guards_silent(用户令07-20): 守卫只报不拦(skip全部放行提交,唯一例外=非有限梯度仍跳),证书流程废除;唯一执法=控制器自身折半;判官=用户盯CE; simplev2王冠直发链排plain2后(--guards_silent --kretry 0,无cap/wsync)

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

1 files changed, 14 insertions, 4 deletions

diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py
index b378c07..c77b92c 100644
--- a/ep_run/casc_eq_train.py
+++ b/ep_run/casc_eq_train.py
@@ -68,6 +68,11 @@ ap.add_argument('--drift_adapt', type=float, default=0.0)  # >0: adaptive drift
 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
+ap.add_argument('--guards_silent', action='store_true')   # USER ORDER 07-20: guards REPORT but never
+                                                          # block — no step skips, no kretry escalation
+                                                          # needed, gn/drift pass-through. The only acting
+                                                          # mechanism is the beta controller's own halving
+                                                          # (that's its measurement, not a guard). Judge = CE.
 ap.add_argument('--beta_simple', type=float, default=0.0) # >1: the user-spec adaptive beta setter, NOTHING else:
                                                           # clean step -> beta *= this (e.g. 1.01, next step);
                                                           # illegal step -> the beta_sync halvings PERSIST.
@@ -590,6 +595,9 @@ def ep_step(x, y):
                 # CODEX FIX: K=8 rescue ran at beta/2^beta_sync — persist those halvings too,
                 # else the next step jumps straight back to the failed beta (upward bias)
                 GOV['bsimp'] = GOV.get('bsimp', 1.0) * 0.5 ** args.beta_sync
+        if not ok_retry and args.guards_silent:
+            GOV['skd'] = GOV.get('skd', 0) + 1   # counted for telemetry, but the step COMMITS
+            ok_retry = True
         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
@@ -669,8 +677,9 @@ def ep_step(x, y):
         gdrift2 = ddp_max_scalar(drift2)
         if (not math.isfinite(gdrift2)) or (gdrift2 > 0.5 and not args.noguard):
             GOV['skd'] = GOV.get('skd', 0) + 1   # second-pass drift reject -> skip step (synced)
-            for p in all_params: p.grad = None
-            return free_ce, beta_t, GOV.get('kuse', GOV['K']), False
+            if not args.guards_silent:
+                for p in all_params: p.grad = None
+                return free_ce, beta_t, GOV.get('kuse', GOV['K']), False
         E2 = 0.0
         for z, o in zip(zpb, lob): E2 = E2 + 0.5 * ((z.detach().float() - o.float()) ** 2).sum()
         gsE2 = torch.autograd.grad(E2 / (NBT * b2), all_params, allow_unused=True)
@@ -691,8 +700,9 @@ def ep_step(x, y):
     GOV['gn'] = gn
     if not math.isfinite(gn) or (gn > 8 * GOV['gema'] and not args.noguard):
         GOV['skg'] = GOV.get('skg', 0) + 1   # gn-EMA-guard reject (gradient-magnitude spike)
-        for p in all_params: p.grad = None
-        return free_ce, beta_t, GOV.get('kuse', GOV['K']), False
+        if not (args.guards_silent and math.isfinite(gn)):
+            for p in all_params: p.grad = None
+            return free_ce, beta_t, GOV.get('kuse', GOV['K']), False
     for p, g in zip(all_params, gs):
         p.grad = g
     if args.beta_simple > 1.0 and _ok0: