2.2 KiB
B15 one prediction-only continuation
Claude1794907848 follows the explicit B15 projection stop, after E2 actual start. Original model/algorithm/parameters are unchanged. Total8000 CPU-s (+800) is final; prior3185.972043089 is counted once. No refit, no geometry work, no second prediction of the completed first part. Copy and hash-verify the fitted model and first query/full-field predictions; reference immutable431 geometry outputs. Remaining110 validation predictions run once, then the original B15 score function once after full111 prediction freeze. Original stops and partial data remain.
The new per-part computation is AST-identical to the old prediction loop, except progress records add cumulative completed-part CPU and the iteration starts after the reused first row. Same15 features, model, positive-speed checks, normalized geometry, FIM solver, original gates, volume clock, query selection and A baseline. Three tests passed: scientific loop identity, prior accounting and completed-cost projection, invalid-support rejection.
Projection keeps factor1.25 and the8000 stop: charged cumulative CPU +1.25*(completed prediction internal CPU / completed node weight)*remaining node weight. The reused first part supplies its measured32.152705187 CPU-s only to forecast calibration, not a second budget charge. The earlier forecast included in-flight next-part CPU in the completed-part denominator; this engineering correction aligns cost and support. New startup/in-flight/guard/coordinator CPU still enters actual cumulative accounting. No forecast adjustment changes prediction science or budget. Actual total reserves120 CPU-s for control and stops before exceeding8000. A new prediction worker has only the remaining total cap, not an extra stage allowance.
One prediction worker onCPU4, coordinator29; original8GiB RSS. The12GiB storage cap counts the old referenced root plus new files, preventing a symlink from hiding reused geometry storage. No network or H2/B14 changes. Local model/feature/contract proofs precede prediction, and full111 prediction/scorer proof precedes scoring. If forecast, actual total or disk limit fails, stop and report without further extension. Actual start is a separate receipt.