results: glm-4.7-flash on SadeedDiac-25 — 13.0035 raw / 13.2256 zero-skip; axis complete

[ronaldtse]

glm-4.7-flash completes the SadeedDiac-25 frontier regression axis:

• 13.0035 raw / 13.2256 zero-skip DER (w/o-CE 10.0510/10.3206; NFDW 17.60) — 1,200/1,200 paragraphs, thinking-disabled (the last plain-completion GLM row), zero empties after 12 resumable passes past sustained provider 429s; the fix(eval): retry empty checkpoint rows instead of resuming them as done #65 guard dropped sentinels between passes (249 -> ... -> 0).
• Attribution (rates over GT-marked positions; scorer validated by reproducing the recorded 5.2 wrong 2.64% / 5.3 missing 3.75 wrong 8.59 exactly): missing 6.67% — worst of the family — wrong 9.01%, extra 0.96%; U+0670 convention effect 0.04pp.
• Paired bootstrap vs GLM-5.2 (paragraph-level simplified scorer, 10k resamples, seed 42): +7.945pp, 95% CI [+7.516, +8.370], p<1e-4.
• Axis: GLM-5.2 2.5060 -> GLM-5.3-Flash 8.5721 -> GLM-5.3 9.9760 -> glm-4.7-flash 13.0035 raw. Every measurable GLM successor lost classical-Arabic mark knowledge its predecessor had; 4.7-flash regresses on both missing and wrong.

RESULTS.md section + protocol-ledger row + bootstrap row; results dir with preds CSVs; register statuses. Paper row follows in interscript-ml; site note already amended (interscript.github.io #149).