JooYoon/riidolaya-shortclaim-data-effect-failed-79

Failed data-only claim model79 — inactive research archive

This archive preserves a completed experiment that did not improve candidate inspection order. The model needs31 candidate checks where the existing lexical control needs27, and gets the correct candidate first on2/10 answerable validation requests where the control gets5/10. Keep the lexical control and existing verification path. These weights are inactive; qualification and production flags remain false.

한국어 설명 · Results · Configuration · Provenance · License scope

What this artifact is for

The research asks whether a very small model can suggest which candidate to verify first. It does not decide whether code is correct, generate code or text, eliminate candidates, or replace an independent verifier. Inspection counts here are simulated ranks under saved labels, rather than actual LLM requests, token bills, or verifier executions.

This is a scratch-trained,8192-dimensional lexical feature-hash linear model. Request/document unigrams and ordered bigrams produce signed cross features, plus a lexical-overlap feature. Its learned associations do not establish semantic reasoning or rules. It is not Laya fine-tuning, a transformer, a GPU/MPS model, or a1.58-bit/ternary checkpoint. No pretrained Laya weights enter this experiment.

What changed, and what stayed fixed

The previous76-parent development corpus is preserved, with3 qualified source80 training requests and9 candidates appended. The new requests concern Google UUID parsing/scanning behavior and a selected go-humanize Ordinal slice. Their labels come from bounded behavior comparisons;72 candidate executions are not72 independent user requests or a universal proof for all inputs.

The joined corpus has79 requests and235 candidates:41 answerable,17 no-answer and21 unknown, across21 whole-source groups. Roles remain47 training,20 validation and12 calibration requests. Training has100 candidate rows, of which18 retain zero loss weight; validation has45 rows, of which1 retains zero loss weight. Unknown supervision remains nullable and excluded from fitting and truth utility. Calibration remains excluded from fitting, epoch choice and this utility check. Existing masks and the old validation candidates were retained.

One fresh FP32 fit uses seed1729,50 epochs,batch128,learning rate0.1,L2=0.0001,8192 coefficients and no ranking term. The earliest strict minimum validation BCE selects epoch50. Training NLL is0.656502358851809 and validation NLL is0.6662350441005636. There is no warm start, seed sweep, loss search, threshold relaxation or extra fit during packaging. The trainer retains FP64 shadow/gradient/loss calculations while scoring coefficients are rounded to FP32. Both the old91-row and new100-row training sets are below batch128, so each epoch uses exactly one batch:50 epochs still mean50 updates. The fixed8192-dimensional signed hash-feature schema is unchanged; there is no learned vocabulary or new model-class schema. Adding nine unit-weight rows changes the gradient inputs and the sample-weight normalization denominator from73 to82; the seeded row sequence can also change. These numerical/input changes and the three new semantic cases occur together. This run does not identify which change caused the observed result.

The model is a fresh data-effect sibling of the failed first FP32 model71. The rank-plus-BCE model72 is an objective sibling with a different loss. Neither is a parent checkpoint inherited by79, and79 is not their compressed child.

Result and limits

The reused validation contains15 known requests:10 answerable and5 no-answer. Five unknown validation requests remain audit-only and excluded from truth utility. No-answer requests exhaust their candidate list. Top1/Top3 denominators include only the10 answerable requests.

Method	Candidate checks	Top1	Top3
Existing lexical control	27	5/10	10/10
First FP32 fit71	31	2/10	10/10
Rank-plus-BCE fit72	33	1/10	10/10
This data-only fit79	31	2/10	10/10

Relative check reduction for79 is −14.8148%: it requires more work than the control and fails the unchanged5% reduction and Top1 preservation criteria. The oracle needs21 checks. Matching71's aggregate outcome does not mean its weights or every candidate order are identical.

On the3 newly appended training-origin requests, the model needs6 checks and gets1/3 first choices correct; lexical control needs4 checks and gets2/3 correct. This is a diagnostic, not a held-out or generalization gate. Each request has3 candidates, making Top3=3/3 trivial.

The same validation selects the epoch and measures utility, and its earlier outcomes influenced subsequent experiments. These are developmental regressions, not protected-final generalization results. A fresh final corpus of at least2400 distinct semantic requests per domain remains outstanding. There is no measured Codex cost saving, no production claim, and no default router activation.

Size and CPU observation

The unchanged custom file is32792 bytes: a24-byte header plus8192×4 FP32 coefficient bytes. The reference Go decoder expands coefficients into65536 bytes of FP64 backing storage. Neither number is whole-process memory usage or proof that a smaller disk format makes this scorer faster.

The actual macOS child, including controls,fit,trace,model I/O and checkpoints, finished in1.244151917 seconds; OS MaxRSS was36782080 bytes, or35.078125 MiB. User CPU time was0.203462 seconds and system CPU time0.081739 seconds. The soft Go heap setting was256MiB, with one Go scheduler CPU. OS RSS is an observed process measure, separate from Go heap and GPU attribution. This is one CPU observation, not a distribution of inference latencies or a training-kernel-only benchmark. See original OS receipt.

How to use the archive

First read the failure,configuration,provenance and checksums. Once the separate Root publication proof supplies an immutable HF commit, download that revision and verify SHA256SUMS. Keep the experiment's runtime registration and routing activation off. This custom .hbin is readable by the project's Go research reader; it is not a Transformers AutoModel, GGUF, ONNX model or hosted text-generation pipeline. The decoder/feature contract is internal to the repository; no standalone stable inference SDK is promised here.

The model file SHA256 is dff05140098845943ece87c3ab8a31a15564a4b170a59ab017ee393501158004. Raw plans,private host paths,tokens,pprof traces and GPU traces are excluded. Model bodies belong on Hugging Face; GitHub receives numerical results,license/provenance metadata and immutable HF references.

License and provenance

The project's authored scratch artifact and documentation use Apache-2.0; LICENSE and the retained project notice accompany this archive. Upstream source fixtures retain their own terms. Complete original MIT notices for semver/doublestar from the earlier corpus are retained. Google UUID revision2d3c2a9cc518326daf99a383f07c4d3c44317e4d carries BSD-3-Clause; selected go-humanize Ordinal revisiona1b4e66b9a6d890e9e15e7091cf16c8032367d6e carries MIT. Their complete licenses are in licenses/ and their scope is described in NOTICE.

Humanize number.go and its separate gorhill/WTFPL lineage are excluded from the selected composition; this archive claims no whole-package clearance. Attribution is not endorsement. Existing public source and archived license pins support this scoped redistribution review; it is not a complete ancestry or legal certification. No change to the failed scientific qualification is implied by licensing or archiving.

The reusable Go implementation is pinned to source commit01ce47dcca16c8f82022c97354d82105b6819875 and source CI37015530233. Later CI37023314793 passed before this run with the relevant Go sources unchanged. Publication of this experiment's records still requires the separate exact-head Git CI/merge evidence and Root's inactive-archive publication plan; preparation itself makes no external write.

Go research source · Project · Failed71 immutable archive · Failed72 immutable archive

Integrated evidence and inactive archive gate

PR105,headbb19a422a7a88e58230bf7c6424cc8488d65f19b,merged as cf648a5e6489d864e05fdf4b81700ce265170b3f. Root verified that final CI37032449547 passed all four required checks. The first Ubuntu attempt failed a dependency download with HTTP/2 INTERNAL_ERROR before tests; one failed-jobs rerun used the same head without changing code,gates or metrics. The final official API marks all four job records run_attempt2,including macOS/secrets records whose original timestamps were retained under new job IDs. The separate ROOT-CI-FINAL.v1.json and ROOT-INACTIVE-ARCHIVE-POLICY.v1.json bind Root's formal evidence and inactive-archive authority before final checksums/upload. This preparer performs no external publication; the CPU/BCE utility failure and activation=false are unchanged.