ara-diac-small: full-set correction — 8.259 DER-CE; subset 3.658 unrepresentative

[ronaldtse]

Summary

• Full-set SadeedDiac-25 measurement of ara-diac-small-1.0: 8.259 DER-CE on all 1,200 paragraphs vs 3.658 on the first-300 subset. Teacher reproduces 2.5815 against its documented 2.5793, confirming protocol consistency. The subset was not representative — it sits in the student's training-domain neighborhood; the remaining 900 paragraphs expose a domain-generalization gap
• Catalog number is now the full-set figure (+5.68pp against the strict gate, disclosed); the leaderboard reading becomes "behind Sadeed-1.5B" (the earlier between-Gemini-Flash-and-GPT-4 placement was a subset artifact, withdrawn)
• Standing rule recorded: student-tier numbers publish from full benchmark sets only; subset figures labeled at first publication
• Paper integrity fix: frontier finding 3 now matches the results log — the Arabic 33M collapse ran on corrupted labels (retracted there), so the clean-label pretrained-or-collapse evidence is Thai; abstract/contribution/scope statements aligned
• Metrics chain: one mapping entry with four table specs (subset + full-set) for fp32, plus a new entry covering the int8 variant's metadata; imf metrics passes for all 10 models

Test plan

• imf cli metrics — all models trace to RESULTS.md (validated against the pushed branch ref; flip ref to main post-merge)
• asciidoctor render clean, no stale 3.66/2.34 references outside the disclosed-context sentences
• teacher reproduction within 0.003pp rules out a harness change