feat: log judge, target, and auditor model metadata in saved JSONs

simpleaudit/experiment.py

@@ -370,7 +370,7 @@ async def _run_single_rep(
             else:
                 fresh = {r.scenario_name: r for r in partial}
                 merged_results = [fresh.get(r.scenario_name, r) for r in result]
-                result = AuditResults(merged_results)
+                result = AuditResults(merged_results, meta=partial.meta)
                 result.timestamp = partial.timestamp
             errored = [r.scenario_name for r in result if r.severity == "ERROR"]
             if not errored:
@@ -524,6 +524,7 @@ async def run_streamed(
                 "judge_model": self.judge_model,
                 "judge_base_url": self.judge_base_url,
                 "judge_provider": self.judge_provider,
+                "judge_name": self.judge if isinstance(self.judge, str) else None,
             }.items() if v is not None
         } or None
 
@@ -696,6 +697,7 @@ async def run_async(
                 "judge_model": self.judge_model,
                 "judge_base_url": self.judge_base_url,
                 "judge_provider": self.judge_provider,
+                "judge_name": self.judge if isinstance(self.judge, str) else None,
             }.items() if v is not None
         } or None
 

simpleaudit/model_auditor.py

@@ -1683,7 +1683,30 @@ async def _run_one(scenario: Dict) -> AuditResult:
                     if bar.total is not None and bar.n < bar.total:
                         bar.update(bar.total - bar.n)
 
-        return AuditResults(results)
+        return AuditResults(results, meta=self._run_meta(scenarios, max_turns, language))
+
+    def _run_meta(
+        self,
+        scenarios: Union[str, List[Dict]],
+        max_turns: Optional[int] = None,
+        language: str = "English",
+    ) -> Dict[str, Any]:
+        """Build the run-level metadata dict stored in the saved JSON."""
+        meta: Dict[str, Any] = {
+            "target_model": self.target_model,
+            "target_provider": self._target_client_config["provider"],
+            "judge_model": self.judge_model,
+            "judge_provider": self._judge_client_config["provider"],
+            "auditor_model": self.auditor_model,
+            "auditor_provider": self._auditor_client_config["provider"],
+            "max_turns": max_turns or self.max_turns,
+            "language": language,
+        }
+        if self.judge_name is not None:
+            meta["judge_name"] = self.judge_name
+        if isinstance(scenarios, str):
+            meta["scenario_pack"] = scenarios
+        return meta
 
     def run(
         self,

simpleaudit/reframing.py

@@ -1215,7 +1215,7 @@ async def grade(result: AuditResult):
                 target_output_tokens=result.target_output_tokens,
             )
         )
-    return AuditResults(rejudged)
+    return AuditResults(rejudged, meta=source.meta)
 
 
 def rejudge(

simpleaudit/results.py

@@ -86,9 +86,12 @@ class AuditResults:
         "pass": "🟢",
     }
 
-    def __init__(self, results: List[AuditResult]):
+    def __init__(self, results: List[AuditResult], meta: Optional[Dict] = None):
         self.results = results
         self.timestamp = datetime.now().isoformat()
+        # Run-level metadata (judge/target/auditor models, max_turns, etc.).
+        # Populated by ModelAuditor.run_async / SingleTurnAuditor.run_async.
+        self.meta: Dict = meta or {}
 
     def __len__(self) -> int:
         return len(self.results)
@@ -273,7 +276,7 @@ def summary(self):
         print()
 
     def to_dict(self) -> Dict:
-        return {
+        d: Dict = {
             "version": "1.0",
             "timestamp": self.timestamp,
             "summary": {
@@ -291,6 +294,9 @@ def to_dict(self) -> Dict:
             "recommendations": self.all_recommendations,
             "results": [r.to_dict() for r in self.results],
         }
+        if self.meta:
+            d["meta"] = self.meta
+        return d
 
     def save(self, filepath: str):
         """Save results to JSON file (atomically, so interrupts can't corrupt it)."""
@@ -309,7 +315,7 @@ def load(cls, filepath: str) -> "AuditResults":
             AuditResult(**r) for r in data["results"]
         ]
 
-        instance = cls(results)
+        instance = cls(results, meta=data.get("meta"))
         instance.timestamp = data.get("timestamp", instance.timestamp)
         return instance
 

simpleaudit/single_turn.py

@@ -657,4 +657,4 @@ async def _run_one(scenario: Dict) -> AuditResult:
                     await asyncio.gather(*tasks, return_exceptions=True)
                     raise
 
-        return AuditResults(results)
+        return AuditResults(results, meta=self._run_meta(scenarios, max_turns, language))