Fix step reward flow, wire investigation API, and refresh README scores

README.md

@@ -237,10 +237,10 @@ python inference.py --env-url http://localhost:8000
 
 | Task | Difficulty | F1 Score | Precision | Recall |
 |------|-----------|----------|-----------|--------|
-| Expense Audit | Easy | 0.4285 | 0.31 | 0.68 |
-| Invoice Match | Medium | 0.2667 | 0.17 | 0.67 |
-| GST Reconciliation | Hard | 0.1212 | 0.07 | 0.33 |
-| Fraud Detection | Expert | 0.0384 | 0.02 | 0.14 |
+| Expense Audit | Easy | 0.1200 | 0.07 | 0.43 |
+| Invoice Match | Medium | 0.1800 | 0.11 | 0.44 |
+| GST Reconciliation | Hard | 0.0100 | 0.01 | 0.01 |
+| Fraud Detection | Expert | 0.1100 | 0.11 | 0.10 |
 
 **Why did it fail?**
 The model frequently drops to 0.00 on the tasks because it struggles with abstract rules (like date math for weekend expenses, or tracking cumulative limits). It actively gets tricked by "red herrings"—perfectly legal expenses that it hallucinates as errors—which entirely destroys its precision score.

financial_audit_env/server/app.py

@@ -128,6 +128,7 @@ class StepRequest(BaseModel):
     """Request body for the /step endpoint."""
     action: AuditAction
     session_id: Optional[str] = None
+    request_categories: Optional[List[str]] = None
 
 
 class BaselineResponse(BaseModel):
@@ -240,7 +241,10 @@ async def step_endpoint(request: StepRequest):
     """
     try:
         env = _get_env(request.session_id)
-        obs = env.step(request.action)
+        obs = env.step(
+            request.action,
+            request_categories=request.request_categories or [],
+        )
         _metrics["total_steps"] += 1
         if obs.done:
             _metrics["total_episodes_completed"] += 1
@@ -304,24 +308,18 @@ async def get_grader_score(session_id: Optional[str] = None):
             "message": "No episode completed. Call /reset then /step with submit_final=True.",
         }
 
-    def final_clamp(val: float) -> float:
-        """Keep score-like fields within a stable open interval."""
-        return max(0.01, min(0.99, val))
-
     return {
         "status": "completed",
         "task_id": env.state.task_id,
-        # Primary score fields with a final stability clamp.
-        "score": final_clamp(result["score"]),
-        "precision": final_clamp(result["precision"]),
-        "recall": final_clamp(result["recall"]),
+        "score": result["score"],
+        "precision": result["precision"],
+        "recall": result["recall"],
         "true_positives": result["true_positives"],
         "false_positives": result["false_positives"],
         "false_negatives": result["false_negatives"],
         "total_errors": result["total_errors"],
-        # Enhanced scoring fields with the same stability clamp.
-        "weighted_score": final_clamp(result.get("weighted_score", result["score"])),
-        "partial_credit_score": final_clamp(result.get("partial_credit_score", result["score"])),
+        "weighted_score": result.get("weighted_score", result["score"]),
+        "partial_credit_score": result.get("partial_credit_score", result["score"]),
         "partial_matches": result.get("partial_matches", 0),
         # Confusion matrix
         "confusion_matrix": result.get("confusion_matrix", {}),

financial_audit_env/server/environment.py

@@ -234,14 +234,15 @@ class FinancialAuditEnvironment(Environment):
         # Check if episode should end
         is_final = action.submit_final or step_num >= self._task["max_steps"]
 
-        # Compute step reward.
-        # Keep cumulative episode rewards aligned with the final task score.
-        if not is_final:
-            step_reward = 0.0
-        else:
-            final_grader = compute_f1_score(self._findings, self._ground_truth)
-            # Use the bounded final score as the terminal reward.
-            step_reward = max(0.01, min(0.99, final_grader["score"]))
+        # Compute dense per-step reward using only new findings plus episode context.
+        step_reward = compute_step_reward(
+            new_findings=new_finding_dicts,
+            all_findings_so_far=self._findings,
+            ground_truth=self._ground_truth,
+            step_number=step_num,
+            is_final=is_final,
+            max_steps=self._task["max_steps"],
+        )
         
         self._episode_reward += step_reward
 

test_http.py

@@ -17,7 +17,7 @@ print(f"GET /health → {r.status_code} {r.json()}")
 r = requests.get(f"{BASE}/tasks")
 assert r.status_code == 200
 data = r.json()
-assert data["total_tasks"] == 3
+assert data["total_tasks"] == 4
 print(f"GET /tasks → {r.status_code}, {data['total_tasks']} tasks")
 for t in data["tasks"]:
     print(f"  - {t['id']}: {t['name']} ({t['difficulty']})")
@@ -56,8 +56,8 @@ assert r.status_code == 200
 grader = r.json()
 print(f"GET /grader → score={grader['score']}, P={grader['precision']}, R={grader['recall']}")
 
-# 6. Test all 3 tasks work
-for task_id in ["expense_audit", "invoice_match", "gst_reconciliation"]:
+# 6. Test all 4 tasks work
+for task_id in ["expense_audit", "invoice_match", "gst_reconciliation", "fraud_detection"]:
     r = requests.post(f"{BASE}/reset", json={"task_id": task_id, "seed": 42})
     assert r.status_code == 200
     print(f"POST /reset({task_id}) → {r.status_code} ✓")