Expose TOTEM model response diagnostics

app.py

@@ -3122,6 +3122,44 @@ def _apply_llm_dashboard_state(
     return render_dashboard(state), log_df, _score_summary(log_df), gate_summary
 
 
+def _bridge_verification_notes(debug: dict, llm_state: dict) -> dict:
+    actual = debug.get("actual_tokens") if isinstance(debug.get("actual_tokens"), dict) else {}
+    usage_reported = actual.get("total_tokens") not in (None, "", 0)
+    warnings: list[str] = []
+    if debug.get("provider") == "fake":
+        warnings.append("Fake audit provider returned these numbers; this should never be enabled on Hugging Face Spaces.")
+    if not usage_reported:
+        warnings.append("Provider did not report token usage, so API completion cannot be cost-audited from usage fields.")
+    if not str(llm_state.get("evidence_summary") or "").strip():
+        warnings.append("Model returned no evidence_summary; score grounding is weak.")
+
+    scores = llm_state.get("scores") if isinstance(llm_state.get("scores"), dict) else {}
+    rounded_scores = [score for score in scores.values() if isinstance(score, int) and score % 5 == 0]
+    if scores and len(rounded_scores) == len(scores):
+        warnings.append("All dashboard scores are multiples of 5; this may indicate coarse model judgement rather than workbook-derived precision.")
+
+    queue = llm_state.get("revision_priority_queue") or []
+    generic_blocks = []
+    if isinstance(queue, list):
+        for item in queue:
+            if not isinstance(item, dict):
+                continue
+            block = str(item.get("block") or "").strip().lower()
+            if block in {"page 1", "page 2", "page 3", "page 4", "page 5", "opening", "middle", "ending"}:
+                generic_blocks.append(item.get("block"))
+    if generic_blocks:
+        warnings.append(f"Revision queue uses generic block labels: {generic_blocks}. Ask for stronger manuscript evidence.")
+
+    return {
+        "api_call_verified": bool(debug.get("provider") and debug.get("provider") != "fake" and debug.get("schema_pass")),
+        "usage_reported": usage_reported,
+        "raw_response_chars": debug.get("raw_response_chars"),
+        "token_audit_transaction_id": debug.get("token_audit_transaction_id"),
+        "token_audit_path": debug.get("token_audit_path"),
+        "warnings": warnings,
+    }
+
+
 def analyze_manuscript_and_refresh(
     file_obj,
     author_name: str,
@@ -3234,6 +3272,9 @@ def run_totem_analysis_from_context(active_path: str, manuscript_context_json: s
             "actual_tokens": debug.get("actual_tokens"),
             "token_audit_path": debug.get("token_audit_path"),
             "token_audit_transaction_id": debug.get("token_audit_transaction_id"),
+            "verification": _bridge_verification_notes(debug, llm_state),
+            "validated_payload": debug.get("validated_payload") or llm_state,
+            "raw_model_payload": debug.get("raw_model_payload"),
         }
         return dashboard_html, log_df, summary, json.dumps(debug_log, indent=2)
     except Exception as exc:

src/totem_bridge.py

@@ -223,6 +223,8 @@ Use the workbook matrix as scoring authority and manuscript text as evidence.
             "schema_pass": True,
             "token_audit_path": audit_path,
             "token_audit_transaction_id": transaction_id,
+            "raw_model_payload": raw_payload,
+            "validated_payload": validated,
         }
 
     provider_config = _select_llm_provider()
@@ -350,6 +352,8 @@ Use the workbook matrix as scoring authority and manuscript text as evidence.
         "actual_tokens": actual_tokens,
         "token_audit_path": audit_path,
         "token_audit_transaction_id": transaction_id,
+        "raw_model_payload": parsed,
+        "validated_payload": validated,
     }
     return validated, debug