From bd3bc03cece613a69ad0e0a2577d895035e68849 Mon Sep 17 00:00:00 2001 From: Ronald Tse Date: Thu, 3 Sep 2026 08:40:16 +0200 Subject: [PATCH] =?UTF-8?q?docs(EXPERIMENTS):=20GKD=20rung=20registered/ar?= =?UTF-8?q?med=20=E2=80=94=20gate=20+=20prediction=20before=20any=20launch?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The last untested reading of the domain attribution; spec design, gate (4.5218), and prediction (4.30-4.65) pre-registered. Launch is the owner call. Also records the run-006 re-eval discrepancy (5.0821 vs canonical 4.8218) so the next verdict compares within one eval pass. --- docs/EXPERIMENTS.md | 25 +++++++++++++++++++++++++ 1 file changed, 25 insertions(+) diff --git a/docs/EXPERIMENTS.md b/docs/EXPERIMENTS.md index 4435f3b..f3f7f79 100644 --- a/docs/EXPERIMENTS.md +++ b/docs/EXPERIMENTS.md @@ -297,6 +297,31 @@ All rows passed the CER parity gate at release. Readings: shows a convergence-speed effect distinguishable from noise on the surviving curves; the DER gates remain the only verdicts. +## GKD — on-policy distillation rung (REGISTERED/ARMED 2026-09-03) + +- **Status:** REGISTERED, launch pending owner ordering (the last + lever with no data; TODO.substantiate/03). Implementation follows + this registration verbatim when launched. +- **Hypothesis:** the client tier's remaining ~2.0pp over the teacher + is domain-shaped exposure: the student never trains on its own + decode distribution. On-policy GKD (sequences sampled from the + student during training, scored against the frozen r7 teacher's + targets/logits) attacks exactly that. +- **Design:** spec `ara-diac-small-2-gkd`, control = run-006 verbatim + (30k units, identical steps/schedule/optimizer); delta = GKD loss + mixing — student-sampled sequences (temperature-matched to greedy + inference) blended into the sequence-KD objective at a fixed ratio, + annealed to zero over the final third. +- **Pre-agreed gate:** adopt at <= 4.5218 full-set windowed DER; + honest-report band [4.5218, 4.8218); investigate if worse. +- **Prediction (registered):** 4.30-4.65 — the E6 swap-negative says + register mix at constant budget is not the lever; on-policy + exposure is the remaining untested reading of the domain + attribution. Note: run-006's canonical control is 4.8218 (E4-era + eval); a later auto-chain re-eval printed 5.0821 for the same run — + the ledger's 4.8218 stays canonical; a GKD verdict compares within + one eval pass. + ## Parked - **Speculative decoding** (LongCat converts sparsity→speed): revisit