Task Results 53 EN

PDCA 53: repeated behavioral attempts and public Go source acquisition

한국어 · Frozen plan · Pre-outcome Laya observation · Numeric results

One new behavioral request, repo-keyword-language-guard, was attempted twice per explicit requested profile. All four fixed CLI attempts passed independent requirements and had zero exit with complete core usage. Four executions remain one new request. Their mutable source bytes were identical. The verified keyword guard is also integrated into the actual repository-selection runtime.

Cumulative coding evidence is four distinct development requests,12 owned records,two code families,one repository. It is not an executed120-request development corpus or a2,400-request protected final per domain.

Plan

Plan SHA 2205505aeaf03e331ff1d3872c2aee6fe11a1f9668c3c7574f0d30f41cdb2390 was frozen in pre-outcome commit ebd1507c957e8ac3991a2e71cf58eeb5b8f8f8f4. Fixed order is Luna / Sol6 / Sol6 / Luna, concurrency one, main120s, independent checks45s, no executor retry/resume/substitution. Repetitions were preselected to inspect variability.

The executor is a clean -trimpath build of unchanged CI-passed main ee72334166e2962b0821d5198a50fdd13f92ab29. Go metadata recorded that revision and vcs.modified=false; executable SHA is 5b801da856ed824a11372a9fb705901f3624f09bd1b52283c46efe28a3420c6c. The public task retains separate source revision 146b02b9c37e6d90a11386050ccfdffc036c113e and its already published version1 contract. The new parser contract and acquisition inventory were not used for these attempts.

A read-only catalog refresh in an isolated authenticated home without settings/cache listed both profiles with low. CLI and Go executable pins were unchanged. Metadata does not attest access or backend identity; every observed_model remains unknown. Existing ChatGPT login was used, with no API-key billing,new endpoint,cloud training job,or purchased credits. Money,subscription usage,and backend call/retry counts remain unknown.

Laya before coding outcomes

The prior candidate inventory authored a fast hypothesis. Separately frozen plan53 proposed standard before this Laya observation or coding outcomes. Both are authored hypotheses, not measured labels.

The exact task prompt was processed once by fixed base Laya INT8,CPU one thread,threshold0.9. It proposed standard with maximum score0.6692456483 and abstained. The strong result naming Astra is a configured fallback; no Astra coding attempt occurred. The new fast/Luna,standard/Sol6,strong/Astra mapping applies only to this observation. Earlier predictions and their6.1 Sol mapping remain unchanged.

Whole-process wall time was1,273ms; application-reported ms was860.411ms. OS peak RSS was1,503,657,984bytes(about1.50GB/1.40GiB),peak memory footprint1,473,218,696bytes. This single cold observation includes native loading and is not warm latency,Go heap,GPU memory,or remote coding-model resources. It does not meet the ultra-small-resource target. All11 model/runtime/attribution files matched the unchanged manifest.

The router artifact is a -trimpath -buildvcs=false build from an immutable public Git source archive and has no VCS metadata; distinguish it from the cleanVCS executor. The prediction was frozen in pre-coding commit 469dc8ffcf72d55046a8bb0ec498a50e94ada099. Confidence is not calibrated coding-success probability.

Do / Check

Ordinal	Requested profile	Main ms	Verification ms	Input	Cached subset	Output	Reasoning subset
1	Luna low	36,006	4,108	79,359	61,184	1,131	0
2	Sol6 low	38,435	4,105	80,278	64,256	1,112	9
3	Sol6 low	58,658	4,254	115,974	104,192	1,646	67
4	Luna low	38,788	4,391	64,194	50,944	1,001	0

All attempts recorded accepted source closure,exit0,complete capture,process-group cleanup,and copied-auth removal. Reported core CLI usage totals339,805 input,280,576 cached input,4,890 output,and76 reasoning. Cached input/reasoning are subsets and never added twice. Main171,887ms and verification16,858ms exclude parent preparation and are separate. Token totals do not estimate money or subscription quota.

Changed candidate test files were hashed and preserved but never used for acceptance. Only the declared mutable source was executed with pinned base tests and the independent contract. Other files,caches,and additions are unassessed. Each public terminal JSON was byte-identical to its private durable record.json. Raw traces,authentication,absolute local paths,and model weights remain unpublished.

Both Sol attempts passed yet used38.435/58.658s and80,278/115,974 input. Luna used36.006/38.788s and79,359/64,194 input. Two observations do not estimate a population distribution. Luna had lower total input, but uncached input reversed the ordering:31,425 for Luna versus27,804 for Sol. Total input alone cannot establish which profile cost less.

Act

Runtime router.go adopts the first completed accepted source, SHA e0d9ddb2373604e41b35e151e7b593174dbdc005410ab5132567f15285d7883b. It was compared with the unchanged public runtime base. The published independent contract became runtime regression coverage with test names renamed; candidate-written tests were not adopted. Only selected-candidate keywords are inspected, preserving nil-Judge behavior,query guards,and lexical fallback. No allocation,shared state,or lock was added; scanning remains within existing8-candidate×32-keyword×128-byte limits.

The public Go acquisition survey links120 authored behavioral candidates across8 repositories to revisions,actual files,and license evidence. This is not120 execution-eligible requests. Independent acceptance,closed source scope,file attribution,dependencies,isolation,and invocation budgets still need verification. Candidates have no training labels or final eligibility. Connect templates,shared code,parents/children,and forks before splitting.

A separate new parser task defines top-level JSONL key uniqueness,count64,and decoded-byte length128. Independent prepublication review reproduced six false acceptances: skipping keys before type,overrestricting nested startup keys,changing legacy usage-error categories,retaining unknown key names,hiding retained names only from JSON,and retaining non-string unknown values. The strengthened contract rejects17 authored incorrect implementations and accepts a correct control. Final independent reruns also rejected all three privacy mutants and accepted the same correct control. Error precedence for inputs combining multiple invalid conditions is outside the assessed scope. The parser task has zero actual model attempts. Previously frozen contracts and these four attempts remain unchanged. Authored validation controls are neither model outcomes nor additional golden requests.

No fitting,calibration,weight release,or Codex policy activation occurred. Protected search final2,402 remains unscored. Next actual execution needs distinct requests,repositories,families,and paired success/failure differences. Both profiles passing this one request cannot establish useful training signal or a savings policy.