The Arabic frontier, closed: every rung measured, two verdicts negative, one 95 MB browser model

Two weeks ago the phonological layer shipped with a disclosed gap: the Arabic client model scored 8.26 on the full benchmark. This week every rung was measured on the full test set, with confidence intervals on each separation, and the two remaining hypotheses were both tested. One test came back negative, and that result is published here with the rest.

The Arabic size–quality frontier, every rung full-set with a paired bootstrap interval on the gap to the teacher:

Adjacent intervals are disjoint end to end: the frontier’s separations are statistically real. The new releases — ara-diac-small-2.1 and the layerdrop trio — resolve from index-v3 in every runtime.

The lite rung is the headline for the browser: encoder depth halved by verbatim layer copy, quantized to 4 bits, ~95 MB — and the quantization is nearly free by measurement: 0.036% of argmax decisions flip against fp32, every one of them at near-tie margins. The live demo leads with it: download once, verify the checksum, run offline, watch the decoder stream with per-step confidence.

The residual gap between the 2.1 student and its teacher had one live attribution left: classical-domain coverage. Both tests of it failed. Swapping news-domain training units for classical Tashkeela at constant budget made the student worse (+0.72pp over the corrected 2.0 control). Adding five times the classical corpus — the maximal version of the lever — left it statistically flat (4.82 vs 4.57, intervals overlapping). The residual is not a coverage deficit the corpus can reach; it lives in the teacher–student interaction itself.

The second negative traveled further. The depth cut that costs 1.21pp on Arabic was replicated single-variable on the Hebrew lineage — and collapsed: 77.48 DER against the full-depth 30.38. Depth-compressibility is a property of the recipe, not of the pretrained backbone. Both findings ship in the results log with their intervals; a claim scoped by its counterexample beats a generalization nobody tested.

Every leaderboard number we publish can now be re-derived by anyone:

The tool reproduces our published verdicts exactly — it re-derives this week’s 4.8231 run from its raw predictions and the public benchmark, intervals included. The comparison is a command, not a claim.

Twenty-two models across four languages and three tiers, every artifact checksummed and index-pinned, every metric traceable to a logged harness, and now a protocol tool that puts the scoring in your hands instead of ours.