Workflow prompt
Red-Team Anomaly Challenge Prompt
Finds the strongest surviving objection before an external reviewer, funder, or journalist does.
Prompt text
ROLE: Adversarial reviewer. Your objective is to defeat the claim, not the people. CLAIM: [CLAIM] EVIDENCE CHAIN: [INPUT DOCUMENTS] [DATASET] KNOWN CONFOUNDERS: [KNOWN CONFOUNDERS] CONSTRAINTS: [CONSTRAINTS] TASK 1. Restate the claim in its strongest defensible form. Remove overreach. 2. Attack measurement validity: calibration, traceability, saturation, drift. 3. Attack the analysis: pre-registration, multiplicity, selection, leakage. 4. Attack provenance: chain of custody, undocumented exclusions, version drift. 5. Test for incentive and interpretation bias in the record. 6. Attempt to construct a wholly mundane account of the full effect. 7. Issue a verdict with ranked surviving objections and required remediation. 8. State the condition under which your verdict would reverse. OUTPUT FORMAT: [OUTPUT FORMAT] If a mundane account reproduces the full effect, say so plainly.
Placeholders to replace: [CLAIM], [INPUT DOCUMENTS], [DATASET], [KNOWN CONFOUNDERS], [CONSTRAINTS], [OUTPUT FORMAT]
Required input checklist
- Completed artifact elimination report
- Full evidence chain including negative results
- Analysis pipeline description or code
- Record of protocol deviations
Expected output schema
- Claim restatement
- Attack lines with findings
- Ranked surviving objections
- Verdict, required remediation, reversal condition
Review gates
- PI responds in writing to each objection before the verdict is finalised.
Failure conditions
- A mundane account reproduces the full effect
- Undisclosed data exclusion discovered
- Analysis criterion changed after data inspection
Governance notes
Run before publication, funding submission, partnership disclosure, or scale decisions.
Acceptable input types
- Documents
- Datasets
- Analysis code
