asks "is this good?" Adversarial review asks "how could this be wrong?"
Run this against a compiled edition after Step 3 (compile) and before Step 3.5 (capability review). It catches problems that survive three rounds of happy-path reading: contradictions between articles, claims no ledger source backs, civic stats that only hold under one framing, citizen quotes readable multiple ways, and narrative gaps the paper pretends don't exist.
Source: Fulton, "Agent Skills: The Cheat Codes for Claude Code" (April 2026) — adversarial review mode catches issues that pass normal review.
When to Run
- Before every publish — after compile, before capability review
- After major revisions — when reporters rewrite after Rhea/Mara feedback
- On demand — when something feels off but you can't name it
Inputs
- Compiled edition —
editions/cycle_pulse_edition_{XX}.txt
- Engine audit JSON —
output/engine_audit_c${XX}.json (for cross-checking claims against data)
- Previous edition —
editions/cycle_pulse_edition_{XX-1}.txt (for contradiction detection)
The Five Probes
Run all five against the compiled edition. Each probe is a different adversarial lens.
Probe 1: Cross-Article Contradictions
Read every article in the edition. For each factual claim (names, numbers, dates, roles, outcomes), check whether any other article in the same edition contradicts it.
What to look for:
- Same citizen described with different roles in two articles
- Approval ratings that don't match between civic and front page
- Initiative status described differently (e.g., "passed" in civic, "under review" in letters)
- Timeline conflicts (event happened "two cycles ago" in one article, "last cycle" in another)
- Score/stat conflicts between sports articles
Output per finding:
CONTRADICTION: [article A headline] says [claim A]. [article B headline] says [claim B].
Severity: HIGH/MED
Fix: [which is correct, or both need checking]
Probe 2: Unsourced Claims
For every quantitative claim, approval rating, population stat, economic figure, or specific outcome:
- Can it be traced to the engine audit JSON?
- Can it be traced to a citizen lookup or canon search?
