Phase G: verifier-derived training rewards with kernel-observed divergence penalty

docs/NEXT_STATUS.md

@@ -31,10 +31,10 @@ language — this table is the current status source of truth.
 | A | partial | Brief + runnable lumen task v1 + strict verifier (#33-35), reference-policy A3 comparison (#44) | A3 trained-checkpoint comparison; A5 external reproduction (blocked: partner) |
 | B | partial | Threat model + env scrubbing (#36), container backend (#38), bounded transfer (#39), remainder + attestation + CI image (#40), probes (#41) | B5 cloud mint/storage (blocked: cloud owner); registry image publication (commercial decision) |
 | C | partial | Dossier + cited gate (#37), orphan-label refusal + provenance + fault robustness (#42) | C4 world-native distributions (need SOFA/GPU world revisions); C5 phantom (blocked: partner) |
-| D | partial | Paired comparison + CIs (#43), compare CLI (#45), split manifests with patient/site grouping (#60, #61, #62), scorecard uncertainty views | D5 benchmark-level reporting with external study data |
+| D | partial | Paired comparison + CIs (#43), compare CLI (#45), split manifests with patient/site grouping (#60-62), D5 explicit cohort subgroup validation, head-covered trial bindings, clustered continuous bootstrap, and cloud parity (#64, cloud #5) | D5 benchmark-level reporting with external study data |
 | E | partial | Obs/action contract + A3 (#44), interactive streams (#46), LeRobot reader + MONAI (#51), honest video adapter (#54) | E4 media alignment (no test media; mp4 banned from VC); semantic output schemas (needs E contract design) |
 | F | partial | Prefix-replay branching proof + seed caveat (#47), trajectory-backed forecast task + null baseline (#56), planning utility measurement and recipe (#58) | F1 trajectory benchmark imports; F4 closed-loop simulator branching |
-| G | partial | SB3 PPO recipe + measurement (#48), runnable fixes (#50, #52), kernel-observed episode divergence (#57, #59) | Verifier-derived training rewards (projection withdrawn); Prime interop |
+| G | partial | SB3 PPO recipe + measurement (#48), runnable fixes (#50, #52), kernel-observed episode divergence (#57, #59), verifier-derived training rewards with divergence penalty (#65) | Prime interop |
 | H | partial | Episode resume (#49), crash-safe writes + partial recovery (#53), strict trial provenance + pre-write drift rejection (#59) | Fleet queue/autoscaling (blocked: hosted capacity decision); vectorized stepping (not justified by measured need) |
 
 ### Status and evidence rules

docs/examples/tasks/lumen-nav-safe/verifier.py

@@ -35,6 +35,8 @@ def score(self, context: dict[str, Any]) -> dict[str, Any]:
         safety_max_pen = float(context["safety_max_pen"])
         max_pen = _finite_number(info.get("max_pen"))
         diverged = _exact_bool(info.get("diverged"))
+        if diverged is None and context.get("diverged_observed") is True:
+            diverged = True
         raw_unsafe = info.get("unsafe")
         if raw_unsafe is None:
             unsafe = None if max_pen is None else max_pen > safety_max_pen

src/or_audit/eval/export_verifiers.py

@@ -43,6 +43,7 @@
     GymEnv,
     GymFactory,
     assert_perturbations_applied,
+    episode_diverged,
     make_gym,
     run_gym_episode,
     sample_action,
@@ -995,6 +996,7 @@ def action_fn(env: GymEnv, observation: Any, step: int) -> Any:
             "info": info,
             "trajectory": [dict(step) for step in steps],
             "safety_max_pen": safety,
+            "diverged_observed": episode_diverged(steps),
         },
         runtime=verifier,
     )

src/or_audit/eval/wrap.py

@@ -1062,6 +1062,8 @@ def _render_verifier(spec: WrapRequest) -> str:
         "        info: dict[str, Any] = raw_info if isinstance(raw_info, dict) else {}",
         '        raw_success = _boolean(info, "raw_success", "success", "is_success")',
         '        diverged = _boolean(info, "diverged")',
+        '        if diverged is None and context.get("diverged_observed") is True:',
+        "            diverged = True",
         "        breached: list[str] = []",
         "        unreported: list[str] = []",
         "        for gate_id, signal, predicate in GATES:",

tests/test_eval_run.py

@@ -635,3 +635,38 @@ def test_nonfinite_kind_numpy_scalars() -> None:
     assert nonfinite_kind(np.float64("inf")) == "+inf"
     assert nonfinite_kind(np.float32("-inf")) == "-inf"
     assert nonfinite_kind(np.float32(1.5)) == ""
+
+
+def test_lumen_verifier_consumes_diverged_observed_and_fails_gate() -> None:
+    import importlib.util
+
+    path = ROOT / "docs/examples/tasks/lumen-nav-safe/verifier.py"
+    spec = importlib.util.spec_from_file_location("lumen_verifier", path)
+    assert spec is not None
+    assert spec.loader is not None
+    mod = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mod)
+    verifier = mod.LumenVerifier()
+    res = verifier.score(
+        {
+            "info": {"success": True, "max_pen": 0.05},
+            "safety_max_pen": 0.3,
+            "diverged_observed": True,
+        }
+    )
+    assert res["metrics"]["diverged"] is True
+    assert res["metrics"]["safe_success"] is False
+    assert res["gates"]["wall_penetration"]["status"] == "fail"
+    assert "episode diverged" in res["gates"]["wall_penetration"]["reason"]
+
+    # When diverged_observed is None -> diverged is None (unassessed)
+    res_none = verifier.score(
+        {
+            "info": {"success": True, "max_pen": 0.05},
+            "safety_max_pen": 0.3,
+            "diverged_observed": None,
+        }
+    )
+    assert res_none["metrics"]["diverged"] is None
+    assert res_none["metrics"]["safe_success"] is None
+    assert res_none["gates"]["wall_penetration"]["status"] == "pass"