Adversarial Review Cards

What conclusion does this wording make easier? Which alternative frames were excluded? What classification would reduce scrutiny? Who disappears if the current category is treated as natural?

Reframe or reclassify when a materially different framing changes evidence needs, standing, safeguards, or feasible options.

Who gains, who bears cost or risk, who cannot exit, and who is easiest to miss? Is visibility being mistaken for importance? Does an aggregate gain conceal a concentrated or rights-sensitive burden?

Expand the stakeholder or distribution record, seek missing evidence, or trigger rights and standing review before authorization.

Which institution may perform the proposed act? What is the legal or constitutional source and limit? Is expertise, popularity, model consensus, or urgency being used to invent authority?

Block implementation claims until authority is verified; route the recommendation to the legitimate institution or lawful emergency pathway.

Does the measure capture the concept? Who is outside the denominator or sampling frame? Was the method selected before the desired result was known? Which missingness mechanism could reverse the inference?

Redesign measures or sampling, narrow the claim, collect missing evidence, or mark the inference unresolved.

Whose interpretation dominates? Were contradictory cases sought? Are quotations decontextualized? Did recruitment, translation, positionality, or coding choices systematically shape the finding?

Revisit sampling or coding, add negative cases, preserve rival interpretations, or limit transfer beyond the observed context.

Is a descriptive association being treated as causal? What counterfactual supports the claim? Which assumption—parallel trends, no unmeasured confounding, no spillovers, stable measurement—does the estimate need?

Downgrade causal language, run robustness or falsification tests, seek a stronger design, or preserve causal uncertainty.

What data, labels, transformations, objective, model or version, and validation produced the output? Where could correlated error or subgroup failure hide? Is fluency being mistaken for evidence?

Validate against source or ground truth, test subgroups, preserve provenance, seek independent methods, or prohibit consequential use.

Are multiple sources genuinely independent or repeating the same dataset, report, or model? What credible contrary evidence was omitted? Were incomparable outcomes forced into one score?

Rebuild claim-level synthesis, de-duplicate lineage, preserve dissent and non-comparability, or revise evidence strength.

Who could participate and who could not? What does participation actually establish? Are vocal participants being called representative? Which information or moderation conditions shaped the expressed preferences?

Correct representation claims, add missing participation modes, disclose information conditions, or separate voice from authorization.

Does the implemented policy match the authorized baseline? Which metric can be gamed? Could measured improvement coexist with unmeasured harm? Is monitoring being treated as proof of cause?

Audit implementation divergence, add counter-metrics, investigate mechanisms, or trigger formal reopening without automatic reversal.

Who controls framing, routing, panel construction, aggregation, stopping, publication, and review? Is the same principal evaluating its own safeguards under a different role label?

Require principal separation, external review, conflict disclosure, procedural audit, or explicit acknowledgement that independence has not been established.