docs: E6 verdict — gate FAILED at 5.8057; register-swap negative

docs/EXPERIMENTS.md

@@ -271,6 +271,20 @@ All rows passed the CER parity gate at release. Readings:
 - **Prediction (registered):** 4.45-4.75 — a register-diversification
   effect at constant budget should be real but second-order; if the
   domain attribution is right, this should beat E5's MTP-aux.
+- **Status:** COMPLETE 2026-09-02 — **GATE FAILED: 5.8057** full-set
+  windowed DER (gate <=4.5218; control 4.8218; E5's failed 5.0853).
+  NOT ADOPTED. Teacher reproduces at 2.289 (n=1200, protocol
+  consistent). Paired bootstrap student-teacher: +3.2388pp, CI
+  [2.939, 3.575]. The prediction (4.45-4.75) missed badly. Read:
+  swapping news-domain units OUT for classical-register units at
+  constant budget HURT (−0.98pp vs control) — the news-domain
+  teacher labels match the benchmark surface better than classical
+  units do, and 30k total was already enough budget that register
+  mix was not the binding constraint. Together with G2a (4.5701 at
+  6 epochs) the ladder now reads: longer training and MORE total
+  units move the rung; register swaps at constant budget do not.
+  Clean negative; direction shifts to G2b (48k add). Labels sha256
+  39b4213b5ab4c9f7273da90a9a1954f7701320b62870a954e6e2c02e810617a4.
 
 ## Parked
 

docs/PUBLICATION-NOTES.md

@@ -70,7 +70,19 @@ Paper: frontier section (fourth question). RESULTS.md run-003-pkm.
 RL teacher polishing flat/negative ×3; microkimi bridges improve
 structure but not accuracy; teacher beam-search unnecessary for
 Arabic; per-channel int8 rejected on measurement; the 30 MiB tier
-closed as infeasible without pretraining.
+closed as infeasible without pretraining; **E5 MTP-aux (2026-09-01):
+5.0853 vs the 4.8218 control — multi-token-prediction as a training
+auxiliary HURT at this scale (+0.26pp), with a disclosed preemption
+confound (fresh aux head for the final 23% of steps); E6
+constant-budget register swap (2026-09-02): 5.8057 — replacing news
+units with classical Tashkeela at constant total HURT (−0.98pp vs
+control), the domain-shaped-residual hypothesis's causal test
+failing in the swap direction; the add direction (G2b) remains
+open.** The E5/E6 pair is the paper's data-vs-architecture exhibit:
+two levers from the frontier-LLM literature (MTP, register
+diversification at constant budget) both regressed on byte-level
+student distillation — the levers that moved the rung were optimizer
+(E3), fresher teacher labels (E4), and epochs (G2a).
 
 ### 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
@@ -217,8 +229,11 @@ The decomposition for paper B, every line full-set with brackets:
 | depth halved (lite, 6ep) | 5.784 | delta 3.25 [3.03, 3.49] |
 
 Paper-B framing: levers compose roughly additively (optimizer >>
-teacher > epochs), the remaining ~2.0-2.2pp is domain-shaped (the
-label-scale rung is the causal test), and capacity appears ONLY as
+teacher > epochs), and the E6 causal test came back NEGATIVE —
+swapping 8k news units for classical-register Tashkeela at constant
+30k total scored 5.8057 (−0.98pp vs control), so the residual is
+not fixed by register mix at constant budget; the add direction
+(G2b, 48k total) is the live test, and capacity appears ONLY as
 depth — width is load-bearing (both stitch ratios collapsed) while
 depth trades 1.21pp for 37% of the artifact. The subset-overstatement
 phenomenon (five instances, up to 3.2x inflation) is the