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

44 lines
5.6 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.
# B13-E2 result: no confirmed increment from the fixed type partition recipe
**COMPLETE / PRUNE_FIXED_TYPE_CONTEXT_RECIPE.** On the frozen552 test parts /204 components, the raw full-roster T−G field and last-fill contrasts do not pass Claude1794904249's deployment gate. This prunes this fixed K8/seed0/readout/model recipe. It does not refute all type organisation, topology, connectivity or alternative operators. E1b's confirmed generic-context improvement over A remains a separate result; a crossing interval is not equivalence.
| Component-weighted mean | G generic | T true-type | C one frozen shuffle |
|---|---:|---:|---:|
| Field relative L2 | .069702214305 | .069979340102 | .069731025504 |
| Last-fill distance/Rg | .387589687577 | .406967897735 | .391364883691 |
| Kendall tau-b | .928112243828 | .927540088075 | .928272917633 |
Raw paired-component differences use2000 shared resamples, seed2026111408. Field/last negatives favor the first arm; tau positives favor it. Every contrast has204 defined components and zero failed raw replicates.
| Pair | Field difference [95% CI] | Last-fill difference [95% CI] | Tau difference [95% CI] |
|---|---|---|---|
| T−G | +.000277125797 [−.000752536853,+.001182127879] | +.019378210157 [−.020869784012,+.064202312812] | −.000572155753 [−.001619798512,+.000527980831] |
| T−C | +.000248314598 [−.000629333506,+.001138688097] | +.015603014044 [−.040656365713,+.070638579663] | −.000732829558 [−.001714123025,+.000177368349] |
| C−G | +.000028811199 [−.000830569853,+.000875057407] | +.003775196113 [−.042824699146,+.050936945873] | +.000160673805 [−.000814158218,+.001263054978] |
The raw T−G field upper bound is+.00118213, and the last-fill upper bound is+.06420231. Neither branch of field.upper<0 OR(last.upper<0 AND field.upper<.001) passes. Adjusted comparisons never replace this gate.
## Descriptive adjustment and its support limitation
The full-panel origin is supported and estimable for all three pairwise regressions (rank3). Each pair has479 dual-valid parts, but their original component weights differ: T−G .8669297571; T−C .8765497574; C−G .8621412042. Regressions use the original uncentered log cut-area and log RMS-diameter ratios; all invalid/descriptor-undefined parts keep their raw differences and original weights. Each component resample refits the regression.
| Pair | Adjusted field | Adjusted last-fill | Failed replicates |
|---|---|---|---:|
| T−G | point+.001190114666; **CI unavailable** | point+.073816066637; **CI unavailable** |274/2000=13.7% |
| T−C | −.000234446234 [−.002131825878,+.001495600603] | +.042868084472 [−.038054301974,+.151865477716] |4/2000=.2% |
| C−G | −.000425699574 [−.001814382072,+.000549689575] | −.012627550664 [−.124047325222,+.066587709137] |1/2000=.05% |
Every failed replicate above is ZERO_POINT_OUTSIDE_SUPPORT. T−G exceeds the prespecified5% ceiling, so no adjusted interval is reported or used. The remaining adjusted intervals cross zero. These conditional descriptions do not identify a direct causal type effect; the fixed shuffle also changes spatial autocorrelation, and the bootstrap excludes retraining/alternative-shuffle variation.
## Actual intervention, missingness and execution
G/T/C valid parts are510/513/507. The eight validity combinations (GTC order) are000:7,001:1,010:7,011:27,100:4,101:27,110:27,111:452. No cross-arm filtering occurred. Each pair has479 dual-valid readout comparisons; zero are bitwise identical. Per-fold learning/test component coverage, descriptors, cut-interface differences, FP32/readout differences and missing-category contributions are retained in coverage.json and the complete result rows. Identical-readout counts are not a claim of sufficient causal intervention.
Geometry ran once:552 guard PASS,3868.411711768 CPU-s. Postfit ran ten fresh T/C fits, ten prediction jobs producing1104 arrays, then one score job; all21 guards PASS. Five G models,552 G predictions and original G metrics were reused by hash. No new G fit, geometry, label decoding, validation gate, GPU work or retry. All ten checkpoints were locally frozen atd0f2b7d78e747ba183e42fc2cd6c4dc4185a11e0; all prediction hashes and scorer at16657833f5c0f9fcf3f9c8c2cf3a6257c20b5349 before test access; terminal science at88809b00abc27449d3acef57642d55636c859c94.
Cumulative CPU including the terminal freeze is6631.254902427/14400 seconds (geometry counted once). The pre-freeze complete receipt is6629.985108592; post-terminal-freeze-cost.json preserves the additional cost rather than rewriting it. New worker costs arefit2106.523620,predict77.891149,score12.391026 CPU-s; guards/setup/coordinator/barriers are included in the cumulative ledger. Peak fit tree RSS2741313536B is below8GiB. Fourteen synthetic checks passed before launch.
The terminal read-only audit checked source pins, all five local barriers, ten model hashes and actual learning roles,1104 prediction hashes, the bootstrap archive,552 distinct test accesses and absence of E2 processes. It did not recompute metrics or bootstrap. Contract22b37303de730e69b06726fb7ab099138e869929cdfd284736f644993b5bfa65; result SHA256ce93404a879747d8129698b84d929897cc4c9725a4236e9bba11858b18671de1. Root:artifacts/phase2-b13-e2-postfit-2026-11-17. Remote publication remains pending the accepted transport change; local freezes and Slack review delivery are not claims of remote Git completion.
The next decision belongs to the existing Claude-led stage review. Preserve all three arms, original missingness and failures, raw and adjusted results, and the independent type/connectivity hypotheses. No restart, alternate shuffle, threshold change or new scientific computation is implied by this terminal report.