Create consensus/metrics.py

consensus/metrics.py

@@ -0,0 +1,23 @@
+import numpy as np
+
+def cosine_similarity(a: np.ndarray, b: np.ndarray) -> float:
+    a = a / (np.linalg.norm(a) + 1e-12)
+    b = b / (np.linalg.norm(b) + 1e-12)
+    return float(a @ b)
+
+def trinary_quantize(x: np.ndarray, threshold: float) -> np.ndarray:
+    """
+    Map to {-1, 0, +1}. Symmetric threshold around 0.
+    """
+    out = np.zeros_like(x, dtype=np.int8)
+    out[x >  threshold] =  1
+    out[x < -threshold] = -1
+    return out
+
+def weighted_average_text(responses: list[str], weights: np.ndarray) -> str:
+    """
+    Simple consensus surface: pick the response with max weight (argmax).
+    Placeholder; could switch to RLAIF-style merge, majority vote over facts, etc.
+    """
+    idx = int(np.argmax(weights))
+    return responses[idx]