gg2

ai-engine/models/calibration.py

@@ -149,10 +149,15 @@ class Calibrator:
                 except Exception as e:
                     print(f"[Calibrator] Warning: Failed to load metrics for {market}: {e}")
 
-    # Below this sample count, blend isotonic with raw_prob to dampen overfit jumps.
-    # Above this count, trust isotonic fully.
-    TRUSTED_SAMPLE_FLOOR = 30
-    TRUSTED_SAMPLE_CEILING = 200
+    # Below this sample count, the isotonic model is treated as untrained
+    # (raw_prob is returned). Between MIN and FLOOR we ramp from 0 to ~15%
+    # trust. Between FLOOR and CEILING we ramp to full trust.
+    # Rationale: 12-sample calibrators are statistical noise; even 30%
+    # blending on them propagates the noise into the confidence value the
+    # betting_brain reads downstream.
+    HARD_MIN_SAMPLES = 50
+    TRUSTED_SAMPLE_FLOOR = 100
+    TRUSTED_SAMPLE_CEILING = 400
     # Hard cap on how far calibration can move probability in either direction.
     MAX_DELTA = 0.20
 
@@ -198,15 +203,21 @@ class Calibrator:
                 # Sparse models barely move probability; mature models dominate.
                 metrics = self.metrics.get(market_key)
                 n_samples = metrics.sample_count if metrics else 0
+                if n_samples < self.HARD_MIN_SAMPLES:
+                    # Below 50 samples isotonic fit is unreliable — bypass it
+                    # entirely and return raw_prob. The heuristic shrinkage
+                    # below would still apply a model-version multiplier elsewhere.
+                    return float(np.clip(raw_prob, 0.01, 0.99))
                 if n_samples >= self.TRUSTED_SAMPLE_CEILING:
                     iso_weight = 1.0
                 elif n_samples <= self.TRUSTED_SAMPLE_FLOOR:
-                    # Very sparse: at least 30% trust to surface the signal
-                    iso_weight = max(0.30, n_samples / self.TRUSTED_SAMPLE_CEILING)
+                    # Linear ramp from 0% at HARD_MIN_SAMPLES to ~25% at FLOOR
+                    span = self.TRUSTED_SAMPLE_FLOOR - self.HARD_MIN_SAMPLES
+                    iso_weight = 0.25 * (n_samples - self.HARD_MIN_SAMPLES) / span
                 else:
-                    # Linearly ramp 30% → 100% between floor and ceiling
+                    # Linearly ramp 25% → 100% between floor and ceiling
                     span = self.TRUSTED_SAMPLE_CEILING - self.TRUSTED_SAMPLE_FLOOR
-                    iso_weight = 0.30 + 0.70 * (n_samples - self.TRUSTED_SAMPLE_FLOOR) / span
+                    iso_weight = 0.25 + 0.75 * (n_samples - self.TRUSTED_SAMPLE_FLOOR) / span
                 blended = iso_weight * iso_pred + (1.0 - iso_weight) * raw_prob
 
                 # Cap delta to avoid huge swings on noisy calibrators