5.3 KiB
R4bc final postfit: original all-pool continuation
The original queue is actually terminal: FIT_QUEUE_TERMINAL,1005 attempts=1004PASS+original job90 process_count_limit,105 final modelsPASS, accounting complete, no global stop/cut/monitor error, no never-attempted jobs. Original coordinator521868 and all R4bc processes were absent at handoff; no existing postfit process/output was found. No fit is restarted. Training fit97267.128466CPU-s, original setup99.840649; guard6096.497806331 and coordinator10618.968657329 are separate host overhead.
Root artifacts/r4bc-final-postfit-2026-11-17. Copies of the original contract, actual batch-complete and actual models-terminal are byte-identical, not manufactured. terminal-input-audit checks all1005 guard records, successful result hashes and all1004 model checkpoints;105 final checkpoints and35 original100% models match their prior frozen identity. Checkpoint/metadata CPU2.490779 is covered by a60CPU-s conservative reserve. Seven new synthetic adapter checksPASS/0.505717CPU-s: full terminal/failure/model identity, missing/duplicate/extra-failure refusal, unchanged interim models, cumulative CPU/shared wall/score reserves, no100% regeneration, no duplicate stage, cryptographic Git proofs without subprocess inside the single-process guard.
New tools/r4bc_final.py only orchestrates immutable original r4bc_postfit.predict/score, with the previously accepted pure Git-object proof verifier to avoid creating Git subprocesses inside a single-process guard. No original pinned source changes. All original training evidence and checkpoints are Git-frozen before prediction. Reuse4968 frozen100% entries (3864 model predictions+historical H0/EIK) under exact hashes; generate only70 remaining final-model predictions once, sequentially onCPU24 with coordinator25. Then freeze the complete prediction manifest and source before any new scoring access to cached labels. Score with original functions over all100/50/25 pools. Finally tools/r4bc_interim_compare.py verifies contract,35 model hashes,100% prediction arrays, part scores and summary/contrasts/bootstrap at zero tolerance; any difference is DEFECT, never overwrite interim.
Original post14400CPU-s remains shared. Interim346.870387+new synthetic0.505717+metadata reserve60=initial407.376104CPU-s; all new guard CPU accumulates, failures retained, no retry. Each prediction1800CPU/3600wall/8GiB; scoring7200CPU/10800wall/8GiB; comparison600CPU/1800wall/8GiB. Prediction dispatch retains7200 scoring+600 comparison CPU reserve; scoring retains600 comparison reserve. Total worker144000 and original fit/setup caps remain unchanged. Shared post wall starts at original interim predict-start1794859566.1669283, ending six hours later (10:06:06 Asia/Shanghai); waits/setup/publication do not reset it. Aggregate disk100GiB; single worker plus coordinator is within36GiB. No newGPU, fit, target decode, feature computation, retry, hyperparameter or test-based selection.
This document initially describes validated orchestration and freeze. Actual predict/score/compare-started and guard receipts determine stage state; completion is not inferred from service creation. Main authorization is the original R4bc contract and accepted interim continuation. Claude/TRON received actual terminal handoff1794874315.210029; no repeated user-go or new scientific direction.
Latest controlling ruling: Claude1794874796.184549 supersedes1794874336.044949. Preserve the already launched70 chain: new70→freeze→score→exact INTERIM100. Afterwards independently re-predict the35 full-pool models into a labelled determinism root and compare arrays bitwise. Earlier discussion of inserting35 before this scoring was not implemented and is withdrawn. No running source changes. All70 prediction guards nowPASS, worker412.869932CPU-s; post cumulative820.246036. Complete manifest14904 entries=7728 new70 model arrays+3864 interim model arrays+3312 historical H0/EIK entries; missing=[]. Full scoring is still gated by the next Git freeze.
Claude1794874951.316109 explicitly fixed the final sequence and withdrew the crossed earlier message. Executed accordingly: all-pool predictions locally frozen248542d23de8e737b50029b4502b577da4b098aa, original score guardPASS/38.284985CPU-s, exact INTERIM_100 comparePASS/7.881043CPU-s.35models/4968predictions plus100% part scores/summary/bootstrap EXACT_MATCH at zero tolerance. Main-chain post cumulative866.412064CPU-s. This alone did not test fresh100% prediction determinism because those entries were reused; the separately authorized35 check follows in artifacts/r4bc-determinism-100-2026-11-17. Its5 synthetic checksPASS; first fixture-only missing-directory failure0.271408CPU-s preserved, accepted attempt0.818123, both charged. No live or old source modified.
FINAL SCIENTIFIC TERMINAL: the separate35 new predictions allPASS; bitwise3864 model arraysPASS, no defects. Main70+score+exact comparison alreadyPASS. All original/new coordinators absent at final host check. Post1149.437161/14400CPU-s and total worker98516.406276/144000; four metrics undefined0/bootstrap undefined0 across all552parts/204components and three pools. Scientific functions and old failures unchanged. Report docs/r4bc-final-results-2026-11-17.md. Original full training push54821 exited0 and remote05343e13 verified; final scores/determinism publication follows once, no duplicate training push. Do not restart/re-score/retrain.