Upload zimg-fdloss/RECIPE.md with huggingface_hub

zimg-fdloss/RECIPE.md

@@ -51,13 +51,24 @@ holdout = sorted-enumeration idx % 12 == 0.
 | b16 recipe (old champion) | 4 | 0.899 | 0.842 | 1.612 | 1.519 |
 | + batch 64 | 4 | 0.910 | 0.853 | 1.621 | 1.447 |
 | **+ 250-step finishing (CHAMPION)** | **4** | **0.910** | **0.858** | — | — |
-| 1-NFE specialist (TFC full-run, b16) | 2 | 0.843 | 0.879 | 1.909 | 1.596 |
+| **1-NFE frontier: b16 + dose-controlled TFDH-adv** | **4** | **0.908** | **0.892** | — | — |
+| (superseded) 1-NFE specialist, TFC full-run b16 | 2 | 0.843 | 0.879 | 1.909 | 1.596 |
 
 McNemar: champion vs b16 recipe +2.91 @8 / +2.42 @1 (4 seeds, pre-registered
-promotion rule); b16 recipe vs turbo z=+7.0/+8.2 (Holm-robust). No teacher,
-no discriminator, no fake-score network, no CFG at train or inference;
+promotion rule); b16 recipe vs turbo z=+7.0/+8.2 (Holm-robust). The champion
+path uses no teacher, no discriminator, no fake-score network, no CFG;
 ~4.5 h on one 8×H200 node (b16) / ~1.5 h wall on 4 nodes (b64 + finishing).
 
+**1-NFE frontier variant (2026-08-05)**: champion-protocol FD + a hinge
+discriminator on frozen iMF-teacher SigLIP2 tokens, dose-held at
+gnorm-ratio 0.2 by a median-of-3 closed-loop controller (+ approximated-R1
+and head-EMA stabilizers; equilibrium weight drifts 25× over 2000 steps —
+no static weight works). 4-seed: @1 z=+6.42 over b16 / +3.99 over the
+champion at @8 parity (z=−0.08); strictly dominates the old TFC 1-NFE
+specialist (z=+6.00 @8 / +1.51 @1). Did not displace the champion under
+the pre-registered both-endpoint rule (@8 tie ≠ gain). Pick it when
+1-step latency is the deployment target.
+
 ## The three laws (measured)
 
 1. **Division of labor**: mid-training must not spend mode coverage
@@ -79,6 +90,7 @@ no discriminator, no fake-score network, no CFG at train or inference;
 ## Checkpoints in this folder
 
 - `champion-I1combo/` — **the promoted champion** (b64 FD + 250s finishing)
+- `frontier-TFDHA2-1nfe/` — **1-NFE frontier** (b16 + dose-controlled TFDH-adv)
 - `champion-D1c-clean/` — the b16 recipe (FD ckpt + iMF init)
 - `champion-D1c-b64/` — batch-64 stage (the promoted champion's base)
 - `T1-1nfe-record/` — best pure-1-NFE model (0.879@1)