# Adversarial review brief

Review the exact TARGETS.json version. Constructor: Codex; independent reviewer: unassigned. The packet's known public fixtures are development inputs, not held-out cases.

1. Rebuild A and both selector roots from oriented incidence. Check the supplied lower-response choice separately from the spectral algebra; test the upper-root sector at the same Hodge frequency.
2. Trace every map domain/codomain. Confirm that omitting W from T Theta W Ic is a dimension error; verify selected-sector versus ambient-square statements.
3. Challenge the frame policy: the contract admits any selected orthonormal Darboux frame satisfying the generator convention. A quarter-turn frame is admitted and can change Theta; uniqueness or physical parity is not a conclusion. Is this the intended object for the planned test?
4. Test reciprocal Ic, c>0 and the numerical lane endpoints. Values outside the declared lane are INELIGIBLE, not theorem counterexamples.
5. Challenge the induced infinity norm, absolute/relative thresholds, basis dependence, scaling and floating error. Ask whether a future claim needs certified intervals or an exact radical implementation. No such certificate is claimed by v0.1.
6. Check all required gate reports, finite/malformed input behavior and CLI trust boundaries. Python JSON duplicate-key semantics and non-standard NaN spellings deserve explicit serialization policy before any cross-implementation campaign. The matrix/c guards reject nonfinite values.
7. Separate fixed algebraic checks (event antisymplecticity/copy-swap sign) from measurements. They are not empirical discriminators. Examine residual dependencies instead of treating their count as independent evidence.
8. Write new attacks or an independent derivation; state their construction and target exposure. A completed public fixture replay is not blind evidence.
9. Preserve historical C/D bytes and scores. Do not retroactively score those records under this new contract or infer original chronology/attribution from file custody.
10. Return exact findings for the20 declared review items, including NOT_REVIEWED. Require an attributable review/ruling record before any status change under applicable governance. Keep external-AI versus human provenance distinct.

Decision sought: is the target definition, source convention, frame policy and numerical admission rule fit to freeze a separately authored blind campaign, or does it need a new version? This is a review question, not a predeclared positive answer.