Files
MeltMind/docs/s1-real-scoring-delivery-2026-11-14.md
T

2.3 KiB
Raw Blame History

S1 real scoring delivery — submitted for Claude result review

Fixed result commit: 90b8c66f3c21dda5e144e888d535a9b09b66f460.

All 24 sequential folds PASS. No training. Actual 785,808 logical 8NN queries plus 785,808 radius searches, matching every planned fold. Total whole-process fold time 270.5283 s; longest 13.2551 s; peak RSS 146.0078 MiB. No 1800 s / 1.25 GiB limit was enlarged. Polling target 20 ms, observed maximum gap 0.586656 s; full traces/HWM/final peaks retained.

Primary five-seed-mean result: B−(A+S) ranks behind, C−ordinary-edge-shuffle ranks ahead, in both whole and near-wall slices. Each has 5/5 matching point directions and 5/5 matching interval classifications. These opposite primary directions do not satisfy the proposed joint organized-input familiarity interpretation. The normalized secondary edge contrast is negative but much smaller (about −0.00055). All point values, paired query intervals and fixed-LOPO source-reweighting intervals are published; no new-source coverage guarantee or prediction-performance claim.

ZIP: 212,704,694 bytes, SHA256 82dc70af62ad1b4f7ac42b71cc105bf32f89597c5d5311e67359dd8ad6266355. All 286 archived evidence files verified after fresh extraction, and 10 replayed tables/audits are byte-identical. Local receipt: artifacts/s1-real-scoring-2026-11-14/fresh_archive_replay.json.

Remote verification: all 37 non-ZIP files fully read back. ZIP verified by immutable Git blob identity, byte length and first 65,536 bytes; full remote ZIP download/rehash was not performed. This exact boundary was disclosed in Slack. Receipt: artifacts/s1-real-scoring-2026-11-14/publication_receipt.json.

Claude received the full result review request. Grok's independent seed/formula check was already complete and no duplicate full review was requested. This handoff does not claim Claude has accepted the numerical results. No further computation is authorized merely by this delivery.