--beta_sync: synchronous acceptance (this-step relax telemetry gates the commit; reject -> halve beta, retry same batch; failed trial doubles as ceiling measurement for the ride cap)

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

1 files changed, 32 insertions, 3 deletions

diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py
index 10e3515..f115a91 100644
--- a/ep_run/casc_eq_train.py
+++ b/ep_run/casc_eq_train.py
@@ -54,6 +54,9 @@ ap.add_argument('--ddp_grad_test', action='store_true')  # one-step grad equival
 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_sync', type=int, default=0)       # >0: SYNCHRONOUS acceptance — this step's own
+                                                          # relax telemetry gates the commit; on reject,
+                                                          # halve beta and retry same batch (N halvings max)
 ap.add_argument('--ride_ema', type=int, default=0)        # ride-v2(b): rho-EMA before governor decisions
 ap.add_argument('--ride_cool', type=int, default=0)       # ride-v2(c): post-attack climb cooldown (steps)
 ap.add_argument('--drift_adapt', type=float, default=0.0)  # >0: adaptive drift ceiling = this x trailing
@@ -474,9 +477,35 @@ def ep_step(x, y):
         de = GOV.get('drift_ema')
         if de is not None:
             dthr = min(0.5, max(0.05, args.drift_adapt * de))
-    if (not math.isfinite(gdrift)) or (gdrift > dthr and not args.noguard):
+    def _legal(gd):
+        # SYNCHRONOUS acceptance (--beta_sync): judge THIS step by THIS step's own relax
+        # telemetry — an out-of-window beta can be attempted but can never COMMIT.
+        if (not math.isfinite(gd)) or (gd > dthr and not args.noguard): return False
+        if args.beta_sync > 0:
+            res_s = ddp_bcast_scalar(GOV.get('res', 0.0))
+            rho_s = ddp_bcast_scalar(GOV.get('rho') or 0.0)
+            if res_s > 0.02 and rho_s > args.beta_cap_rho: return False
+        return True
+    if not _legal(gdrift):
         ok_retry = False
-        if args.kretry > 0 and math.isfinite(gdrift) and not args.noguard:
+        if args.beta_sync > 0 and not args.noguard:
+            for _h in range(args.beta_sync):     # halve beta, retry the SAME batch
+                beta_t = beta_t * 0.5
+                GOV['skr'] = GOV.get('skr', 0) + 1
+                z0, zs, ins, outs = free_states_graphed(x_in)
+                zs_free = [z.clone() for z in zs]
+                zp, last_outs = relax(z0, zs, ins, outs, y_in, +beta_t, GOV['K'], x_in, bmask=bmask)
+                with torch.no_grad():
+                    drift = _drift(zp, zs_free)
+                gdrift = ddp_max_scalar(drift)
+                if _legal(gdrift):
+                    ok_retry = True
+                    if args.beta_ride > 1.0:     # the failed trial IS the ceiling measurement
+                        GOV['cap'] = max(GOV.get('cap', 1.0) * 0.5 ** (_h + 1), 0.05)
+                    break
+        if ok_retry:
+            pass
+        if not ok_retry and args.kretry > 0 and math.isfinite(gdrift) and not args.noguard:
             GOV['skr'] = GOV.get('skr', 0) + 1   # marginal batch: retry once with deeper relaxation
             z0, zs, ins, outs = free_states_graphed(x_in)
             zs_free = [z.clone() for z in zs]
@@ -484,7 +513,7 @@ def ep_step(x, y):
             with torch.no_grad():
                 drift = _drift(zp, zs_free)
             gdrift = ddp_max_scalar(drift)
-            ok_retry = math.isfinite(gdrift) and gdrift <= 0.5
+            ok_retry = _legal(gdrift)
         if not ok_retry:
             GOV['skd'] = GOV.get('skd', 0) + 1   # drift-guard reject (relaxation non-convergence)
             for p in all_params: p.grad = None