Executive summary
Five live claims receive direct native evidence: the shared-prototype mechanism and Val5 best-case gain reproduce; the 0.7472 macro and S=2 ablation peak fail to reproduce under the printed algorithm; and the exact prototype-sharing work is 6.08% of per-worker confusion-matrix state. Ten public datasets match all registered scale fields exactly. Paired warning-strict runs and a packaged replay are byte-identical; no peer artifact, source-only abstention, or substitute benchmark is counted.
Primary records: OpenReview and arXiv.
<iframe src="../../poster_embed.html" style="width:100%;height:430px;border:0" title="PTBCC reproduction poster"></iframe>