Archive diagnostic checkpoint from indeterminate Stage 7b v3 run

stage7/stage7b_v3_treatment_indeterminate_l40s_v3a/README.md

@@ -0,0 +1,49 @@
+# Stage 7b v3 treatment: immutable indeterminate stop
+
+This directory records the first formal Stage 7b v3 treatment attempt on the
+L40S server. The sealed v3 source gate had passed before treatment started.
+The run completed the Audi specialist's frozen 432 updates, wrote its final
+checkpoint, and passed every Audi training/checkpoint gate. It then stopped
+before the first Tesla update because the next arm's source-state identity gate
+detected an in-memory Adam-state mutation.
+
+The formal status is permanently `stage7b_v3_indeterminate`. Capacity,
+specialization, and oracle-recovery decisions are unavailable. No final endpoint
+evaluation was executed, the Audi checkpoint was not evaluated for attacked AP,
+and neither the Audi arm nor its training metrics may be used as a partial
+scientific result.
+
+## Root cause
+
+The v3 treatment runner loaded the source checkpoint once and passed the same
+Python mapping to sequential training arms. PyTorch's
+`Optimizer.load_state_dict` retained references to Adam state tensors in that
+mapping. The Audi optimizer therefore advanced both its own state and the
+in-memory source mapping from step 1728 to step 2160. The frozen source file on
+disk remained byte-identical and still loads at step 1728. When Tesla attempted
+to initialize, the required source-state validator observed step 2160 and
+stopped the run before Tesla training.
+
+`root_cause_diagnostic.json` records the frozen-source identity, the valid Audi
+training checkpoint identity, the exact failure, and an independent minimal
+PyTorch aliasing reproduction. A subsequent audit may address only this runtime
+nuisance by using a new audit ID and fresh output and by independently reloading
+the byte-identical frozen source checkpoint for every arm. It must not resume or
+reinterpret this v3 run, reuse its Audi checkpoint, or change any scientific
+input, schedule, threshold, loss, seed, optimizer setting, or claim boundary.
+
+## Included evidence
+
+- `decision.json` and `failure.json`: identical formal indeterminate record.
+- `treatment_started.json`: sealed-source treatment boundary with zero prior
+  updates.
+- `partial_specialist_training_results.json`: Audi training/checkpoint audit.
+- `audi_training_metrics.csv`: all 432 frozen training updates, retained only as
+  failure evidence.
+- `runtime_file_access.json`: excluded-path access audit.
+- `launcher.log`: process log through the mechanical stop.
+- `root_cause_diagnostic.json`: read-only post-failure diagnosis.
+- `SHA256SUMS`: hashes of this evidence packet.
+
+The public Hugging Face release metadata is recorded separately after upload and
+immutable-revision verification.

stage7/stage7b_v3_treatment_indeterminate_l40s_v3a/SHA256SUMS

@@ -0,0 +1,3 @@
+50be39699d0450db37a65ec820fd3a8a54fb926add4266e33fb6cb77be19005a  README.md
+1fa39378a52910dda9fded94ea8dd54675a5d90835872bc3a9a8ff396ddd477c  audi_etron_diagnostic_final.pt
+e19995db3358667426352ccb19fb80c837686de49b78dbdbf18914f872bd42cf  root_cause_diagnostic.json

stage7/stage7b_v3_treatment_indeterminate_l40s_v3a/audi_etron_diagnostic_final.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fa39378a52910dda9fded94ea8dd54675a5d90835872bc3a9a8ff396ddd477c
+size 420526562

stage7/stage7b_v3_treatment_indeterminate_l40s_v3a/root_cause_diagnostic.json

@@ -0,0 +1,68 @@
+{
+  "audit_id": "stage7b_per_vehicle_vs_shared_continuation_control_v3",
+  "diagnostic_scope": "post_failure_read_only_runtime_aliasing_diagnosis",
+  "formal_v3_status": "stage7b_v3_indeterminate",
+  "formal_v3_status_is_permanent": true,
+  "capacity_decision": "unavailable",
+  "specialization_decision": "unavailable",
+  "oracle_recovery_decision": "unavailable",
+  "treatment_started_at_utc": "2026-08-04T13:00:46.162866+00:00",
+  "failure_recorded_at_utc": "2026-08-04T13:06:58.772311+00:00",
+  "completed_specialist_training_vehicles": [
+    "audi.etron"
+  ],
+  "completed_final_endpoint_executions": 0,
+  "shared_training_completed": false,
+  "failure": {
+    "error_type": "ValueError",
+    "stage": "tesla_specialist_initial_source_state_validation_before_first_update",
+    "expected_source_adam_step": 1728,
+    "observed_in_memory_source_adam_step": 2160,
+    "expected_source_optimizer_state_sha256": "234f93882fa5aa16c99501d44d4bbcec5dc32252b10d33767a89160b0a94b003",
+    "observed_in_memory_source_optimizer_state_sha256": "7883735ad9300887a50311152fb4de3e175b207e84fd0a8741888a97cd4abf4b"
+  },
+  "frozen_source_on_disk_after_failure": {
+    "checkpoint_sha256": "c9bfa671a06924aa51e2e7ea6036f54b3d761da61ea93b776effa545ef0b6631",
+    "generation_step": 1728,
+    "adam_step_min": 1728.0,
+    "adam_step_max": 1728.0,
+    "optimizer_state_sha256": "234f93882fa5aa16c99501d44d4bbcec5dc32252b10d33767a89160b0a94b003",
+    "unchanged": true
+  },
+  "completed_audi_training_arm": {
+    "status": "training_arm_valid",
+    "all_validity_gates_pass": true,
+    "updates": 432,
+    "failed_validity_gates": [],
+    "final_generation_step": 2160,
+    "final_adam_step_min": 2160.0,
+    "final_adam_step_max": 2160.0,
+    "checkpoint_bytes": 420526562,
+    "checkpoint_sha256": "1fa39378a52910dda9fded94ea8dd54675a5d90835872bc3a9a8ff396ddd477c",
+    "scientific_result_role": "none_failure_evidence_only",
+    "attacked_endpoint_evaluated": false
+  },
+  "minimal_pytorch_alias_reproduction": {
+    "source_state_step_before_loaded_optimizer_step": 1.0,
+    "source_state_step_after_loaded_optimizer_step": 2.0,
+    "alias_mutation_observed": true
+  },
+  "root_cause": {
+    "source_checkpoint_loaded_once_for_all_sequential_arms": true,
+    "same_python_mapping_passed_to_each_training_arm": true,
+    "optimizer_load_state_dict_retained_adam_tensor_references": true,
+    "first_arm_optimizer_steps_mutated_in_memory_source_mapping": true,
+    "next_arm_identity_gate_detected_mutation_before_update": true,
+    "disk_source_checkpoint_corrupted": false,
+    "training_numerical_failure": false
+  },
+  "permitted_minimal_repair_boundary": {
+    "new_audit_id_required": true,
+    "fresh_output_required": true,
+    "v3_resume_permitted": false,
+    "v3_artifact_reuse_as_formal_endpoint_permitted": false,
+    "independent_frozen_source_reload_per_arm_required": true,
+    "scientific_inputs_or_thresholds_may_change": false,
+    "external_narrow_review_before_new_formal_execution_required": true
+  }
+}