Data and offline verification materials for Checking decisions and diagnostic conclusions in software fault diagnosis: two controlled studies using explicit rules and recorded states

Existing inputs, public rules, saved-state checks, separate evaluator reference sets, complete actual final answers, and minimal offline reference, strict-scoring and table-verification code for two controlled software fault-diagnosis studies. Study 1 comprises 12 groups, 36 initial views, 432 research dialogues and 719 actual research requests. Study 2 comprises 15 base inputs in three presentations (45 records) and 270 research answers. Truncated and empty final answers are retained. Original prompts are Chinese. The verifier reproduces existing descriptive tables without a model, network connection or GPU. These are author-controlled synthetic tasks, not natural defects or field-validation data. The recorded services were deepseek-v4-pro and qwen3.7-plus-2026-05-26. The generation limit per request was 4096 tokens in Study 1 and 8192 in Study 2. Exact model settings and collection intervals are included in the archive. Licence scope: CC BY 4.0 applies to author-licensable materials and documentation, and MIT applies only to the six minimal offline Python files. Original final model-answer strings, actual assistant-answer history and service-returned fields have a separate source and rights notice. RIGHTS_NOTICE.txt and LICENSE_SCOPE.json in the archive identify the applicable file and field scopes. Neither standard licence is offered as an alternative for every element of the archive; no additional all-purpose author relicence of the original service-generated fields is granted. The PDF guide explains file links, scoring, denominators, and the evidence underlying manuscript Tables 3, 5 and 6. The archive includes the complete recorded final answers and the original descriptive results. Hidden reasoning text, credentials, live API runners and development responses are excluded. The files preserve the supporting-materials snapshot prepared on 2 November 2026. Dated preparation-status statements inside the files describe that snapshot. The status of this repository record identifies whether the deposit has been published.