Files
MeltMind/docs/phase2-b13-e2-postfit-contract-2026-11-17.md
T

12 lines
3.7 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# E2 postfit continuation: frozen geometry to raw and adjusted comparison
Claude1794904249 and1794907827 authorize this continuation. The original geometry stage is complete:552 guard PASS,3868.411711768 CPU-s. No geometry or G fit/prediction is repeated. Two arms T/C, five frozen 100% learning pools, yield exactly ten new fits and1104 predictions on the original552 test parts/204 components. G predictions and metrics are reused by hash from E1b. All test parts remain regardless of arm validity; invalid two-column readouts stay nativeNaN. Learning logs record the actual fold and arm, correcting the old helper's metadata-only hardcoding without changing the model recipe.
Order: validate geometry results/feature hashes and all frozen fold identities; locally freeze contract/features/scorer; fit ten XGBoost100/depth8/.1/seed0/one-thread models with21 FP32 columns and unit-node weights; freeze all ten checkpoints; predict T/C in seconds on queries identical to frozen G including multiplicities/wall masks; freeze all1104 prediction hashes and original G manifest; then access cached test seconds once per part for metrics and run the fixed paired-component statistics. No new label decoding, HPO, validation gate, retries or GPU. Independent local Git barriers follow the already accepted6657 engineering backend; the remote writer/upload is untouched.
The cumulative14400 CPU-s cap includes original geometry once,60 seconds setup provision, all new guard/worker/coordinator/barrier costs. Original first-ten forecast passed at11299.812906. Actual geometry plus the unchanged6876.301638 postfit forecast and60 setup is10804.713350, below the same cap. Maximum four single-thread workers onCPU8/10/12/14, coordinator30–31; each fit1800CPU-s, prediction120CPU-s, score2400CPU-s; wall=CPU cap+180s;8GiB per worker,32GiB concurrent;12GiB combined geometry+postfit directory budget. Admission reserves full active job caps and120 CPU-s shutdown headroom; failures stop without retry. Final barrier cost is reported separately and remains inside the same cap.
Deployment decision uses raw full-roster T−G: field CI upper<0 OR(last CI upper<0 AND field CI upper<.001), otherwise PRUNE; undefined full metric produces PARK_UNDEFINED. T−C/C−G and Kendall tau-b are also reported. Adjusted field/last comparisons are descriptive only. The original component weights1/(204×component original part count) remain on all parts; dual-valid geometry contributes an uncentered weighted regression against log cut-area and log RMS-diameter ratios. No valid-subset renormalization. Invalid/descriptor-undefined parts retain raw differences and weights. SVD rcond1e-12 and zero-support/estimability tolerance1e-10 are fixed; an origin outside the convex hull or non-estimable intercept makes the adjustment unavailable. Bootstrap2000/seed2026111408 jointly resamples components, carries all their original parts, and refits each pair/endpoint regression in each replicate. At least5% failed replicates suppresses that adjusted interval; no replacement draws. Report eight validity combinations by fold, full-component coverage, descriptors and original interface/readout comparison evidence. No causal or universal topology conclusion follows.
Fourteen synthetic checks passed before real fits: component weighting/resampling, retained missing-arm contributions, uncentered adjustment, zero columns and rank deficiency, hull support, strict5% boundary, strict raw decision/full roster, refit-per-replicate behavior, cumulative admission, nativeNaN, query multiplicities/wall identity, prediction-before-target barrier, and actual-fold training metadata/full missing rows. Scientific source files of completed geometry and other live branches remain immutable.