Keep unconfigured capability audits operational and report actual helper pacing

Author: Mantiee

rlm/v100/capabilities.py

@@ -40,6 +40,7 @@ def audit(root: Path) -> dict:
             "alerts": "local paper proposals, including rejection flag",
         },
         "helper_endpoint": helper.get("runtime", {}).get("base_url"),
+        "helper_active_time_target_percent": helper.get("resources", {}).get("helper_duty_percent"),
         "requested_context": profile.get("runtime", {}).get("context_window"),
         "flash_attention_requested": profile.get("server", {}).get("flash_attention"),
         "kv_cache_requested": profile.get("server", {}).get("cache_type"),
@@ -66,11 +67,11 @@ def audit(root: Path) -> dict:
         ],
         "master_architecture": {
             "automatic_activation": True,
-            "backend": profile["runtime"].get("backend", "llama.cpp"),
+            "backend": profile.get("runtime", {}).get("backend", "llama.cpp"),
             "custom_scratch_scope": "Arbitrary build(config) network, full-weight warm-start children, isolated byte-token inference and the same fixed/fresh/public gates. Not an automatic universal GGUF converter; inference/training may fail their resource/time limits.",
             "active_expert": profile.get("resources", {}).get("active_foundation_expert"),
             "history": profile.get("resources", {}).get("architecture_history", []),
-            "gate": "Task improvement, no retained regressions, fresh audit, public gate when configured and native boot probe; predecessor kept",
+            "gate": "Task improvement, no retained regressions, fresh audit, public gate when configured and actual inference probe; predecessor kept",
         },
         "owned_external_compute": evidence["external_compute"],
         "colab": {
@@ -91,7 +92,7 @@ def audit(root: Path) -> dict:
             "automatic account creation or bypassing browser authentication/quotas",
         ],
         "outcome_learning_scope": "Causal earlier-input preferences from fully resolved audited paper trades enable LoRA DPO, including losses; observational hindsight labels and CPU shadow reward optimization are not unbiased policy RL or proof of profit",
-        "helper_limit_scope": "30% request active-time target in R&D, not a hard per-process GPU utilization/power cap; game guard pauses only the isolated helper",
+        "helper_limit_scope": "Configured request active-time target in R&D, not a hard per-process GPU utilization/power cap; game guard pauses only the isolated helper",
         "runtime_scope": "Requested configuration; actual execution requires native logs and on-device measurements",
     }
     atomic_json(root / "research/mission/capabilities.json", result)

tests/test_v100_hardware_acceptance.py

@@ -4,7 +4,7 @@
 
 import pytest
 
-from rlm.v100 import competition, hardware_acceptance, mission
+from rlm.v100 import capabilities, competition, hardware_acceptance, mission, progress
 
 
 def test_required_hardware_failure_records_report_and_does_not_start_master(tmp_path, monkeypatch):
@@ -27,3 +27,26 @@ def test_hardware_acceptance_never_runs_beside_active_mission(tmp_path, monkeypa
     with pytest.raises(ValueError, match="stopped"):
         hardware_acceptance.run(tmp_path, tmp_path / "unused.json")
     assert not (tmp_path / "research/hardware-acceptance").exists()
+
+
+def test_capability_audit_before_any_mission_has_no_fake_hardware_evidence(tmp_path, monkeypatch):
+    monkeypatch.setattr(mission, "status", lambda root: {"running": False})
+    monkeypatch.setattr(
+        progress,
+        "report",
+        lambda root: dict.fromkeys(
+            (
+                "last_learning_cycle",
+                "accepted_weight_updates_this_run",
+                "training_metrics",
+                "paper_blockers",
+                "available_audited_paper_outcomes",
+                "drones",
+                "external_compute",
+            )
+        ),
+    )
+    result = capabilities.audit(tmp_path)
+    assert result["mission_running"] is False
+    assert result["helper_active_time_target_percent"] is None
+    assert result["master_architecture"]["backend"] == "llama.cpp"