From 8fc7df5b26888c78cba66859de6fd70fe072877f Mon Sep 17 00:00:00 2001
From: sora <2075279110@qq.com>
Date: Tue, 15 Sep 2026 03:46:14 +0000
Subject: [PATCH] Gate: continuous scaling on shared endpoints; adaptive dwell;
 MIN_OK 5

User-directed changes:
- steady no longer pins the converged level: the endpoint is shared,
  other tenants move its capacity mid-run, so steady keeps judging
  forever (+1 when rate beats reference by 5%, -1 when 15% below,
  reference drifts by EWMA). Large drops are still handled by the
  failure channel's multiplicative x0.7; the +-1 path tracks drift.
- dwell fallback scales with the OBSERVED completion cadence:
  max(120s, 3x inter-completion gap EMA). A 25s timer judged 60s-per-
  request benches on one lone sample.
- MIN_OK floor raised to 5 (evidence = max(5, 2x level)).

Simulated capacity drift 8->3->8: gate follows down then recovers.

Co-Authored-By: Claude <noreply@anthropic.com>
---
 evalharness/model/pool.py | 48 ++++++++++++++++++++++++++++-----------
 1 file changed, 35 insertions(+), 13 deletions(-)

diff --git a/evalharness/model/pool.py b/evalharness/model/pool.py
index 599633a..275e37c 100644
--- a/evalharness/model/pool.py
+++ b/evalharness/model/pool.py
@@ -162,8 +162,11 @@ class AdaptiveGate:
     # bisect narrows to the knee; steady holds there. Any failure x0.7s
     # immediately and restarts probing from the shrunken level.
     GAIN_EPS = 1.1      # rate must beat the previous level by 10% to keep doubling
-    MIN_OK = 3          # baseline completions needed at a level before judging
-    MAX_AT_LEVEL_S = 25  # ... or this many seconds, whichever comes first
+    MIN_OK = 5          # baseline completions needed at a level before judging
+    # dwell fallback: JUDGE also when the level has been held this long (with
+    # >= 1 completion) -- scaled by observed inter-completion gap so a bench
+    # whose single request takes 60s is not judged on one lone sample at t=25s
+    DWELL_BASE_S = 120.0
     # robust judging: sample count scales WITH the level (a 2-completion
     # estimate at level 8 is pure quantization noise), plus a minimum dwell
     # so one lucky tick cannot speak for the whole level
@@ -256,6 +259,13 @@ class AdaptiveGate:
         if ok:
             self._interval_ok += 1
             self._level_ok += 1
+            now = time.monotonic()
+            last = getattr(self, '_last_ok_t', None)
+            if last is not None:
+                gap = now - last
+                ge = getattr(self, '_gap_ema', None)
+                self._gap_ema = gap if ge is None else 0.6 * ge + 0.4 * gap
+            self._last_ok_t = now
         else:  # multiplicative decrease -- survival first
             self._interval_fails += 1
             # capacity moved (or we overshot): shrink now and restart the
@@ -310,12 +320,17 @@ class AdaptiveGate:
             lvl = max(1, int(self.limit))
 
             # not enough evidence yet at this level: keep measuring.
-            # need = max(MIN_OK, level): rate noise shrinks only with
+            # need = max(MIN_OK, 2x level): rate noise shrinks only with
             # samples proportional to the concurrency being judged
             need_ok = max(self.MIN_OK, lvl * 2)
-            if self._level_ok < need_ok and dt < self.MAX_AT_LEVEL_S:
+            # dwell fallback scales with the OBSERVED completion cadence:
+            # a 60s-per-request bench needs minutes, not 25s, before a
+            # single-sample judgment is acceptable
+            need_dt = max(self.DWELL_BASE_S,
+                          3.0 * (getattr(self, '_gap_ema', None) or 0.0))
+            if self._level_ok < need_ok and dt < need_dt:
                 return
-            # zero completions in MAX_AT_LEVEL_S: hang or overloaded -> hold
+            # zero completions so far: hang or overloaded -> hold
             if self._level_ok == 0:
                 return
 
@@ -356,17 +371,24 @@ class AdaptiveGate:
                 else:
                     self._enter_level('steady')   # hit HI with gains: stay
             elif self._mode == 'steady':
-                # capacity estimates are noisy: periodically re-probe upward
-                self._steady_ticks = getattr(self, '_steady_ticks', 0) + 1
-                if self._steady_ticks >= 12:   # ~60s at PROBE_S=5
-                    self._steady_ticks = 0
-                    self._mode = 'probe'
-                    self._prev = (lvl, rate)
-                    # try one level up, not a full double, from settled state
+                # the endpoint is SHARED: other tenants change its capacity
+                # while we run -- keep judging forever, nudge +-1 against the
+                # reference rate instead of pinning the converged level
+                ref = self._good_rate or rate
+                if rate >= ref * 1.05 and lvl < self.HI:
                     self.limit = float(min(lvl + 1, self.HI))
+                    self.stats['ramp_demand'] += 1
                     self._push_limit()
                     self._enter_level()
-                    return
+                elif rate <= ref * 0.85 and lvl > self.LO:
+                    self.limit = float(max(self.LO, lvl - 1))
+                    self.stats['backoff_queue'] += 1
+                    self._push_limit()
+                    self._enter_level()
+                else:
+                    # reference drifts with fresh measurements (slow EWMA)
+                    self._good_rate = 0.7 * ref + 0.3 * rate
+                    self._enter_level()   # restart the measurement window
             elif self._mode == 'bisect':
                 lo, hi = self._bis
                 if rate >= self._good_rate * 0.9: