Next experiment56c: scoped property claims and utility before fitting

한국어 · 56b results · Existing acquisition plan

56b's17 total and16 labeled groups clear preparation minima. The data still come from one synthetic English authoring process; roles, splits and new baseline utility have not been frozen. Clearing a minimum alone does not authorize fitting or replacing existing weights.

The fictional user is an agent inspecting code. Given a small request such as which candidate clears output on error, a cheap tool proposes only inspection order. An external independent verifier checks actual code; preserve every candidate and fallback. Repeated hints are not correctness probabilities, execution approvals or whole-function certificates.

Prepare two separate paths.

Path	Required preparation	What passing does not establish
One-property claims	A single scope, observation and literal oracle for preserving quoted/escaped delimiters or zero output on error; new schema, plan and caption review	Code satisfying one property also satisfies the whole function contract
Utility before fitting	Fixed-order, BM25, lexical, narrow-rule and oracle comparison with separate answerable/no-answer/unknown sets, common cost denominators and retained candidates	Oracle headroom is achieved model improvement or reduced LLM usage

The same source can have different acceptable sets for different properties. A candidate may clear output on initial non-ASCII validation but retain partial output after a later syntax error. Do not reuse full-contract correct/wrong control roles as property labels. Author literal observations for each scope; unsupported cases, incomplete captions and unreliable observers remain unknown. Preserve56b's4 unknown delimiter requests.

Multiple properties of one function remain joined by parent, code, negative and helper relations. More properties, paraphrases, seeds, siblings or INT8/PTQ children do not create independent groups. Testing a narrower semantic model requires diverse public-source development requests120→240 alongside language, condition and authorship coverage. Verify rights, fixed revisions, closed compile scope and independent truth first; preserve the historical120-candidate source survey.

After sufficient data and headroom exist, freeze training/validation/calibration roles and global group relations before outcomes. Keep the minimum15 and5% necessary utility condition. Record FP32 parents, their INT8/PTQ children and equal-budget ternary STE siblings separately. Execute the existing proposal of at most8 fits/16 derivative artifacts only under a new training execution plan. Measure archive bits separately from expanded inference arrays in memory.

Resource work also needs a separate plan. Start with uncached resident JSONL processing of identical versus distinct public requests; separate startup, warmup, actual processing, pipes, controller work and whole-child RSS/CPU. Then consider open-loop load with a bounded queue. Do not claim cache, lock, SIMD or GPU gains before observed reuse and end-to-end benefit. This document does not record already completed experiments or new measurements.

Preserve7 distinct actual coding requests,20 records and exhausted call ledgers. New paid attempts require a separate execution plan and budget record. Do not read existing protected final or CoSQA reserves; at least2,400 distinct protected-final requests per domain remains a separate target. Publish only new models passing utility, rights and CI checks as immutable Hugging Face versions. Git contains code, plans, public numbers and model references only.