From e4f35a2d8b24c21db7733396639036072e09f512 Mon Sep 17 00:00:00 2001
From: KoalaTrapLord <265133236+KoalaTrapLord@users.noreply.github.com>
Date: Fri, 3 Apr 2026 16:08:50 -0400
Subject: [PATCH] debug: log top-5 action probabilities on every decide() call

Adds a single INFO log line after each successful inference showing:
  - chosen action + confidence
  - top-5 actions sorted by probability
  - turn, phase, vector_source

This lets us diagnose the 'only plays lands' behaviour without a Python
REPL. Remove or downgrade to DEBUG once the training issue is resolved.
---
 routes/policy.py | 18 ++++++++++++++++++
 1 file changed, 18 insertions(+)

diff --git a/routes/policy.py b/routes/policy.py
index 32e2d65..2af9f3d 100644
--- a/routes/policy.py
+++ b/routes/policy.py
@@ -342,6 +342,24 @@ async def decide(req: DecideRequest):
     if "error" in result:
         raise HTTPException(500, detail=result["error"])
 
+    # ----------------------------------------------------------------
+    # Probability diagnostic — logs top-5 actions on every decision.
+    # Remove or downgrade to logger.debug once training looks healthy.
+    # ----------------------------------------------------------------
+    probs = result.get("probabilities", {})
+    if probs:
+        top5 = sorted(probs.items(), key=lambda kv: kv[1], reverse=True)[:5]
+        top5_str = ", ".join(f"{a}:{p:.3f}" for a, p in top5)
+        logger.info(
+            "[decide] turn=%d phase=%s → %s (conf=%.3f) | top5=[%s] | src=%s",
+            req.resolved_turn(),
+            req.phase,
+            result["action"],
+            result["confidence"],
+            top5_str,
+            vector_source,
+        )
+
     log_prob = 0.0
     try:
         conf = result.get("confidence", 0.5)