3.4 KiB
One labelled gate outcome v2 continuation
Claude1794927353.190299 supersedes the95k cap after the original reservation stop, authorizing one labelled continuation after the S2 scorer review handoff. That handoff was posted1794927225.586049, with remote delivery8354a2f6. The actual original terminal has333 PASS and160 never attempted, correcting the332/161 discussion snapshot. Original root /home/frank14f/MeltMind-runs/phase2-gate-outcome-calibration-v2-2026-11-17 remains STOPPED, CPU66936.020178167, accounting_complete=true, no new labels; coordinator666520 exited. Do not modify or restart it.
New unique root /home/frank14f/MeltMind-runs/phase2-gate-outcome-v2-continuation-2026-11-17 reuses all333 terminal result/array/guard hashes in place. Only original indices333..492 may execute. The unchanged phase2_gate_outcome_features.prepare(root, original_contract, index) supplies the exact original algorithm and sealed gate IDs; no solver/threshold/feature changes, retries, targets, fits or predictions. CPU16–27 for up to12 workers, coordinator30–31, leaves S2 training resources unchanged. Each new part retains1800 CPU/1800 wall seconds,4GiB RSS, one process. Entire original plus new output remains within original12GiB;48GiB aggregate live RSS.
The final measured total cap is104639.31073428741 CPU-seconds, below110000. Formula: old settled66936.020178167 once +1.15×(old settled−original fixed setup600)/9360267 completed nodes×4096096 remaining nodes +new setup600 +terminal120 +postfit3600. Variable remaining estimate33383.29055612041 includes original worker, guard and coordinator cost; the historical setup is already in the old ledger and excluded from the per-node variable rate. Worker+guard-only observed62000.949475436 is separately lower and is not used to omit orchestration costs. The measured total, not just remaining cost or the110000 ceiling, becomes the contract cap; geometry cap101039.31073428741 reserves3600 for postfit. No further extension is authorized.
Before every dispatch reserve a full1800 per active/new job plus120 terminal overhead. Waiting for other jobs is allowed; a deficit when idle stops without shortening a job or retrying. Runtime cumulative CPU includes old ledger once,600 new setup, current coordinator and all worker/child costs using maximum of recorded and reaped costs, plus active process trees. Failures retain original and new evidence. A fixed started.json, exclusive lock and actual original/continuation process checks reject duplicate launch.
At completion the new canonical manifest explicitly identifies333 original hash-reused results and160 new results. The new complete receipt has original_reused_terminal, new_records and493 canonical count; it never manufactures493 executions in the original batch. New status is CANONICAL_FEATURES_COMPLETE_POSTFIT_REVIEW_PENDING, after which all features and cumulative costs are frozen. The52 exact-set reuses and7 invalid cases remain unchanged in the552 denominator. No postfit is launched by this queue.
The preparation-root interface and accounting change means the downstream postfit needs a new exact identity and full semantic review; it is not merely a budget-constant patch. The prior72000 PASS remains historical and cannot be transplanted. Actual launch and worker evidence are recorded separately after preparation and synthetic acceptance; this contract is not proof of execution.