Evidence Conflicts Before Verdicts

Across three synthetic inputs and six archived model responses, an equal-authority evidence conflict was incorrectly classified as valid under both a full task and a validity-only task. Post-hoc suppression withheld both incorrect conflict responses but reduced coverage and left other errors visible. Correct conflict rejections were absent from that archive, so their loss rate is unknown. A constructed routing demonstration exposed two correct conflict mentions alongside four false sentence units and one unsupported claim in a blocked review channel. A pre-sprint certificate check supported three known open judgments and rejected three defective controls. A separately authorized 23 October offline restart evaluated six synthetic views prepared for an earlier technically aborted attempt. Four source-bound, manually AI-typed in-scope mappings yielded 16 reference-matching labels, including three open judgments with compatible opposite-valued witnesses. Two further views were stopped by manual scope decisions. A separate 24 October language-transfer exercise used six new prose cases, prepared on 23 October, in the same two contexts. Four in-scope mappings produced 16 reference-matching labels and five open axes with opposing compatible witnesses; two manual scope stops were supported. Neither exercise establishes general transfer, automatic extraction, model improvement, or collusion detection. There is no independent holdout or human reference validation. Source authority and completeness are stipulated. The original abort remains documented. Earlier diagnostics and checks are pre-sprint work; both transfer exercises are dated sprint-period work. Sanitized L01-L06 reading evidence is now available as L1; the V01-V06 restart records remain local.