Finite-input claim audit56e results

A separately scoped v2 connects actual function observations to limited truth. Freeze source/prose/plan first, then complete one official audit with zero retries: four sources×nine inputs=36 observations. Eleven of12 derived request rows have a matching candidate in their finite scope; one has no answer. Learned-model training, utility and LLM cost savings remain unfinished.

A source can satisfy the quoted-comma inputs while retaining partial output on syntax errors. Copying its whole-function wrong role into every smaller property would create incorrect labels. Instead state which property is checked on which inputs, and link observations only to that property's labels.

Observations and denominators

Listed input set	Sources agreeing on every literal / four	Matching candidates in four derived rows
Quoted commas: two inputs	3/4	2,3,2,2
Outside escapes: two inputs	2/4	1,2,1,2
Syntax error and zero output: five inputs	1/4	1,1,0,1

Six of12 source/property combinations agree with all listed literals. Actual source/input observations are18 matching/18 mismatching among36. Referencing them in36 row/candidate labels also yields18/18: the identical counts are coincidental, and the denominators differ. These intentionally defective source controls are not50% model accuracy. Every input is checked after an earlier mismatch.

Observer dispatch attempts36, actual candidate calls36, completed36, observer failures0 and verified Git blobs36. Exact Error, Count, all eight Tokens and Panicked comparisons reject zero output that omits a required syntax error. Complete collection does not mean every candidate is correct.

Frozen links and raw evidence

• Source/prose/review/unit controls/modules freeze: 1f9631c2819abf24ca57e6317ed48b8397be0881.
• Input/plan freeze: ca5ff83b847e984dab628be536cc2462ddff8221.
• Actual Go1.27.1 Darwin/arm64 CGO0 trimpath executable:8,824,130bytes, SHA b5b5f97acb48dd3c1d6f472effee6e56d6d094e7e2b25db4ab0c68d06c792612. Executable excluded from Git.
• Input:30,846bytes, SHA 86556bd96599ae013441f199f90c8cd619a011d52212d1ccbfdbe98b6f6a1ec5.
• Plan:5,901bytes, SHA 6a21ff6da7772476e070509b45bf9506b95f50473841bf3183d17a2fca7ded69.
• Raw result:59,125bytes, SHA 4a50160b99ea059a8de6e7be92ba22b5017f94dec2d3ca2fea644bda759a1be8.

The three JSON files total95,872bytes. Before observation, verify21 compiled sources plus13 support files against source-commit blobs, and two inputs against a separate input commit. Executable/canonical JSON/package-list/text/scope/closure mismatches refuse rather than supply truth. One official collection occurred. The later publication_test.go is not retroactively among the34 pre-observation pins. CI replays the observations without adding official attempts or samples. Actual CI/merge completion is linked in subsequent issue19 execution records.

Progress and remaining requirements

Keep56c's12 requests and36 candidate captions pending and unchanged. New v2 is an AI-assisted synthetic problem explicitly declaring hex inputs; it does not solve general natural-language function semantics or the original problem. Document an independent agent source-reading review, not nonexistent human approval. Preserve four original unknowns and72 parents/17 total/16 labeled groups. All four sources share one helper family: zero new independent parents/families.

Each prepare/audit preflight's192 normalizations validate bounds only; numeric encoders/ranking features/rankings0. Roles/splits/fits/new models/weights/model calls/paid calls/protected-final/CoSQA reads/activation0. GOMAXPROCS1/256MiB soft heap are settings; CPU/RSS/GPU/latency are unmeasured. Keep56d's unexplained first577ms and all-BM25-fallback evidence unchanged.

Useful tiny-model progress now needs diverse real sources and utility. The fixed semver/glob investigation completes reading and concrete contract proposals, with zero newly acquired requests/truth. Actually acquire sources toward120→240, closed compilation/rights, independent contracts/authorship and relationship groups, then freeze a model-free utility test separately. Preserve minimum15 groups,5% necessary utility and2400 distinct protected-final requests per domain. Training eligibility remains false; no new HF weights.