- Status
- Experimental research
- Published
- 2026-11-17
- Modified
- 2026-11-17
- Category
- confidence
Included changes
- 0111 November · The r4 audit candidate addresses assessment reuse, SDK compatibility, durable failure reviews, confirmation binding and backup consistency. All 29 findings have recorded dispositions. Local validation: 1,624 tests passed and 14 skipped; six GitHub checks passed before merge.
- 0216 November · Fourteen missing Google passages were caused by joined paragraph boundaries in table cells. The corrected candidate matches all 1,060 archived reference passages. Fifty-eight focused tests and an independent Python replay passed with network access denied.
- 0316 November · All 120 originally unsuccessful or held scopes were classified. Fixes recover thirteen cases from identical captures, including ten false CAPTCHA detections, with zero losses among the earlier 380 extraction successes. All 222 focused tests passed offline; responsibility remains undetermined for 46 transport outcomes.
- 0416 November · The frozen 500-scope catalogue produced 380 extractions in the original local trial. The fixed candidate produced 397/500 locally in 459.555 seconds and 367/500 in the Hostinger SSH account in 772.602 seconds. The target diagnostic ran outside the served website; its 133 unsuccessful or held scopes remain recorded.
- 058–9 November · Two free-API connectivity probes passed. A separate semantic trial received six analyst outputs but no usable verifier verdict across eleven attempts. The isolated worker remained queued at verification. These failures remain part of the research record.