diff options
Diffstat (limited to 'protocol/README.md')
| -rw-r--r-- | protocol/README.md | 5 |
1 files changed, 3 insertions, 2 deletions
diff --git a/protocol/README.md b/protocol/README.md index 770d582..f275883 100644 --- a/protocol/README.md +++ b/protocol/README.md @@ -133,8 +133,9 @@ The defaults match the paper: healthy networks (~1e-5). - `h_norm_explosion_ratio = 50.0` — BP-trained networks are ~1-3× per layer in our measurements; failure modes show ~10⁵-10⁶× from layer 0 to layer L. -- `stability_drift_ceiling = 0.30` — calibrated for ~128-sample batches. - BP/EP-trained networks: 0.05-0.18. DFA/SB drift mode: 0.43-0.99. +- `stability_drift_ceiling = 0.30` — calibrated for ~128-sample batches at K=10. + BP/EP-trained networks (6 audited values): [-0.04, +0.12]. DFA/SB/CB drift + mode (9 audited values): up to +0.99, with 5/9 above the 0.30 cutoff. - `frozen_acc_margin_pp = 2.0` — DFA-on-ResMLP undercuts frozen baseline by 4 pp; the threshold is loose enough to allow noise on healthy networks. |