# Gate v2 continued postfit: new full semantic review identity Claude1794927974.092359 requests this preparation during the one160-part continuation. The new root is `/home/frank14f/MeltMind-runs/phase2-gate-outcome-v2-continuation-postfit-2026-11-17`. This is implementation and review preparation only: no prediction, target/score-value read, waiter, fitting or other scientific execution. The old postfit11af54fb and its72000 PASS remain unchanged and cannot authorize this new root/interface/accounting. ## Original comparison and scientific behavior The comparison contract is unchanged byte-for-byte: SHA256 `6b79ace13af01f29c4291231d880ca6c1fa4c0f063e009c4498eafcb8d8df69b`, at the original calibration root. Its historical original preparation/budget references remain intact; the new execution budget is in the new postfit contract, not patched into the comparison. Original five A19/G21 models, model hashes, test roles, query identities/multiplicities, t_set conversion and native NaN G extras remain unchanged. No refits, feature computation or candidate-label reads occur here. All493 canonical non-exact cases receive new predictions from the original frozen models. The333 old feature results and160 continuation feature results are not old predictions to reuse: they are the two sources for these493 new predictions. The separate52 exact gate cases reuse their original prediction hashes and per-part metrics without target reads or recomputation. Seven invalid cases stay in the552 denominator, with eligibility loss7/552. Negative model values are retained, never clipped. Use unchanged `phase2_gate_outcome_statistics.two_metrics/summarize`: each model/metric uses only paired parts with both predictions defined, averages those paired parts within each component, then weights components equally. Partial components remain; only components with zero defined paired parts are excluded and counted. No imputation.2000 paired component bootstrap replicates, seed2026111408. Both A and B13-E1b must have field CI inside[-.001,.001] and last-location/Rg CI inside[-.010,.010], including boundaries. Failure is not proof of non-equivalence or failure of every gate route. The original exact-set FAIL remains preserved. ## Canonical input and review barriers Original root remains STOPPED with333 actual PASS/CPU66936.020178167. The continuation root must have CANONICAL_FEATURES_COMPLETE_POSTFIT_REVIEW_PENDING,333 explicitly reused originals,160 unique new PASS at indices333..492,493 canonical count, complete finite accounting, feature Git freeze, settled post-freeze cost and no global stop. Both original and continuation queue/guard/worker processes must be absent. No original493-complete receipt is fabricated. The canonical manifest must reproduce each original333 entry exactly plus its explicit reuse origin and original path;160 new entries must use continuation paths and origin. Check result/guard receipt identities, gate IDs, output hashes and original case/index for all493. Preserve original52 exact entries and7 invalid cases. The implementation can be reviewed before these future outputs exist, but execution cannot adopt partial output or guessed hashes. New `expected-review-binding.json` contains the implementation commit, contract SHA256, implementation-proof SHA256, unchanged comparison SHA256 and full code map. Require a new actual MeltAssist full semantic PASS matching all five fields, frozen with the binding and contract as implementation-review, followed by actual validation. No old PASS transplant. The runner validates it before adoption, each predictor before model deserialization, and score before any original score values or target arrays. After493 new plus52 reused predictions are frozen, the scorer reads exactly493 existing target caches;52 exact/7 invalid target reads are prohibited. New decodes and new fits remain zero. ## Shared budget and execution Final total cap remains104639.3107342874 CPU-seconds; postfit is at most3600 within it. Adopt the continuation's complete settled cumulative cost once, already including the original66936.020178167; the original amount is provenance, not added again. Require the continuation's post-freeze cost to remain below101039.3107342874 so3600 is still reserved. Reject undercounted worker/guard cost, doubled old prior, nonfinite ledger, changed caps and incomplete accounting. Postfit counts120 conservative preparation/setup CPU-seconds, coordinator CPU and the maximum of explicitly recorded worker CPU and actual reaped child CPU to avoid double counting. Five sequential fold predictors each have300 CPU/480 wall seconds; score has600 CPU/780 wall seconds,4GiB RSS,3GiB postfit output. CPU8, no GPU. Before every step reserve all remaining prediction budgets plus600 scoring and120 terminal work against both the3600 and total caps. No shortening jobs, retries or extensions. Single exclusive lock and started receipt prevent duplicate execution. Terminal/post-terminal accounting includes freezes. The entry launches no next stage. ## Preparation validation 24 CPU synthetic tests passed: ten new canonical ledger/provenance/remaining-budget/read-order cases, eight unchanged statistics cases, and six real temporary Git/file exact-review validator cases. Tests include333/160/52/7 accounting, stale original493 receipt rejection, wrong origins/paths/hashes, partial or duplicate new terminals, changed reuse flags, missing feature freeze and new PASS before model/target reads. No actual dataset arrays or scientific job were run. Independent review remains pending until an actual new matching receipt arrives.