feat(scoring): hard-zero credibility above the timeout limit

allways/constants.py

@@ -63,6 +63,10 @@
 VOLUME_WEIGHT_ALPHA: float = 0.5
 # Closed swaps required for full credibility (0 → 100% linear ramp).
 CREDIBILITY_RAMP_OBSERVATIONS: int = 10
+# More than this many timed-out swaps within CREDIBILITY_WINDOW_BLOCKS hard-zeros
+# a miner's credibility (and thus their whole reward) until the old timeouts age
+# out of the rolling window. 0-2 tolerated; the 3rd timeout zeros credibility.
+CREDIBILITY_MAX_TIMEOUTS: int = 2
 
 # ─── Emission Recycling ────────────────────────────────────
 RECYCLE_UID = 53  # Subnet owner UID

allways/validator/scoring.py

@@ -17,6 +17,7 @@
 
 from allways.chains import canonical_pair
 from allways.constants import (
+    CREDIBILITY_MAX_TIMEOUTS,
     CREDIBILITY_RAMP_OBSERVATIONS,
     CREDIBILITY_WINDOW_BLOCKS,
     DIRECTION_POOLS,
@@ -249,6 +250,7 @@ def calculate_miner_rewards(self: Validator) -> Tuple[np.ndarray, Set[int]]:
             wt.record_credibility(
                 closed_swaps=sum(success_stats.get(hotkey, (0, 0))),
                 ramp_target=CREDIBILITY_RAMP_OBSERVATIONS,
+                timed_out=success_stats.get(hotkey, (0, 0))[1],
             )
             crown_share_dir = blocks / total_crown_dir
             vol_dir = volumes_dir.get(hotkey, 0)
@@ -384,9 +386,14 @@ def success_rate(stats: Optional[Tuple[int, int]]) -> float:
 
 def credibility_ramp(stats: Optional[Tuple[int, int]]) -> float:
     """Linear ramp to full credibility at CREDIBILITY_RAMP_OBSERVATIONS closed
-    swaps. Applied linearly to the reward, not cubed. Zero observations → 0."""
+    swaps. Applied linearly to the reward, not cubed. Zero observations → 0.
+    More than CREDIBILITY_MAX_TIMEOUTS timed-out swaps in the window hard-zeros
+    credibility — a rolling penalty that recovers as old timeouts age out."""
     if not stats or stats == (0, 0):
         return 0.0
+    _completed, timed_out = stats
+    if timed_out > CREDIBILITY_MAX_TIMEOUTS:
+        return 0.0
     return min(1.0, sum(stats) / CREDIBILITY_RAMP_OBSERVATIONS)
 
 

allways/validator/scoring_trace.py

@@ -10,7 +10,12 @@
 import bittensor as bt
 import numpy as np
 
-from allways.constants import CREDIBILITY_RAMP_OBSERVATIONS, RECYCLE_UID, TAO_TO_RAO
+from allways.constants import (
+    CREDIBILITY_MAX_TIMEOUTS,
+    CREDIBILITY_RAMP_OBSERVATIONS,
+    RECYCLE_UID,
+    TAO_TO_RAO,
+)
 
 if TYPE_CHECKING:
     from allways.validator.scoring import DirectionTrace
@@ -45,9 +50,12 @@ def record_volume(self, vol_rao: int, total_volume_rao: int, crown_share: float,
         self.participation = min(1.0, self.volume_share / crown_share) if crown_share > 0 else 1.0
         self.volume_factor = factor
 
-    def record_credibility(self, closed_swaps: int, ramp_target: int) -> None:
+    def record_credibility(self, closed_swaps: int, ramp_target: int, timed_out: int = 0) -> None:
         self.closed_swaps = closed_swaps
-        self.credibility_ramp = min(1.0, closed_swaps / ramp_target) if ramp_target > 0 else 1.0
+        if timed_out > CREDIBILITY_MAX_TIMEOUTS:
+            self.credibility_ramp = 0.0
+        else:
+            self.credibility_ramp = min(1.0, closed_swaps / ramp_target) if ramp_target > 0 else 1.0
 
 
 def log_scoring_trace(

tests/test_scoring_v1.py

@@ -158,14 +158,20 @@ def test_none_is_zero(self):
         assert credibility_ramp((0, 0)) == 0.0
 
     def test_scales_linearly_below_threshold(self):
-        """1 closed swap → ramp 0.1; 5 closed → 0.5. Counts timeouts too."""
+        """1 closed swap → ramp 0.1; 5 closed → 0.5. Counts tolerated timeouts too."""
         assert credibility_ramp((1, 0)) == 0.1
         assert credibility_ramp((5, 0)) == 0.5
-        assert credibility_ramp((2, 3)) == 0.5
+        assert credibility_ramp((3, 2)) == 0.5  # 2 timeouts tolerated
 
     def test_caps_at_full_ramp(self):
         assert credibility_ramp((10, 0)) == 1.0
-        assert credibility_ramp((90, 10)) == 1.0
+
+    def test_excess_timeouts_hard_zero(self):
+        """More than CREDIBILITY_MAX_TIMEOUTS (2) timeouts → 0, regardless of volume."""
+        assert credibility_ramp((8, 2)) == 1.0  # 2 tolerated, fully ramped
+        assert credibility_ramp((8, 3)) == 0.0  # 3rd timeout zeros it
+        assert credibility_ramp((90, 10)) == 0.0  # high volume can't rescue it
+        assert credibility_ramp((0, 3)) == 0.0
 
 
 class TestCrownHoldersHelper:
@@ -1693,6 +1699,15 @@ def test_record_credibility_zero_target_is_unity(self):
         wt.record_credibility(closed_swaps=5, ramp_target=0)
         assert wt.credibility_ramp == 1.0
 
+    def test_record_credibility_excess_timeouts_zero(self):
+        """Trace mirrors the reward rule: >3 timeouts → ramp shown as 0."""
+        from allways.validator.scoring_trace import WeightingTrace
+
+        wt = WeightingTrace()
+        wt.record_credibility(closed_swaps=100, ramp_target=10, timed_out=4)
+        assert wt.closed_swaps == 100
+        assert wt.credibility_ramp == 0.0
+
 
 class TestVolumeWeighting:
     """End-to-end volume weighting via calculate_miner_rewards."""
@@ -2036,8 +2051,27 @@ def test_full_ramp_at_threshold(self, tmp_path: Path):
         np.testing.assert_allclose(rewards[0], POOL_BTC_TAO, atol=1e-6)
         v.state_store.close()
 
-    def test_timeouts_advance_ramp_but_hurt_raw_rate(self, tmp_path: Path):
-        """5 completed + 5 timed_out → ramp = 1.0, raw = 0.5 → 0.5³ × 1.0 = 0.125."""
+    def test_tolerated_timeouts_advance_ramp_but_hurt_raw_rate(self, tmp_path: Path):
+        """8 completed + 2 timed_out (within the 2-timeout tolerance) → ramp = 1.0,
+        raw = 0.8 → 0.8³ × 1.0."""
+        hotkeys = pad_hotkeys_to_cover_recycle(['hk_a'])
+        v = make_validator(tmp_path, hotkeys, baseline_credibility=False)
+        self.seed_btc_tao_crown(v, 'hk_a')
+        for i in range(8):
+            v.state_store.insert_swap_outcome(
+                swap_id=i + 1, miner_hotkey='hk_a', completed=True, resolved_block=100 + i
+            )
+        for i in range(2):
+            v.state_store.insert_swap_outcome(
+                swap_id=100 + i, miner_hotkey='hk_a', completed=False, resolved_block=200 + i
+            )
+        rewards, _ = calculate_miner_rewards(v)
+        np.testing.assert_allclose(rewards[0], POOL_BTC_TAO * 0.8**3, atol=1e-6)
+        v.state_store.close()
+
+    def test_excess_timeouts_zero_reward(self, tmp_path: Path):
+        """5 completed + 5 timed_out → 5 timeouts > 3 → credibility hard-zeroed →
+        reward 0 despite a high raw fulfillment count."""
         hotkeys = pad_hotkeys_to_cover_recycle(['hk_a'])
         v = make_validator(tmp_path, hotkeys, baseline_credibility=False)
         self.seed_btc_tao_crown(v, 'hk_a')
@@ -2050,7 +2084,7 @@ def test_timeouts_advance_ramp_but_hurt_raw_rate(self, tmp_path: Path):
                 swap_id=100 + i, miner_hotkey='hk_a', completed=False, resolved_block=200 + i
             )
         rewards, _ = calculate_miner_rewards(v)
-        np.testing.assert_allclose(rewards[0], POOL_BTC_TAO * 0.5**3, atol=1e-6)
+        np.testing.assert_allclose(rewards[0], 0.0, atol=1e-6)
         v.state_store.close()
 
     def test_unearned_ramp_portion_recycles(self, tmp_path: Path):