beta_servo: 贴上界伺服(亮区反演cap*=servo·cap_rho/ρ̂双向一步落点,暗区复用ride爬升;常数仅0.8安全系数) — 开火证书:plain s195000死亡区2.5k步CERT_PASS(β双向移动25/26行,best 3.5703→3.4260,零崩溃;贴边开销+50%墙钟为最恶劣工况); fw72m_servo王冠从头跑已发射(servo+ride4+cap_floor0+wsync3)

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

1 files changed, 12 insertions, 1 deletions

diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py
index b3239f1..3000461 100644
--- a/ep_run/casc_eq_train.py
+++ b/ep_run/casc_eq_train.py
@@ -68,6 +68,13 @@ 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('--beta_servo', type=float, default=0.0)  # >0: CEILING-HUGGING SERVO. Meter lit (res above the
+                                                          # v2 absolute gate): deadbeat inversion onto the ceiling
+                                                          # — cap *= servo*cap_rho/rho^ (rho ~= G*beta near the
+                                                          # edge, so one step lands beta at servo*ceiling; grows
+                                                          # toward it when under, shrinks when over). Meter dark:
+                                                          # existing ride climb probes upward. Value = safety
+                                                          # fraction of the ceiling to sit at (canonical 0.8).
 ap.add_argument('--wsync', type=int, default=0)           # >0: SYNCHRONOUS WEIGHT-STEP ACCEPTANCE — snapshot
                                                           # params+momentum before each opt.step; next step's
                                                           # nudged relax measures the new state through the SAME
@@ -577,7 +584,11 @@ def ep_step(x, y):
         if args.ride_ema > 0:  # ride-v2(b), opt-in: smooth the meter before decisions
             rho_use = GOV['rho_ema'] = 0.9 * GOV.get('rho_ema', rho_g) + 0.1 * rho_g
         GOV['cool'] = max(GOV.get('cool', 0) - 1, 0)
-        if res_g > 0.02 and rho_use > args.beta_cap_rho:
+        if args.beta_servo > 0 and res_g > 0.02 and rho_use > 0:
+            # meter lit -> INVERT onto the ceiling (both directions), replacing the AIMD attack
+            GOV['cap'] = min(max(GOV.get('cap', 1.0) * (args.beta_servo * args.beta_cap_rho / rho_use),
+                                 args.cap_floor), args.beta_ride)
+        elif res_g > 0.02 and rho_use > args.beta_cap_rho:
             GOV['cap'] = max(GOV.get('cap', 1.0) * 0.85, args.cap_floor)  # attack (gentler than v1)
             if args.ride_cool > 0: GOV['cool'] = args.ride_cool  # ride-v2(c), opt-in
         elif (res_g < 0.01 or rho_use < 0.5 * args.beta_cap_rho) and GOV['cool'] == 0: