--beta_ride ceiling-riding governor (two-sided bcap: climb in calm, attack near threshold); smoke: 3e-3 -> 2.4e-2 hover at ride cap, rho calm at 42M tail

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

1 files changed, 6 insertions, 1 deletions

diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py
index 8e246d1..6d94ef5 100644
--- a/ep_run/casc_eq_train.py
+++ b/ep_run/casc_eq_train.py
@@ -51,6 +51,9 @@ ap.add_argument('--data', default='tinystories_bpe')     # dataset dir under ep_
 ap.add_argument('--sync_check', type=int, default=500)   # DDP: verify bitwise param sync every N steps (0=off)
 ap.add_argument('--ddp_backend', default='nccl', choices=['nccl', 'gloo'])  # gloo = correctness tests on shared GPUs
 ap.add_argument('--ddp_grad_test', action='store_true')  # one-step grad equivalence test vs single-GPU big batch, then exit
+ap.add_argument('--beta_ride', type=float, default=1.0)   # cap ceiling: >1 lets the governor RAISE beta
+                                                          # above the schedule, up to ride x schedule
+ap.add_argument('--beta_ride_up', type=float, default=1.02)  # per-step climb rate in the calm branch
 ap.add_argument('--beta_cap_rho', type=float, default=0.0)  # >0: LOOP-GAIN CAP on beta — if per-sweep residual
                                                           # ratio rho^ exceeds this, bscale *= 0.8 (beta backs off
                                                           # under the wall-2 ceiling); recovers x1.02 when rho^ low
@@ -476,7 +479,9 @@ def ep_step(x, y):
         if res_g > 0.02 and rho_g > args.beta_cap_rho:
             GOV['cap'] = max(GOV.get('cap', 1.0) * 0.85, 0.05)  # attack (gentler than v1)
         elif res_g < 0.01 or rho_g < 0.5 * args.beta_cap_rho:
-            GOV['cap'] = min(GOV.get('cap', 1.0) * 1.02, 1.0)   # recover
+            # recover; with beta_ride > 1 the governor CLIMBS past the schedule — beta finds
+            # its own ceiling and hovers there (ride-the-ceiling; 1.0 = legacy defensive cap)
+            GOV['cap'] = min(GOV.get('cap', 1.0) * args.beta_ride_up, args.beta_ride)
     if CF:
         # one-graph centered: [g(+b)+g(-b)]/2 = d[(E+ - E-)/(2b·NBT)]/dtheta; CE-head term from
         # the +beta half only (matches sequential centered's gsC at the +beta top states).