Independent audit: PASSED

Scope: frozen 64-candidate / 10-output / core-six / at-most-two-addition study. All saved new development and held-out policies were audited. No full evaluation or candidate generation was rerun.

Verified:
- Four input SHA-256 values and byte counts match the frozen protocol. Every copied .npy field equals its authoritative NPZ field, including dtype.
- Candidate, moment and truth parent axes match. Held-out query/candidate identities and per-query optimization/check seeds match the archived original study.
- Fold0 tuning and folds1-4 evaluation are disjoint. All 46x256x3 development and 6x3276x3 held-out output lists are unique, in bounds, preserve core/order and complete lists, and use at most two external candidates.
- Actual development policies confirm MMR lambda0 = farthest-first, and rank-based MMR lambda1 = Reference. Independent development losses select MMR_R=.4 and MMR_P=.7.
- Every checked output reconstructs from its same raw proposal and saved single-check decision, or exact Reference on failure. Raw sensitivity rows remain separate and use the same frozen lambdas.
- Eight outcome-independent held-out query replays (first parent in each fold, both degraded conditions) independently reproduce all three proposals and their check gains/LCBs.
- Independent candidate/truth distance calculations reproduce all summary metrics. Independent 10,000-parent paired bootstrap reproduces all contrast points/intervals and the original plan hash.

Sign convention below: EXTERNAL minus TraceRIFT mean regret. Positive numbers favor TraceRIFT. These are the negated TraceRIFT-minus-external rows in the experiment CSV, with interval endpoints reversed.
Sequential: -5.9398471506506666e-05; 95% interval [-0.00051156178460397305, 0.00040152900589778521]
MMR_R: 0.0074200452149915403; 95% interval [0.0055406322432891885, 0.0093507896410760776]
MMR_P: 0.0083657676905730689; 95% interval [0.0066024372434916939, 0.010180944876185441]
FarthestFirst: 0.019152632928177496; 95% interval [0.017065877984227547, 0.021402779502813971]
MMR_R_raw: 0.0078741005635277802; 95% interval [0.0059774547965382391, 0.0098488185104942776]
MMR_P_raw: 0.0084414667507399666; 95% interval [0.0066162189672036758, 0.010285030101224972]
FarthestFirst_raw: 0.025315036345016521; 95% interval [0.022986911848435473, 0.027698041533228555]

Reporting cautions:
- check_not_passed_fraction includes identity proposals; actual rollback is changed_by_check or rejected nonidentity proposals.
- The matched-core/quota assertion covers the new methods, Reference, Sequential and TraceRIFT; historical FixedDonor/TargetGuided remain separate one-exchange comparisons.
- Both raw sensitivities use checked-development-optimal lambdas; they are not independently retuned optimal ungated MMR methods.
- The common predictive check is part of the adapted methods, so checked results test a hybrid with that shared safeguard; raw rows separately test unguarded proposals.
- This is a retrospective extension of the existing fixed test cohort; no new untouched data or candidate generation is asserted.
- Only eight held-out queries were reselected in this audit. Full saved policy invariants/statistics were checked, but full selector replay was not repeated.

Timing benchmark is outside this completed numerical audit.
Details: independent_audit_report.json; independent_method_summary.csv; independent_paired_contrasts.csv; independent_development_grid.csv.
