From 63eba90820c18ca9731b3c7512c4b44d9c841d01 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Tue, 1 Sep 2026 17:48:19 +0200 Subject: [PATCH] =?UTF-8?q?docs(EXPERIMENTS):=20E5=20verdict=20=E2=80=94?= =?UTF-8?q?=20gate=20FAILED=20at=205.0853,=20not=20adopted?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit MTP-aux (beta 0.15, 3 steps) cost +0.26pp vs the 4.8218 control; prediction 4.5-4.75 missed. Teacher reproduces 2.289 (control intact). Confound disclosed: preemption at ~8,650 + resume from a checkpoint whose head write was interrupted — final 23% of steps trained with a re-initialized head, so the delta cannot be cleanly split between MTP-aux and the perturbation. Negative result logged; nothing shipped. --- docs/EXPERIMENTS.md | 17 ++++++++++++++++- 1 file changed, 16 insertions(+), 1 deletion(-) diff --git a/docs/EXPERIMENTS.md b/docs/EXPERIMENTS.md index c55949c..918de0f 100644 --- a/docs/EXPERIMENTS.md +++ b/docs/EXPERIMENTS.md @@ -212,7 +212,22 @@ All rows passed the CER parity gate at release. Readings: ## E5 — MTP-aux distillation rung (run-007-r7-muon-mtp) -- **Status:** REGISTERED 2026-09-01, launching. +- **Status:** COMPLETE 2026-09-01 — **GATE FAILED: 5.0853** full-set + windowed DER (gate <= 4.5218; worse than the 4.8218 control by + +0.26pp; registered prediction 4.5-4.75 missed). Teacher reproduces + 2.289 — harness control intact. **NOT ADOPTED.** The MTP-aux head + at beta 0.15 / 3 steps did not densify supervision into quality at + this rung; it cost a quarter of a point. +- **Confound, disclosed:** the run was preempted at step ~8,650 and + resumed from step-8,500 whose mtp_head.pt write was interrupted — + the final 23% of steps trained with a re-initialized aux head + (student+optimizer state intact). The clean control had no such + event. The +0.26pp cannot be cleanly attributed between MTP-aux + itself and the resume perturbation; a clean rerun is the optional + follow-up if MTP-aux is ever revisited. What is NOT confounded: + the head (0.57% params, discarded at inference) shipped nothing, + and the E6 arm is unaffected. +- Original registration follows. - **Source:** Tencent Hy4-preview carries a native multi-token prediction layer; mapped to our stack as a TRAINING auxiliary (TODO 07-hy4) — per-position multi-step heads densify supervision on the