docs: GKD verdict — 6.0036 GATE FAILED; the lever ladder closes

On-policy reverse-KL is the worst rung measured, +1.18pp over the 4.8218 control; bootstrap +3.4083 CI 3.109-3.743; teacher reproduces 2.289. Registered prediction 4.30-4.65 missed. With MTP, register swap and add, the frontier-training toolkit is 0-for-4 on byte-level student distillation; optimizer, fresher labels, epochs remain the only positives. run-012 absorbed nine preemptions via checkpoint-resume. PUBLICATION-NOTES: lever table row, negatives exhibit extended to the four-way set, framing line closed.

Author: Ronald Tse

docs/EXPERIMENTS.md

@@ -318,30 +318,34 @@ All rows passed the CER parity gate at release. Readings:
   (−0.47pp), and epochs (−0.25pp). What remains is GKD (on-policy
   exposure, run-012 in flight) — after which the ladder is closed.
 
-## GKD — on-policy distillation rung (REGISTERED/ARMED 2026-09-03)
+## GKD — on-policy distillation rung (COMPLETE 2026-09-05)
 
-- **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.
+- **Status:** COMPLETE — **GATE FAILED: 6.0036** full-set windowed DER
+  (n=1200; teacher reproduces 2.289; paired bootstrap student−teacher
+  +3.4083, CI [3.109, 3.743]). NOT ADOPTED. The registered prediction
+  (4.30-4.65) missed badly; honest-report band also breached — this
+  is the worst rung measured, +1.18pp over the 4.8218 control. Run
+  run-012-r7-muon-gkd: 10,995 steps, final CE 0.0076, nine server
+  preemptions absorbed by checkpoint-resume (no measured work lost);
+  labels sha256 e70ce991d15a8c810b83e2b5401f1410293844c623ffefa646b
+  19fb94a6180df.
+- **Hypothesis (registered):** 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 attacks exactly that.
+- **Design (as run):** control run-006 verbatim (30k units, identical
+  steps/schedule/optimizer); delta = reverse-KL on student-sampled
+  sequences (every 4th step, sub-batch 2, temp 1.0, cap 1024) scored
+  by the frozen r7 teacher, ratio 0.3 annealed to zero over the final
+  third (ml #148; run id renamed run-012 in #151).
+- **Read:** the domain-exposure reading of the residual joins the
+  register readings as a THIRD negative (E6 swap 5.8057, G2b add
+  4.8231, GKD on-policy 6.0036) — the program's final datum. The
+  ladder closes: optimizer (−2.96pp), fresher teacher labels
+  (−0.47pp), and epochs (−0.25pp) are the only measured positives;
+  MTP-aux, register diversification (both directions), on-policy
+  distillation, and depth-cut all regressed or went flat. At SFT
+  convergence on clean supervision, byte-level student distillation
+  tolerates nothing but optimization and supervision quality.
 
 ## Parked
 

docs/PUBLICATION-NOTES.md

@@ -81,13 +81,19 @@ failing in the swap direction; G2b label-scale ADD (2026-09-04,
 verdict recorded per TODO.substantiate 04): 4.8231 — 48k total
 (+18k classical) at 6 epochs is FLAT vs the 3ep control and 0.25pp
 worse than news-only G2a (4.5701): the add neutralized the epoch
-gain, closing the domain hypothesis negative in BOTH directions.**
-The E5/E6/G2b set is the paper's data-vs-architecture exhibit:
-two levers from the frontier-LLM literature (MTP, register
-diversification — swap or add) plus the label-scale hypothesis all
-regressed or flat on byte-level student distillation — the levers
-that moved the rung were optimizer
-(E3), fresher teacher labels (E4), and epochs (G2a).
+gain, closing the domain hypothesis negative in BOTH directions; GKD
+on-policy distillation (2026-09-05, the program's final rung,
+run-012): 6.0036 — reverse-KL on student-sampled sequences scored by
+the frozen teacher is the WORST rung measured (+1.18pp over control),
+the domain-exposure reading of the residual joining the register
+readings as the third negative.**
+The E5/E6/G2b/GKD set is the paper's data-vs-architecture exhibit:
+three levers from the frontier-LLM literature (MTP, register
+diversification — swap or add, on-policy GKD) plus the label-scale
+hypothesis all regressed or flat on byte-level student distillation —
+the levers that moved the rung were optimizer
+(E3), fresher teacher labels (E4), and epochs (G2a), and nothing
+else. The program's measured ladder is closed.
 
 ### 9. Muon optimizer A/B on the memory student (E3) — LANDED 2026-08-28
 **4.8287 vs 7.5553 full-set (−2.727pp from the optimizer alone); adopt
@@ -234,6 +240,7 @@ The decomposition for paper B, every line full-set with brackets:
 | register swap (E6, 3ep) | 5.8057 | negative |
 | register add (G2b, 6ep) | 4.8231 | delta 2.37 [2.19, 2.55] |
 | MTP-aux (E5, 3ep) | 5.0853 | negative (confound disclosed) |
+| on-policy GKD (run-012) | 6.0036 | delta 3.41 [3.11, 3.74] — worst rung |
 | depth halved (lite, 6ep) | 5.784 | delta 3.25 [3.03, 3.49] |
 
 Paper-B framing: levers compose roughly additively (optimizer >>