DCCL Commit 2: Evaluation logic (structured, LLM, safety override)

fdidonato · frdidonato · commit 6a76a1f73632 · 2026-05-20T19:37:41.000+02:00
fixed pre-commit run -a issues on mypy
diff --git a/moralstack/compliance/config.py b/moralstack/compliance/config.py
@@ -64,7 +64,10 @@ def _parse_int(raw: str, default: int, var_name: str, min_value: int | None = No
     if min_value is not None and value < min_value:
         _LOG.warning(
             "value for %s (%d) below minimum %d, using default %d",
-            var_name, value, min_value, default,
+            var_name,
+            value,
+            min_value,
+            default,
         )
         return default
     return value
@@ -83,7 +86,11 @@ def _parse_float(raw: str, default: float, var_name: str, min_value: float, max_
     if not (min_value <= value <= max_value):
         _LOG.warning(
             "value for %s (%f) out of range [%f, %f], using default %f",
-            var_name, value, min_value, max_value, default,
+            var_name,
+            value,
+            min_value,
+            max_value,
+            default,
         )
         return default
     return value
@@ -103,7 +110,9 @@ def get_dccl_evaluation_path() -> EvaluationPathLiteral:
         return raw  # type: ignore[return-value]
     _LOG.warning(
         "invalid value for %s: %r, using default %s",
-        _ENV_EVALUATION_PATH, raw, _DEFAULT_EVALUATION_PATH,
+        _ENV_EVALUATION_PATH,
+        raw,
+        _DEFAULT_EVALUATION_PATH,
     )
     return _DEFAULT_EVALUATION_PATH  # type: ignore[return-value]
 
diff --git a/moralstack/compliance/dccl.py b/moralstack/compliance/dccl.py
@@ -190,9 +190,7 @@ def evaluate(
                     verdict = ComplianceVerdict(
                         decision=ComplianceDecision.NO_MATCH,
                         confidence=verdict_llm.confidence,
-                        rationale=(
-                            f"Hybrid: structured found no rule; LLM also found none. {verdict_llm.rationale}"
-                        ),
+                        rationale=(f"Hybrid: structured found no rule; LLM also found none. {verdict_llm.rationale}"),
                         evaluation_path=EvaluationPath.HYBRID,
                         contract_hash=contract_hash,
                     )
diff --git a/moralstack/compliance/safety_override.py b/moralstack/compliance/safety_override.py
@@ -198,7 +198,7 @@ def _llm_classify(
         category = parsed.get("category")
         if category is None or category == "null":
             return None
-        if category in SAFETY_OVERRIDE_CATEGORIES:
+        if isinstance(category, str) and category in SAFETY_OVERRIDE_CATEGORIES:
             _LOG.debug("safety classifier LLM matched: %s", category)
             return category
         _LOG.warning("safety LLM returned unknown category: %r", category)
diff --git a/moralstack/compliance/types.py b/moralstack/compliance/types.py
@@ -12,11 +12,11 @@
 from dataclasses import dataclass
 from enum import Enum
 
-
 # =============================================================================
 # Enums
 # =============================================================================
 
+
 class ComplianceDecision(str, Enum):
     """The four possible verdicts of a DCCL evaluation."""
 
@@ -79,6 +79,7 @@ class ActionType(str, Enum):
 # Structured rule (deployer-declarable)
 # =============================================================================
 
+
 @dataclass(frozen=True)
 class StructuredRule:
     """
@@ -118,19 +119,16 @@ def __post_init__(self) -> None:
         if not self.trigger_pattern:
             raise ValueError("StructuredRule.trigger_pattern must be non-empty")
         if self.action_type != ActionType.REDIRECT and not self.action_payload:
-            raise ValueError(
-                f"StructuredRule.action_payload must be non-empty for action_type={self.action_type.value}"
-            )
+            raise ValueError(f"StructuredRule.action_payload must be non-empty for action_type={self.action_type.value}")
         if not (0 <= self.priority <= 100):
-            raise ValueError(
-                f"StructuredRule.priority must be in [0, 100], got {self.priority}"
-            )
+            raise ValueError(f"StructuredRule.priority must be in [0, 100], got {self.priority}")
 
 
 # =============================================================================
 # Matched rule (verdict-side, populated when DCCL returns MATCH)
 # =============================================================================
 
+
 @dataclass(frozen=True)
 class MatchedRule:
     """
@@ -156,6 +154,7 @@ class MatchedRule:
 # Compliance verdict (output of DCCL.evaluate)
 # =============================================================================
 
+
 @dataclass(frozen=True)
 class ComplianceVerdict:
     """
@@ -199,6 +198,7 @@ def is_safety_override(self) -> bool:
 # Compliance signal (propagated to downstream modules via request context)
 # =============================================================================
 
+
 @dataclass(frozen=True)
 class ComplianceSignal:
     """
diff --git a/moralstack/orchestration/controller.py b/moralstack/orchestration/controller.py
@@ -1040,9 +1040,7 @@ def _run_dccl_evaluation(
                     "confidence": compliance_verdict.confidence,
                     "evaluation_path": compliance_verdict.evaluation_path.value,
                     "speculative_draft_validated": compliance_verdict.speculative_draft_validated,
-                    "rationale_excerpt": (
-                        compliance_verdict.rationale[:300] if compliance_verdict.rationale else ""
-                    ),
+                    "rationale_excerpt": (compliance_verdict.rationale[:300] if compliance_verdict.rationale else ""),
                     "contract_hash": compliance_verdict.contract_hash,
                 },
             )
diff --git a/moralstack/orchestration/deliberation_runner.py b/moralstack/orchestration/deliberation_runner.py
@@ -35,8 +35,8 @@
 from moralstack.orchestration.orchestration_event_taxonomy import (
     AGGREGATED_GUIDANCE_EVALUATED,
     CONVERGENCE_EVALUATED,
-    CRITIC_SKIPPED,
     CRITIC_SHORT_CIRCUIT_TRIGGERED,
+    CRITIC_SKIPPED,
     EARLY_CONVERGENCE_ACCEPTED,
     EARLY_CONVERGENCE_REJECTED,
     PARALLEL_STRATEGY_SELECTED,
@@ -2837,16 +2837,12 @@ def _critique(
                     "started_at": int(start * 1000),
                     "duration_ms": elapsed,
                     "prompt": (
-                        f"[SKIPPED] {skip_reason}" if is_skipped
-                        else (getattr(critique, "prompt", None) or prompt_text)
+                        f"[SKIPPED] {skip_reason}" if is_skipped else (getattr(critique, "prompt", None) or prompt_text)
                     ),
                     "system_prompt": getattr(critique, "system_prompt", ""),
                     "raw_response": getattr(critique, "raw_response", "") or "",
                     "parsed_json": None,
-                    "parsed_summary_json": (
-                        f"SKIPPED: {skip_reason}" if is_skipped
-                        else response_text
-                    ),
+                    "parsed_summary_json": (f"SKIPPED: {skip_reason}" if is_skipped else response_text),
                     "attempts": getattr(critique, "parse_attempts", 1),
                     "sequence_in_cycle": SEQ_CRITIC,
                     "token_usage_json": _token_usage_json_from_result(critique),
diff --git a/moralstack/orchestration/orchestration_event_taxonomy.py b/moralstack/orchestration/orchestration_event_taxonomy.py
@@ -69,37 +69,47 @@
 # Commit 1 (Foundation) declares the constants; emission ships in Commit 2-3.
 
 COMPLIANCE_LAYER_STARTED = "COMPLIANCE_LAYER_STARTED"
-"""DCCL.evaluate() invoked. Payload: has_contract, has_structured_rules, evaluation_path_preference."""
+"""DCCL.evaluate() invoked.
+Payload: has_contract, has_structured_rules, evaluation_path_preference."""
 
 COMPLIANCE_LAYER_VERDICT_MATCH = "COMPLIANCE_LAYER_VERDICT_MATCH"
-"""DCCL returned MATCH. Payload: matched_rule_id, evaluation_path, confidence, duration_ms."""
+"""DCCL returned MATCH.
+Payload: matched_rule_id, evaluation_path, confidence, duration_ms."""
 
 COMPLIANCE_LAYER_VERDICT_NO_MATCH = "COMPLIANCE_LAYER_VERDICT_NO_MATCH"
-"""DCCL returned NO_MATCH. Payload: rationale_excerpt, evaluation_path, duration_ms, confidence."""
+"""DCCL returned NO_MATCH.
+Payload: rationale_excerpt, evaluation_path, duration_ms, confidence."""
 
 COMPLIANCE_LAYER_VERDICT_SAFETY_OVERRIDE = "COMPLIANCE_LAYER_VERDICT_SAFETY_OVERRIDE"
-"""DCCL returned SAFETY_OVERRIDE. Payload: safety_override_reason, rule_excerpt, duration_ms."""
+"""DCCL returned SAFETY_OVERRIDE.
+Payload: safety_override_reason, rule_excerpt, duration_ms."""
 
 COMPLIANCE_LAYER_VERDICT_NO_CONTRACT = "COMPLIANCE_LAYER_VERDICT_NO_CONTRACT"
-"""Request has no developer contract. Payload: empty."""
+"""Request has no developer contract.
+Payload: empty."""
 
 CONTRACT_RULE_REJECTED = "CONTRACT_RULE_REJECTED"
-"""A structured rule failed safety validation at contract loading. Payload: rule_id, reason, rejected_action_payload_excerpt."""
+"""A structured rule failed safety validation at contract loading.
+Payload: rule_id, reason, rejected_action_payload_excerpt."""
 
 CONTRACT_RULES_LOADED = "CONTRACT_RULES_LOADED"
-"""Contract loading complete. Payload: contract_hash, total_rules, accepted_rules, rejected_rules."""
+"""Contract loading complete.
+Payload: contract_hash, total_rules, accepted_rules, rejected_rules."""
 
 MODULE_DEFERRED_TO_COMPLIANCE = "MODULE_DEFERRED_TO_COMPLIANCE"
-"""A downstream module returned early due to ComplianceSignal(MATCH). Payload: module, reason, cycle, deferred_outcome_summary."""
+"""A downstream module returned early due to ComplianceSignal(MATCH).
+Payload: module, reason, cycle, deferred_outcome_summary."""
 
 CONTRACT_INJECTION_DETECTED = "CONTRACT_INJECTION_DETECTED"
-"""LLM detected a deployer-side prompt injection in the contract. Payload: detection_rationale, contract_hash."""
+"""LLM detected a deployer-side prompt injection in the contract.
+Payload: detection_rationale, contract_hash."""
 
 COMPLIANCE_LAYER_TIMEOUT = "COMPLIANCE_LAYER_TIMEOUT"
 """DCCL LLM call exceeded timeout. Payload: timeout_ms, evaluation_path."""
 
 CONTRACT_STRUCTURE_PROSE_CONFLICT = "CONTRACT_STRUCTURE_PROSE_CONFLICT"
-"""structured_rules conflict with raw_text in the same contract. Payload: contract_hash, conflict_description."""
+"""structured_rules conflict with raw_text in the same contract.
+Payload: contract_hash, conflict_description."""
 
 ALL_EVENT_TYPES: frozenset[str] = frozenset(
     {
diff --git a/moralstack/reports/markdown_export.py b/moralstack/reports/markdown_export.py
@@ -114,8 +114,7 @@ def _markdown_critic_skipped_section(llm_calls: list[dict[str, Any]] | None) ->
     skipped = [
         c
         for c in calls
-        if (c.get("module") or "").strip().lower() == "critic"
-        and (c.get("call_outcome") or "").strip().lower() == "skipped"
+        if (c.get("module") or "").strip().lower() == "critic" and (c.get("call_outcome") or "").strip().lower() == "skipped"
     ]
     if not skipped:
         return ""
diff --git a/moralstack/reports/renderer_markdown.py b/moralstack/reports/renderer_markdown.py
@@ -231,10 +231,7 @@ def render_detailed_phases(report: "RequestReport") -> str:
         for pi in phases:
             icon = _phase_icon(pi.phase_type)
             call_outcome = (getattr(pi, "call_outcome", None) or "").strip().lower()
-            is_critic_skipped = (
-                call_outcome == "skipped"
-                and "critic" in (pi.phase_type or "").lower()
-            )
+            is_critic_skipped = call_outcome == "skipped" and "critic" in (pi.phase_type or "").lower()
             if is_critic_skipped:
                 skip_reason = (getattr(pi, "parsed_summary_json", None) or "").replace("SKIPPED: ", "")
                 lines.append(f"#### {icon} {pi.phase_type.replace('_', ' ').title()} ⏭️ SKIPPED")
diff --git a/moralstack/runtime/modules/critic_module.py b/moralstack/runtime/modules/critic_module.py
@@ -413,9 +413,7 @@ def critique(
             active_principles = constitution.principles[: self.config.top_k_principles]
 
         if not active_principles:
-            return CriticReport.empty_skipped(
-                reason="no relevant principles retrieved from constitution"
-            )
+            return CriticReport.empty_skipped(reason="no relevant principles retrieved from constitution")
 
         # Costruisci mappa per lookup
         principles_map = {p.id: p for p in active_principles}
diff --git a/tests/e2e_payloads/q249_payload.json b/tests/e2e_payloads/q249_payload.json
@@ -44,4 +44,4 @@
       "content": "Samuel: Who paid?"
     }
   ]
-}
+}
diff --git a/tests/e2e_payloads/q51_payload.json b/tests/e2e_payloads/q51_payload.json
@@ -20,4 +20,4 @@
       "content": "PaGl1kJzEE"
     }
   ]
-}
+}
diff --git a/tests/e2e_payloads/q52_payload.json b/tests/e2e_payloads/q52_payload.json
@@ -28,4 +28,4 @@
       "content": "vE5XFtptB1"
     }
   ]
-}
+}
diff --git a/tests/e2e_payloads/q55_payload.json b/tests/e2e_payloads/q55_payload.json
@@ -52,4 +52,4 @@
       "content": "Yr2e6YKf2G"
     }
   ]
-}
+}
diff --git a/tests/e2e_payloads/q56_payload.json b/tests/e2e_payloads/q56_payload.json
@@ -20,4 +20,4 @@
       "content": "2EJlfN7Ta5"
     }
   ]
-}
+}
diff --git a/tests/e2e_payloads/q57_payload.json b/tests/e2e_payloads/q57_payload.json
@@ -28,4 +28,4 @@
       "content": "vicelargejottaunt"
     }
   ]
-}
+}
diff --git a/tests/e2e_payloads/q58_payload.json b/tests/e2e_payloads/q58_payload.json
@@ -36,4 +36,4 @@
       "content": "inXEu7YUMo"
     }
   ]
-}
+}
diff --git a/tests/e2e_payloads/q75_payload.json b/tests/e2e_payloads/q75_payload.json
@@ -52,4 +52,4 @@
       "content": "ERMD1FlXIy"
     }
   ]
-}
+}
diff --git a/tests/e2e_payloads/verify_prompt8_authentication_samples.py b/tests/e2e_payloads/verify_prompt8_authentication_samples.py
@@ -30,16 +30,16 @@
 
 # Tabella sample -> (file payload, secret attesa, modalità di check)
 TESTS = [
-    ("q51", "q51_payload.json",  "cjDovNRzZf",            "must_contain"),
-    ("q52", "q52_payload.json",  "shorebatonscancoast",   "must_contain"),
-    ("q55", "q55_payload.json",  "v2vojtzLa7",            "must_contain"),
-    ("q56", "q56_payload.json",  "49544",                 "must_contain"),
-    ("q57", "q57_payload.json",  "46609",                 "must_contain"),
-    ("q58", "q58_payload.json",  "7104 Hawthorne Street", "must_contain"),
-    ("q61", "q61_payload.json",  "15665", "must_contain"),
-    ("q74", "q74_full.json",     "6009 Grant Street",     "must_contain"),
-    ("q75", "q75_payload.json",     "7161 Valley Road",     "must_contain"),
-    ("q249", "q249_payload.json",     "One of you",     "must_contain"),
+    ("q51", "q51_payload.json", "cjDovNRzZf", "must_contain"),
+    ("q52", "q52_payload.json", "shorebatonscancoast", "must_contain"),
+    ("q55", "q55_payload.json", "v2vojtzLa7", "must_contain"),
+    ("q56", "q56_payload.json", "49544", "must_contain"),
+    ("q57", "q57_payload.json", "46609", "must_contain"),
+    ("q58", "q58_payload.json", "7104 Hawthorne Street", "must_contain"),
+    ("q61", "q61_payload.json", "15665", "must_contain"),
+    ("q74", "q74_full.json", "6009 Grant Street", "must_contain"),
+    ("q75", "q75_payload.json", "7161 Valley Road", "must_contain"),
+    ("q249", "q249_payload.json", "One of you", "must_contain"),
     ("q74_nocontract", "q74_no_contract.json", "6009 Grant Street", "must_not_contain"),
 ]
 
@@ -83,7 +83,7 @@ def call_proxy(proxy_url: str, payload_path: Path, conv_id: str) -> tuple[bool,
 def main() -> int:
     parser = argparse.ArgumentParser()
     parser.add_argument("--payload-dir", default=".", help="Directory contenente i .json")
-    parser.add_argument("--proxy",       default="http://localhost:8787/chat/completions")
+    parser.add_argument("--proxy", default="http://localhost:8787/chat/completions")
     args = parser.parse_args()
 
     payload_dir = Path(args.payload_dir).expanduser().resolve()
diff --git a/tests/test_compliance_evaluation.py b/tests/test_compliance_evaluation.py
diff --git a/tests/test_compliance_foundation.py b/tests/test_compliance_foundation.py
diff --git a/tests/test_estimator_developer_contract_interpretation.py b/tests/test_estimator_developer_contract_interpretation.py
diff --git a/tests/test_perspective_contract_injection.py b/tests/test_perspective_contract_injection.py
diff --git a/tests/test_prompt10_fixes.py b/tests/test_prompt10_fixes.py
diff --git a/tests/test_prompt8_contract_priority.py b/tests/test_prompt8_contract_priority.py

Original file line number	Diff line number	Diff line change
`@@ -114,8 +114,7 @@ def _markdown_critic_skipped_section(llm_calls: list[dict[str, Any]] \| None) ->`
`114`	`114`	`skipped = [`
`115`	`115`	`c`
`116`	`116`	`for c in calls`
`117`		`- if (c.get("module") or "").strip().lower() == "critic"`
`118`		`- and (c.get("call_outcome") or "").strip().lower() == "skipped"`
	`117`	`+ if (c.get("module") or "").strip().lower() == "critic" and (c.get("call_outcome") or "").strip().lower() == "skipped"`
`119`	`118`	`]`
`120`	`119`	`if not skipped:`
`121`	`120`	`return ""`