From 941d16750c0b3893a3437f017ad89214dd74fbf9 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Mon, 7 Sep 2026 03:00:25 +0200 Subject: [PATCH] =?UTF-8?q?site:=20finish=20the=202.0=20correction=20?= =?UTF-8?q?=E2=80=94=20negatives=20note=20and=20blog=20table?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Follows 6df21ff (the ladder rung correction): the negatives note keeps each delta as measured, annotated against the corrected 5.08 control, withdraws the dead 'cancelled the epoch gain' reading, and states the withdrawal itself. The blog frontier table's 2.0 row moves to 5.08 with the correction date; its CI matches the re-eval interval [2.36, 2.82]. --- src/content/blog/2026-09-05-frontier-closed.adoc | 5 +++-- src/pages/ml.astro | 16 +++++++++++----- 2 files changed, 14 insertions(+), 7 deletions(-) diff --git a/src/content/blog/2026-09-05-frontier-closed.adoc b/src/content/blog/2026-09-05-frontier-closed.adoc index e1d9589..504fb44 100644 --- a/src/content/blog/2026-09-05-frontier-closed.adoc +++ b/src/content/blog/2026-09-05-frontier-closed.adoc @@ -25,7 +25,7 @@ bootstrap interval on the gap to the teacher: |from-scratch 30M, every lever |30M |73.95 |[70.22, 71.09] |1.0 rung |300M |8.26 |[4.6, 5.7] |lite rung (enc 12→6) |190M |5.78 |[3.03, 3.49] -|2.0 rung |300M |4.82 |[2.36, 2.82] +|2.0 rung |300M |5.08 (corrected 2026-09-05) |[2.36, 2.82] |2.1 rung |300M |4.57 |[1.91, 2.35] |teacher r7 |580M |2.29 |— |=== @@ -52,7 +52,8 @@ 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 (−0.98pp). Adding five times the +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 diff --git a/src/pages/ml.astro b/src/pages/ml.astro index 2848145..ba12f87 100644 --- a/src/pages/ml.astro +++ b/src/pages/ml.astro @@ -338,11 +338,17 @@ const clientModels = [

What didn't move it — every rung pre-registered, measured, and kept in the log: - multi-token-prediction auxiliary 5.09 (+0.26pp); classical-register swap at constant - budget 5.81 (+0.98pp); register add at matched epochs 4.82 (flat — it cancelled the epoch - gain); on-policy GKD 6.00 (+1.18pp, the worst rung); depth-halved 5.78 (63% of the - parameters — shipped anyway as the browser tier). At SFT convergence, supervision quality - dominates. The ladder is closed. + multi-token-prediction auxiliary 5.09 (+0.26pp as measured vs the + then-published control — level with the corrected 5.08); classical-register swap at + constant budget 5.81 (+0.98pp as measured; +0.72pp over the corrected control); register + add at matched epochs 4.82 (statistically flat against news-only 4.57, intervals + overlapping — it sits between the corrected control and the news-only rung); on-policy + GKD 6.00 (+1.18pp as measured; +1.43pp over the rung it was meant to improve — the worst + measured); depth-halved 5.78 (63% of the parameters — shipped anyway as the browser + tier). At SFT convergence, supervision quality dominates. The ladder is closed. The + control's originally published 4.8218 did not reproduce and was withdrawn on + 2026-09-05; two independent re-measurements, one at the artifact level on the exact + released bytes, agree on 5.08.

Hebrew, same discipline: 16.43% DER on Biblical Hebrew, where the modern-Hebrew state of