Freeze approved B15 fixed tet-speed supervised solver gate

This commit is contained in:
Frank14f committed 2026-11-17 16:23:00 +08:00
1 parent 47c85a8a14
commit 7a91633244
7 files changed
+108854

No files matched your search

+11
View File
@@ -0,0 +1,11 @@
# B15 approved fixed one-fold execution
Claude1794901850 approves Mold specification6c33da33 as written, after B14, <=7200 workerCPU-seconds/single launch/no retry. B14 is terminal at its fixed-ten cost gate; it did not fit or score. This does not authorize B13-E1/E2 or another B14 attempt.
Independent CPU28 queue; original learning320/validation111, original A baseline, one XGB100/depth8/.1/seed0. Learning target is negative log norm of FP64 P1 gradient of u=T/t_set in xi=(x−mass centre)/Rg. Geometry kappa2<=1e8; source-containing tet, invalid arrival, numerical plateau and invalid gradient masks preserved separately. Recording resolution unknown:delta0/numerical ties only. SHA256 length-prefixed caseID+original uint64 tetID fixes at most4096 samples per learning part before target access; no resampling. Weights physicalV/pi, global mean1. No clipping, epsilon speed, target infill or validation supervision mask.
Exactly columns4:19 of original A19, dropping F plus its three gradient columns. Frozen gre24_features_v2.COLUMNS and spatial_moment(g,di) establish original column identity and Di-filtered moment neighborhoods; average four vertices per tet. No original F/Df-dependent feature is introduced to the15-column fit.
Predict exp(logspeed) on every original validation tet, require positive finite speed and representable square; use original R4(a) tetFIM on normalized xi, original gates and100000 iteration limit, unchanged coordinate/speed scaling rule and convergence tests. Strict full-node volume CDF preserves ties; original fold0 learning c gives normalized T, no validation time calibration. All original tets retained; failed/incomplete predictions park the gate without dropping hard parts.
Preparation subcap1500CPU-s, first10 cost projection1.25 times node-scaled measured cost, perpart30CPU/60wall. Fit1500, prediction3600, score180, all including guards/setup constrained by7200 total with120 drain reserve; RSS8GiB/output5GiB. All geometry/sampling fixed then Git barrier before learning targets; model frozen before prediction; predictions+scorer fixed before validation scores. Original B12 strict OR gate unchanged; additional Kendall diagnostic does not select model. No test labels or GPU. Seven CPU synthetic checks PASS; no execution claim until real receipts.