release(datasets): publish natural-error-v4-two-stage-confirmatory-2026-10-07 natural-error-v4-confirmatory-protocol batch 1/1

SC-VLA Git revision: bd07e540ebd0f90efaed526349c1cfc51c211c61
Release manifest: releases/natural-error-v4-two-stage-confirmatory-2026-10-07/manifest.json

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_commitment_contract_v2.json

@@ -0,0 +1,94 @@
+{
+  "schema_version": 2,
+  "update_id": "U12",
+  "status": "REGISTERED-AFTER-LABEL-AUDIT-BEFORE-CONTROLLED-VALIDATION",
+  "registered_date": "2026-10-06",
+  "purpose": "Confirm only achievements produced during the recorded episode, using source-time manipulation-completion evidence; reject initially true predicates and separate persistent invalidation from missing visual evidence.",
+  "predecessor": {
+    "path": "configs/updates/natural_error_v4_commitment_contract_v1.json",
+    "sha256": "5c126f37fcccea5e71280400907a707d4c4fec821edd52879fb655e11555d69e"
+  },
+  "revision_evidence": {
+    "excluded_pilot_result": {
+      "path": "configs/updates/natural_error_v4_online_evidence_pilot_result_v1.json",
+      "sha256": "1bed2174540fbac081b17baff226922c75bc394544c94357ed54aef69a55a68e",
+      "status": "PASSED-ONLINE-EVIDENCE-CAPTURE-GATE"
+    },
+    "observed": {
+      "pilot_rollouts": 8,
+      "source_time_audit_passes": 8,
+      "initially_true_turnon_rollouts": 2,
+      "initially_true_turnon_frames": 732,
+      "turnon_production_transitions": 0,
+      "terminal_close_truth_frames_per_rollout": 1
+    },
+    "decision": "An initially true predicate is task context, not a producer achievement. A terminal one-frame truth without stable completion is not a historical commitment. Both are excluded before any development collection is opened."
+  },
+  "formal_contract": {
+    "production_eligibility": "E_t^p = 1 only after NOT Y_s^p has been observed for some s <= t in the same episode. This requires a recorded false-to-true production transition and prevents reset-true context from becoming an achievement.",
+    "stable_achievement": "C_t^p = 1 iff E_t^p, raw predicate truth, and predicate-specific manipulation-completion evidence hold for K consecutive control frames and at least one real action-chunk boundary occurs within that run; historical achievement becomes sticky only after C_t^p.",
+    "invalidation_candidate": "I_t^p = H_t^p AND R_t^p AND NOT Y_t^p, where H is confirmed historical achievement, R is consumer-relative continuation requirement, and Y is privileged world-state truth for offline labels.",
+    "persistent_invalidation": "P_t^p = I_t^p AND consecutive_false_frames(I^p) >= D.",
+    "runtime_trigger": "T_t^p = P_t^p AND A_t^p; NOT A produces missing-evidence/reobservation, never a negative validity measurement.",
+    "require_observed_false_before_achievement": true,
+    "stable_achievement_frames_K": 5,
+    "persistent_invalidation_frames_D": 3,
+    "control_frequency_hz": 20,
+    "physical_interpretation": {
+      "K": "250 ms of manipulation-complete truth after a within-episode production transition, including a controller chunk boundary",
+      "D": "150 ms persistence before repair eligibility; shorter lapses remain transient"
+    }
+  },
+  "predicate_specific_completion_evidence": {
+    "spatial_on_or_in": {
+      "required": [
+        "predicate changed from false to true during the recorded episode",
+        "target object has no gripper contact",
+        "end effector has retreated at least 0.08 m from the target object",
+        "target linear speed is at most 0.02 m/s",
+        "one controller action chunk has completed within the stable run"
+      ],
+      "registered_thresholds": {
+        "minimum_eef_target_distance_m": 0.08,
+        "maximum_target_linear_speed_m_per_s": 0.02
+      }
+    },
+    "articulation_close_open_turnon_turnoff": {
+      "required": [
+        "predicate changed from false to true during the recorded episode",
+        "target articulation has no gripper contact",
+        "absolute target joint speed is at most 0.05 rad/s",
+        "one controller action chunk has completed within the stable run"
+      ],
+      "registered_thresholds": {
+        "maximum_target_joint_speed_rad_per_s": 0.05
+      }
+    }
+  },
+  "consumer_requirement": {
+    "rule": "Continuation requirement is relative to an explicit task dependency or terminal goal deadline, not simply all predicates until first conjunction success.",
+    "development_collection": "The first source-time development collection is support discovery only. Root selection is permitted only from confirmed, within-episode achievements; all root and branch gates remain frozen.",
+    "untouched_natural_validation": "For final-goal-only LIBERO predicates, the terminal episode deadline is the consumer; self-recovery before that deadline is reported separately from recovery-critical persistence."
+  },
+  "handoff_rule": "Never switch producer/continuation controllers at first predicate truth. Handoff may occur only after a within-episode stable achievement confirmation and completion of the action chunk that contains confirmation.",
+  "validation_gate": {
+    "minimum_root_states": 12,
+    "required_paired_branch_types": 5,
+    "achievement_false_confirmation_rate_maximum": 0.05,
+    "target_break_recall_minimum": 0.9,
+    "irrelevant_change_false_trigger_maximum": 0.1,
+    "occluded_valid_false_trigger_maximum": 0.05,
+    "occluded_invalid_reobservation_or_abstention_minimum": 0.9,
+    "rule": "K, D, release distance, speed, and the production-transition guard are frozen before source-time development collection. A failed gate is published unchanged."
+  },
+  "forbidden": [
+    "treating first raw truth or reset-true context as achievement",
+    "handoff in the middle of the producing action chunk",
+    "using observation inadmissibility as false predicate truth",
+    "using replayed contact as source-time completion evidence",
+    "using the eight engineering-pilot rollouts for training or evaluation",
+    "opening old validation or confirmatory-test artifacts for tuning",
+    "changing registered thresholds after development labels are read"
+  ],
+  "claim_boundary": "Contract v2 fixes a semantic defect discovered in an excluded capture pilot. It is not accepted as the final algorithm until a new source-time controlled paired-branch gate passes."
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_confirmatory_continuation_v1.json

@@ -0,0 +1,55 @@
+{
+  "schema_version": 1,
+  "update_id": "U13",
+  "status": "REGISTERED-BEFORE-NATURAL-CONTINUATION-GENERATION",
+  "registered_date": "2026-10-07",
+  "purpose": "Render the five preregistered valid post-commitment continuation controls for every audited confirmatory root before any frozen-model scoring.",
+  "protocol": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_v1.json",
+    "sha256": "60303b2ff86e10c145eca7119f75042b54cac96f8b9985b424e62a802bc41264",
+    "required_status": "REGISTERED-BEFORE-CONFIRMATORY-OPENING"
+  },
+  "root_plan": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_root_plan_v1.json",
+    "sha256": "c6163c97675f019ca7fc5b18ebf7f5dcabb47cc69dafce713bfa6ac86c22fe32",
+    "required_status": "SEALED-BEFORE-CONFIRMATORY-BRANCH-GENERATION"
+  },
+  "controlled_result": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_branch_manifest_v1.json",
+    "sha256": "7739bd63d965c12a887d7d13da8973946d13dfc4ce82831f0bd367624ca77680",
+    "required_status": "PASSED-V4-CONFIRMATORY-BRANCH-GENERATION"
+  },
+  "source_policy": {
+    "allowed_splits": ["confirmatory_test"],
+    "allow_confirmatory_opened": true,
+    "training_allowed": false,
+    "model_scoring_during_generation": false
+  },
+  "selection": {
+    "samples_per_root": 5,
+    "eligible": "Frames strictly after confirmation where source-time historically_achieved, continuation_required, and privileged predicate_truth are all one.",
+    "method": "For N ordered eligible frames choose floor(q*(N-1)) at q in {1/6,2/6,3/6,4/6,5/6}; do not inspect rendered images, admission, or model scores.",
+    "split_unit": "source rollout/root",
+    "minimum_eligible_frames_per_root": 5
+  },
+  "simulator": {
+    "suite": "libero_10",
+    "implementation": "stock",
+    "render_device": 0,
+    "resolution": 128,
+    "cameras": ["agentview", "robot0_eye_in_hand"],
+    "minimum_visible_pixels": 16
+  },
+  "expected": {
+    "roots": 31,
+    "samples": 155,
+    "samples_by_split": {"confirmatory_test": 155}
+  },
+  "forbidden": [
+    "selecting frames by rendered appearance, admission, learned score, or downstream metric",
+    "labelling these valid continuations as natural invalidations",
+    "dropping or replacing a root because it has unfavorable continuation samples",
+    "using any confirmatory sample for training, checkpoint selection, calibration, or threshold selection"
+  ],
+  "claim_boundary": "These are naturally evolving valid post-commitment controls from the fixed confirmatory rollouts. They contain no natural invalidation positive and cannot support natural-invalidation recall or prevalence claims."
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_confirmatory_evaluation_v1.json

@@ -0,0 +1,59 @@
+{
+  "schema_version": 1,
+  "update_id": "U13",
+  "status": "REGISTERED-BEFORE-CONFIRMATORY-MODEL-SCORING",
+  "registered_date": "2026-10-07",
+  "purpose": "Apply the three frozen temporal two-stage checkpoints and their validation-selected thresholds to the complete confirmatory feature panel without training, calibration, tuning, or filtering.",
+  "protocol": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_v1.json",
+    "sha256": "60303b2ff86e10c145eca7119f75042b54cac96f8b9985b424e62a802bc41264",
+    "required_status": "REGISTERED-BEFORE-CONFIRMATORY-OPENING"
+  },
+  "feature_result": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_features_result_v1.json",
+    "sha256": "3996c3a866b5112dd28445b4db77c8b600927e0739ff6b5cf78110a5f90cb055",
+    "required_status": "PASSED-V4-TWO-STAGE-CONFIRMATORY-FEATURE-EXTRACTION"
+  },
+  "model_contract": {
+    "path": "configs/updates/natural_error_v4_two_stage_model_v1.json",
+    "sha256": "b6abfbd7d301de2edb63690b4345e01a104532df8240eb94b8b79fedad3091fe",
+    "required_status": "REGISTERED-BEFORE-TWO-STAGE-TRAINING"
+  },
+  "development_result": {
+    "path": "configs/updates/natural_error_v4_two_stage_result_v1.json",
+    "sha256": "008b1bc4a6319e0f441b000662aca4867fee4c0575c437b2890d8f90f2758849",
+    "required_status": "PASSED-V4-TWO-STAGE-DEVELOPMENT-GATE"
+  },
+  "expected": {
+    "roots": 31,
+    "samples": 620,
+    "seeds": [17, 29, 41]
+  },
+  "evaluation": {
+    "batch_size": 128,
+    "bootstrap_replicates": 5000,
+    "bootstrap_unit": "source rollout/root",
+    "bootstrap_pairing": "Use the same sampled root multiset for all three seeds in each replicate, then form the three-seed mean.",
+    "bootstrap_random_seed": 20261007,
+    "confidence_interval": "2.5th and 97.5th percentiles",
+    "training_allowed": false,
+    "checkpoint_selection_allowed": false,
+    "calibration_allowed": false,
+    "threshold_selection_allowed": false,
+    "root_or_sample_filtering_allowed": false
+  },
+  "natural_panel": {
+    "independent_persistent_invalidation_events": 0,
+    "natural_invalidation_recall_status": "UNSUPPORTED",
+    "policy": "Do not infer zero recall, pool controlled target breaks into the natural panel, or collect top-up confirmatory examples."
+  },
+  "gate": "Use the exact confirmatory_gate and three-seed rule in the preregistered protocol without modification.",
+  "forbidden": [
+    "changing checkpoints, network weights, or per-seed thresholds",
+    "training, recalibration, threshold search, or checkpoint selection on confirmatory features",
+    "dropping a root or sample based on scores or metrics",
+    "relabeling controlled target breaks as naturally occurring invalidations",
+    "revising the current U13 model from confirmatory outcomes"
+  ],
+  "claim_boundary": "Passing supports frozen monitor generalization to new LIBERO-10 rollouts and initial states under the registered controlled perturbations. It does not establish naturally occurring invalidation recall, held-out-task robustness, recovery, or task-success improvement."
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_confirmatory_features_result_v1.json

@@ -0,0 +1,27 @@
+{
+  "checkpoint_id": "openvla/openvla-7b",
+  "checkpoint_revision": "47a0ec7fc4ec123775a391911046cf33cf9ed83f",
+  "claim_boundary": "This feature artifact supports frozen scoring of the registered controlled confirmatory panel only. It does not establish natural-invalidation recall, task-OOD robustness, recovery, or task-success improvement.",
+  "config": {
+    "path": "/HELIOS/Robotics/SCVLA_Cover/SCVLA-update-two-stage-verifier/configs/updates/natural_error_v4_two_stage_confirmatory_features_v1.json",
+    "sha256": "f9ebfab714e0b7761ef4adc994550552709a9ab419ecfa09c01aade0e0311105"
+  },
+  "confirmatory_opened": true,
+  "git_dirty": false,
+  "git_revision": "14091eaad15e0d7bf4653c018e976ea721bc0c6f",
+  "output": {
+    "bytes": 22162746,
+    "path": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/features/update/natural-error-v4-two-stage-confirmatory-openvla-v1.hdf5",
+    "sha256": "9097801943b9dde5ae8a927b31cfa9a2df32d72e560240eac59f1ada3145cea3"
+  },
+  "predicate_feature_dim": 4096,
+  "samples": 620,
+  "samples_by_split": {
+    "confirmatory_test": 620
+  },
+  "schema_version": 1,
+  "seconds": 13.944297526002629,
+  "status": "PASSED-V4-TWO-STAGE-CONFIRMATORY-FEATURE-EXTRACTION",
+  "visual_feature_dim": 2176,
+  "visual_tokens": 4
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_confirmatory_features_v1.json

@@ -0,0 +1,74 @@
+{
+  "schema_version": 1,
+  "update_id": "U13",
+  "status": "REGISTERED-BEFORE-CONFIRMATORY-FEATURE-EXTRACTION",
+  "success_status": "PASSED-V4-TWO-STAGE-CONFIRMATORY-FEATURE-EXTRACTION",
+  "registered_date": "2026-10-07",
+  "purpose": "Extract frozen base-OpenVLA temporal features for the complete preregistered confirmatory controlled and valid-continuation panels without model scoring or training.",
+  "protocol": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_v1.json",
+    "sha256": "60303b2ff86e10c145eca7119f75042b54cac96f8b9985b424e62a802bc41264",
+    "required_status": "REGISTERED-BEFORE-CONFIRMATORY-OPENING"
+  },
+  "inputs": {
+    "controlled": {
+      "path": "configs/updates/natural_error_v4_two_stage_confirmatory_branch_manifest_v1.json",
+      "sha256": "7739bd63d965c12a887d7d13da8973946d13dfc4ce82831f0bd367624ca77680",
+      "required_status": "PASSED-V4-CONFIRMATORY-BRANCH-GENERATION",
+      "artifact_root": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/results/update/natural-error-v4-two-stage-confirmatory-branches-v1"
+    },
+    "continuation": {
+      "path": "configs/updates/natural_error_v4_two_stage_confirmatory_continuation_result_v1.json",
+      "sha256": "89193f324b6d77636a6915fe20ac6423d636d4ec5dbaec56f21f121526a6db06",
+      "required_status": "PASSED-V4-NATURAL-CONTINUATION-GENERATION",
+      "artifact_root": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/results/update/natural-error-v4-two-stage-confirmatory-continuations-v1"
+    },
+    "root_plan": {
+      "path": "configs/updates/natural_error_v4_two_stage_confirmatory_root_plan_v1.json",
+      "sha256": "c6163c97675f019ca7fc5b18ebf7f5dcabb47cc69dafce713bfa6ac86c22fe32",
+      "required_status": "SEALED-BEFORE-CONFIRMATORY-BRANCH-GENERATION"
+    },
+    "confirmatory_opened": true
+  },
+  "source_policy": {
+    "allow_confirmatory_opened": true,
+    "training_allowed": false,
+    "model_scoring_during_extraction": false,
+    "sample_or_root_filtering_allowed": false
+  },
+  "samples": {
+    "unit": "root x candidate frame x camera",
+    "roots": 31,
+    "controlled": 310,
+    "natural_valid_continuation": 310,
+    "total": 620,
+    "by_split": {"confirmatory_test": 620},
+    "split_unit": "source rollout/root"
+  },
+  "frozen_representation": {
+    "checkpoint_id": "openvla/openvla-7b",
+    "checkpoint_revision": "47a0ec7fc4ec123775a391911046cf33cf9ed83f",
+    "encoder": "OpenVLA frozen DINOv2+SigLIP image backbone and Llama predicate backbone",
+    "quantization": "4bit",
+    "visual_pooling_grid": 2,
+    "depth_pooling_resolution": 32,
+    "feature_batch_size": 16,
+    "no_libero_finetuned_checkpoint": true
+  },
+  "targets": {
+    "proposal": "one iff the native route is REPAIR or REOBSERVE",
+    "admission": "per-camera privileged admissibility",
+    "invalidity": "one iff privileged world truth is false",
+    "native_route": {"CONTINUE": 0, "REPAIR": 1, "REOBSERVE": 2},
+    "natural_continuations": "valid hard negatives only",
+    "controlled_target_breaks": "the only invalid positives in this controlled confirmatory panel"
+  },
+  "forbidden": [
+    "training or updating any encoder or verifier weight",
+    "checkpoint, calibration, or threshold selection",
+    "dropping roots or samples after feature inspection",
+    "using a LIBERO-finetuned OpenVLA representation",
+    "calling controlled target-break performance natural-invalidation recall"
+  ],
+  "claim_boundary": "This feature artifact supports frozen scoring of the registered controlled confirmatory panel only. It does not establish natural-invalidation recall, task-OOD robustness, recovery, or task-success improvement."
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_confirmatory_generation_v1.json

@@ -0,0 +1,92 @@
+{
+  "schema_version": 1,
+  "update_id": "U13",
+  "status": "REGISTERED-BEFORE-PAIRED-BRANCH-GENERATION",
+  "success_status": "PASSED-V4-CONFIRMATORY-BRANCH-GENERATION",
+  "failure_status": "FAILED-V4-CONFIRMATORY-BRANCH-GENERATION",
+  "registered_date": "2026-10-07",
+  "purpose": "Generate the preregistered five-way causal panel for all 31 audited confirmatory roots without training, calibration, threshold selection, sample replacement, or model scoring.",
+  "protocol": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_v1.json",
+    "sha256": "60303b2ff86e10c145eca7119f75042b54cac96f8b9985b424e62a802bc41264",
+    "required_status": "REGISTERED-BEFORE-CONFIRMATORY-OPENING"
+  },
+  "pairing": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_pairing_v1.json",
+    "sha256": "ada59172048f22c8fe59499b2c32fc98df90841024bbd036d63fb4d9ca2a5e22",
+    "required_status": "REGISTERED-AFTER-CONFIRMATORY-ROOT-AUDIT-BEFORE-BRANCH-GENERATION"
+  },
+  "root_plan": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_root_plan_v1.json",
+    "sha256": "c6163c97675f019ca7fc5b18ebf7f5dcabb47cc69dafce713bfa6ac86c22fe32",
+    "required_status": "SEALED-BEFORE-CONFIRMATORY-BRANCH-GENERATION"
+  },
+  "contract": {
+    "path": "configs/updates/natural_error_v4_commitment_contract_v2.json",
+    "sha256": "fd5c1c6619bc37aa27ace2b1017209b54837193af89dbb54042661b0c8548500",
+    "required_status": "REGISTERED-AFTER-LABEL-AUDIT-BEFORE-CONTROLLED-VALIDATION"
+  },
+  "simulator": {
+    "suite": "libero_10",
+    "implementation": "stock",
+    "render_device": 0,
+    "resolution": 128,
+    "cameras": ["agentview", "robot0_eye_in_hand"],
+    "minimum_visible_pixels": 16
+  },
+  "temporal_evaluation": {
+    "frames_per_branch": 3,
+    "world_state_policy": "Hold the selected post-intervention simulator state fixed for D frames; this is a controlled confirmatory component test, not a natural physical disturbance trajectory.",
+    "required_match": "frames_per_branch must equal persistent_invalidation_frames_D from the frozen commitment contract"
+  },
+  "achievement_prefix_audit": {
+    "enabled": true,
+    "source": "Use only source predicate truth, live source-time completion evidence, production eligibility, and actual queue-empty chunk boundaries stored in each sealed trace.",
+    "registered_gate": "achievement_false_confirmation_rate <= 0.05 and all 31 roots confirm exactly at their sealed frame"
+  },
+  "interventions": {
+    "no_change": {
+      "world": "Use the byte-verified sealed root state without modification.",
+      "sensor": "Render both registered cameras without modification."
+    },
+    "target_break_visible": {
+      "world": "Use the frozen deterministic spatial-offset candidate order and select the highest target-admission/visible-pixel score; ties retain first candidate order.",
+      "sensor": "Require the target predicate to be false and admissible in at least one registered camera."
+    },
+    "irrelevant_change_visible": {
+      "mode": "sensor_photometric_nuisance",
+      "world": "Use the byte-verified sealed root state so the complete predicate truth vector is identical to no_change.",
+      "sensor": "For every non-monitored-argument RGB pixel, apply channel rotation RGB to GBR, multiply by 0.85, add 12, and clip to uint8. Preserve monitored-argument RGB pixels, depth, and segmentation exactly.",
+      "scope": "Registered controlled hard negative; not evidence of natural irrelevant-object motion."
+    },
+    "occluded_valid": {
+      "world": "Use the unchanged sealed root state.",
+      "sensor": "Apply deterministic privileged sensor dropout independently per camera using the frozen four-pixel padded argument rectangle; if it spans the full frame, use the registered per-channel global-median fill; require target inadmissibility in both cameras."
+    },
+    "target_break_occluded": {
+      "world": "Use the exact selected target_break_visible state.",
+      "sensor": "Apply the identical registered sensor dropout and full-frame fallback as occluded_valid; require target inadmissibility in both cameras."
+    }
+  },
+  "evaluation_policy": {
+    "training_allowed": false,
+    "checkpoint_selection_allowed": false,
+    "calibration_allowed": false,
+    "threshold_selection_allowed": false,
+    "root_or_branch_filtering_allowed": false
+  },
+  "failure_policy": [
+    "Do not replace, drop, or add a root after generation begins.",
+    "Preserve a partial or incomplete panel with FAILED status and exact diagnostics.",
+    "Register any feasibility-policy change before a full rerun; do not silently filter failed roots.",
+    "Do not change intervention search, K, D, visibility, model, or evaluation gates from confirmatory outcomes."
+  ],
+  "expected": {
+    "roots": 31,
+    "roots_by_split": {"confirmatory_test": 31},
+    "branches_per_root": 5,
+    "total_branches": 155,
+    "branches_by_split": {"confirmatory_test": 155}
+  },
+  "claim_boundary": "Passing branch generation establishes that the frozen controlled confirmatory panel is complete and contract-consistent. It does not by itself establish learned-verifier accuracy, natural-invalidation recall, recovery, or task success."
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_confirmatory_pairing_v1.json

@@ -0,0 +1,48 @@
+{
+  "schema_version": 1,
+  "update_id": "U13",
+  "status": "REGISTERED-AFTER-CONFIRMATORY-ROOT-AUDIT-BEFORE-BRANCH-GENERATION",
+  "registered_date": "2026-10-07",
+  "purpose": "Freeze the five-way causal pairing for every root retained by the preregistered confirmatory audit, before generating any confirmatory branch or model score.",
+  "protocol": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_v1.json",
+    "sha256": "60303b2ff86e10c145eca7119f75042b54cac96f8b9985b424e62a802bc41264",
+    "required_status": "REGISTERED-BEFORE-CONFIRMATORY-OPENING"
+  },
+  "root_plan": {
+    "path": "configs/updates/natural_error_v4_two_stage_confirmatory_root_plan_v1.json",
+    "sha256": "c6163c97675f019ca7fc5b18ebf7f5dcabb47cc69dafce713bfa6ac86c22fe32",
+    "required_status": "SEALED-BEFORE-CONFIRMATORY-BRANCH-GENERATION",
+    "roots": 31,
+    "supported_strata": 8,
+    "retention": "Use all 31 audited roots. The singleton butter stratum is retained in pooled/root-bootstrap evaluation but is not counted as a supported stratum."
+  },
+  "contract": {
+    "path": "configs/updates/natural_error_v4_commitment_contract_v2.json",
+    "sha256": "fd5c1c6619bc37aa27ace2b1017209b54837193af89dbb54042661b0c8548500",
+    "required_status": "REGISTERED-AFTER-LABEL-AUDIT-BEFORE-CONTROLLED-VALIDATION"
+  },
+  "data_role": "controlled-confirmatory-only",
+  "paired_branches": [
+    {"id": "no_change", "world_valid": true, "observation_admissible": true},
+    {"id": "target_break_visible", "world_valid": false, "observation_admissible": true},
+    {"id": "irrelevant_change_visible", "world_valid": true, "observation_admissible": true},
+    {"id": "occluded_valid", "world_valid": true, "observation_admissible": false},
+    {"id": "target_break_occluded", "world_valid": false, "observation_admissible": false}
+  ],
+  "total_branches": 155,
+  "split_unit": "source rollout/root",
+  "data_use": {
+    "training": "forbidden",
+    "checkpoint_selection": "forbidden",
+    "calibration_or_threshold_selection": "forbidden",
+    "evaluation": "Apply only the preregistered checkpoints and per-seed thresholds."
+  },
+  "failure_policy": [
+    "Do not replace, drop, or add a root after this registration.",
+    "If any registered branch is infeasible, preserve the failed complete-panel attempt and diagnostics; do not filter the root.",
+    "Do not change perturbation search, sensor intervention, K, D, minimum pixels, checkpoints, or thresholds from confirmatory outcomes.",
+    "Do not call controlled target breaks natural invalidation events."
+  ],
+  "claim_boundary": "This pairing supports a frozen controlled confirmatory test on new rollouts and initial states. It is not natural-invalidation prevalence or recall evidence, and it is not recovery or task-success evidence."
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_confirmatory_result_v1.json

@@ -0,0 +1,231 @@
+{
+  "aggregate_metrics": {
+    "admission_macro_f1": {
+      "all_seeds_passed": true,
+      "maximum": 0.970024174053183,
+      "mean": 0.9644436332225043,
+      "mean_passed": true,
+      "minimum": 0.9600322320709105,
+      "passed": true,
+      "root_bootstrap_95_ci": {
+        "lower_95": 0.9331924833265042,
+        "upper_95": 0.9867607400074128
+      },
+      "sample_sd": 0.005097537820966451,
+      "threshold": 0.9
+    },
+    "irrelevant_change_false_repair": {
+      "all_seeds_passed": true,
+      "maximum": 0.0,
+      "mean": 0.0,
+      "mean_passed": true,
+      "minimum": 0.0,
+      "passed": true,
+      "root_bootstrap_95_ci": {
+        "lower_95": 0.0,
+        "upper_95": 0.0
+      },
+      "sample_sd": 0.0,
+      "threshold": 0.15
+    },
+    "native_route_macro_f1": {
+      "all_seeds_passed": true,
+      "maximum": 0.9682742981588967,
+      "mean": 0.9539298210968373,
+      "mean_passed": true,
+      "minimum": 0.9462726371958552,
+      "passed": true,
+      "root_bootstrap_95_ci": {
+        "lower_95": 0.9248436366539642,
+        "upper_95": 0.9760216247144702
+      },
+      "sample_sd": 0.012432143365075787,
+      "threshold": 0.85
+    },
+    "observable_invalidation_recall": {
+      "all_seeds_passed": true,
+      "maximum": 0.9516129032258065,
+      "mean": 0.903225806451613,
+      "mean_passed": true,
+      "minimum": 0.8709677419354839,
+      "passed": true,
+      "root_bootstrap_95_ci": {
+        "lower_95": 0.8333333333333334,
+        "upper_95": 0.9623655913978495
+      },
+      "sample_sd": 0.042673408242977294,
+      "threshold": 0.75
+    },
+    "observable_validity_macro_f1": {
+      "all_seeds_passed": true,
+      "maximum": 0.9901639344262295,
+      "mean": 0.9885565642380157,
+      "mean_passed": true,
+      "minimum": 0.9853418238615883,
+      "passed": true,
+      "root_bootstrap_95_ci": {
+        "lower_95": 0.9764207995897433,
+        "upper_95": 0.9968035218544523
+      },
+      "sample_sd": 0.002784046832557743,
+      "threshold": 0.9
+    },
+    "proposal_false_positive_rate": {
+      "all_seeds_passed": true,
+      "maximum": 0.009868421052631578,
+      "mean": 0.007675438596491228,
+      "mean_passed": true,
+      "minimum": 0.006578947368421052,
+      "passed": true,
+      "root_bootstrap_95_ci": {
+        "lower_95": 0.0,
+        "upper_95": 0.017435897435897435
+      },
+      "sample_sd": 0.0018991785170711373,
+      "threshold": 0.25
+    },
+    "proposal_recall": {
+      "all_seeds_passed": true,
+      "maximum": 0.9493670886075949,
+      "mean": 0.9261603375527425,
+      "mean_passed": true,
+      "minimum": 0.9113924050632911,
+      "passed": true,
+      "root_bootstrap_95_ci": {
+        "lower_95": 0.8634399559369964,
+        "upper_95": 0.9701986754966887
+      },
+      "sample_sd": 0.020345254770027313,
+      "threshold": 0.9
+    }
+  },
+  "artifacts": [
+    {
+      "checkpoint": {
+        "path": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/checkpoints/update/natural-error-v4-two-stage-v1/seed-17/best.pt",
+        "sha256": "074ef6bc997dfa39310a3ad9056ed6d1f0faf2c7000f8838404b209ef875b1a2"
+      },
+      "predictions": {
+        "bytes": 5435,
+        "path": "seed-17-predictions.npz",
+        "sha256": "9577dee8d0f59771ed867b236b2fcb331c60ec86ce8723566b5fe9e56351f3e1"
+      },
+      "seed": 17
+    },
+    {
+      "checkpoint": {
+        "path": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/checkpoints/update/natural-error-v4-two-stage-v1/seed-29/best.pt",
+        "sha256": "dbdfe0a8da046925b5891a13921bc477a475cace25d1f1d84da0ef98670ab34a"
+      },
+      "predictions": {
+        "bytes": 5443,
+        "path": "seed-29-predictions.npz",
+        "sha256": "cdca477a779d8230c652465698126e0c27027bbb8d028b768c89a95580dc6ed8"
+      },
+      "seed": 29
+    },
+    {
+      "checkpoint": {
+        "path": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/checkpoints/update/natural-error-v4-two-stage-v1/seed-41/best.pt",
+        "sha256": "61954e149c007b9a75a9b2010e8c749768a8c2386484b62e1ec2d05b3fc7c7c3"
+      },
+      "predictions": {
+        "bytes": 5462,
+        "path": "seed-41-predictions.npz",
+        "sha256": "084c46449252ca06cfd7e10ca85522b947a0da6c5132b0545d96bb315b20ef4c"
+      },
+      "seed": 41
+    }
+  ],
+  "bootstrap": {
+    "estimand": "three-seed mean",
+    "paired_across_seeds": true,
+    "random_seed": 20261007,
+    "replicates": 5000,
+    "unit": "source rollout/root"
+  },
+  "claim_boundary": "Passing supports frozen monitor generalization to new LIBERO-10 rollouts and initial states under the registered controlled perturbations. It does not establish naturally occurring invalidation recall, held-out-task robustness, recovery, or task-success improvement.",
+  "config": {
+    "path": "/HELIOS/Robotics/SCVLA_Cover/SCVLA-update-two-stage-verifier/configs/updates/natural_error_v4_two_stage_confirmatory_evaluation_v1.json",
+    "sha256": "6b22652db3161c11d5f158c5189a43369af5ae8213c376ead0fb08ee07e34b64"
+  },
+  "confirmatory_opened": true,
+  "device": "cuda",
+  "feature": {
+    "path": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/features/update/natural-error-v4-two-stage-confirmatory-openvla-v1.hdf5",
+    "sha256": "9097801943b9dde5ae8a927b31cfa9a2df32d72e560240eac59f1ada3145cea3"
+  },
+  "gate_passed": true,
+  "generated_at": "2026-10-06T17:21:21.801612+00:00",
+  "git_dirty": false,
+  "git_revision": "869fcf09a5b04a3cb6dbe9bba2e58d0d279f3b3a",
+  "natural_invalidation_recall": {
+    "independent_persistent_events": 0,
+    "reason": "The preregistered natural panel contained no independent persistent invalidation event.",
+    "status": "UNSUPPORTED"
+  },
+  "peak_gpu_memory_bytes": 158559744,
+  "roots": 31,
+  "samples": 620,
+  "schema_version": 1,
+  "seconds": 7.583765465009492,
+  "seed_results": [
+    {
+      "metrics": {
+        "admission_macro_f1": 0.963274493543419,
+        "irrelevant_change_false_repair": 0.0,
+        "native_route_macro_f1": 0.9462726371958552,
+        "observable_invalidation_recall": 0.8709677419354839,
+        "observable_validity_macro_f1": 0.9853418238615883,
+        "proposal_false_positive_rate": 0.006578947368421052,
+        "proposal_recall": 0.9177215189873418
+      },
+      "seed": 17,
+      "thresholds": {
+        "admission": 0.52734375,
+        "invalidity": 0.984375,
+        "proposal": 0.9921875
+      }
+    },
+    {
+      "metrics": {
+        "admission_macro_f1": 0.9600322320709105,
+        "irrelevant_change_false_repair": 0.0,
+        "native_route_macro_f1": 0.9472425279357601,
+        "observable_invalidation_recall": 0.8870967741935484,
+        "observable_validity_macro_f1": 0.9901639344262295,
+        "proposal_false_positive_rate": 0.006578947368421052,
+        "proposal_recall": 0.9113924050632911
+      },
+      "seed": 29,
+      "thresholds": {
+        "admission": 0.85546875,
+        "invalidity": 0.96875,
+        "proposal": 0.98828125
+      }
+    },
+    {
+      "metrics": {
+        "admission_macro_f1": 0.970024174053183,
+        "irrelevant_change_false_repair": 0.0,
+        "native_route_macro_f1": 0.9682742981588967,
+        "observable_invalidation_recall": 0.9516129032258065,
+        "observable_validity_macro_f1": 0.9901639344262295,
+        "proposal_false_positive_rate": 0.009868421052631578,
+        "proposal_recall": 0.9493670886075949
+      },
+      "seed": 41,
+      "thresholds": {
+        "admission": 0.75,
+        "invalidity": 0.71484375,
+        "proposal": 1.0
+      }
+    }
+  ],
+  "seeds": [
+    17,
+    29,
+    41
+  ],
+  "status": "PASSED-V4-TWO-STAGE-CONFIRMATORY-GATE"
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_confirmatory_v1.json

@@ -0,0 +1,114 @@
+{
+  "schema_version": 1,
+  "update_id": "U13",
+  "status": "REGISTERED-BEFORE-CONFIRMATORY-OPENING",
+  "registered_date": "2026-10-07",
+  "purpose": "Evaluate the frozen temporal two-stage verifier on the untouched v4 confirmatory split without retraining, recalibration, threshold selection, or sample replacement.",
+  "source": {
+    "plan": {
+      "path": "configs/updates/natural_error_v4_trainval_plan_v1.json",
+      "sha256": "04aceb38c0361ecf134d0ba40895a08aff896fe5d6275c0be6c6d0b0854b72aa",
+      "required_status": "SEALED-BEFORE-COLLECTION"
+    },
+    "integrity_result": {
+      "path": "configs/updates/natural_error_v4_trainval_result_v1.json",
+      "sha256": "e4d508e495519ee49d8d65dc7032a1014233b6fd45056852f480aa6eb3fde671",
+      "required_status": "PASSED-V4-TRAINVAL-SUPPORT-GATE"
+    },
+    "contract": {
+      "path": "configs/updates/natural_error_v4_commitment_contract_v2.json",
+      "sha256": "fd5c1c6619bc37aa27ace2b1017209b54837193af89dbb54042661b0c8548500"
+    },
+    "collection_root": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/runs/update/natural-error-v4-trainval-v1",
+    "split": "confirmatory_test",
+    "rollouts": 40
+  },
+  "frozen_model": {
+    "contract": {
+      "path": "configs/updates/natural_error_v4_two_stage_model_v1.json",
+      "sha256": "b6abfbd7d301de2edb63690b4345e01a104532df8240eb94b8b79fedad3091fe"
+    },
+    "development_result": {
+      "path": "configs/updates/natural_error_v4_two_stage_result_v1.json",
+      "sha256": "008b1bc4a6319e0f441b000662aca4867fee4c0575c437b2890d8f90f2758849",
+      "required_status": "PASSED-V4-TWO-STAGE-DEVELOPMENT-GATE"
+    },
+    "huggingface_release": "natural-error-v4-two-stage-development-2026-10-07",
+    "checkpoint_root": "/HELIOS/Robotics/SCVLA_Cover/Benchmarks/SCVLA/checkpoints/update/natural-error-v4-two-stage-v1",
+    "seeds": {
+      "17": {
+        "best_sha256": "074ef6bc997dfa39310a3ad9056ed6d1f0faf2c7000f8838404b209ef875b1a2",
+        "proposal_threshold": 0.9921875,
+        "admission_threshold": 0.52734375,
+        "invalidity_threshold": 0.984375
+      },
+      "29": {
+        "best_sha256": "dbdfe0a8da046925b5891a13921bc477a475cace25d1f1d84da0ef98670ab34a",
+        "proposal_threshold": 0.98828125,
+        "admission_threshold": 0.85546875,
+        "invalidity_threshold": 0.96875
+      },
+      "41": {
+        "best_sha256": "61954e149c007b9a75a9b2010e8c749768a8c2386484b62e1ec2d05b3fc7c7c3",
+        "proposal_threshold": 1.0,
+        "admission_threshold": 0.75,
+        "invalidity_threshold": 0.71484375
+      }
+    }
+  },
+  "root_audit": {
+    "selection": "For each confirmatory rollout, select at most one root: the earliest source-time contract-v2 confirmation, tie-broken by predicate index. Do not select by image, natural invalidation occurrence, branch feasibility, or model score.",
+    "minimum_selected_roots": 20,
+    "minimum_supported_strata": 6,
+    "minimum_roots_per_supported_stratum": 2,
+    "artifact_audit_pass_rate": 1.0,
+    "source_time_provenance_rate": 1.0,
+    "completion_contact_conflicts_maximum": 0,
+    "reset_true_confirmations_maximum": 0
+  },
+  "natural_panel": {
+    "scope": "All post-confirmation frames while the predicate is historically achieved and continuation-required.",
+    "positive": "Contract-v2 persistent invalidation onset after three consecutive false frames.",
+    "report": ["independent_events", "tasks", "predicates", "frame_support"],
+    "minimum_independent_events_for_recall_claim": 20,
+    "failure_policy": "If support is below 20 events, report natural-invalidation recall as unsupported; do not top up, relabel, or tune."
+  },
+  "controlled_panel": {
+    "branches": ["no_change", "target_break_visible", "irrelevant_change_visible", "occluded_valid", "target_break_occluded"],
+    "branches_per_root": 5,
+    "continuations_per_root": 5,
+    "continuation_selection": "Use ordered eligible frames strictly after confirmation at floor(q*(N-1)) for q in {1/6,2/6,3/6,4/6,5/6}; never select by pixels or model score.",
+    "cameras": ["agentview", "robot0_eye_in_hand"],
+    "minimum_visible_pixels": 16,
+    "simulator": "stock LIBERO-10 at the already pinned revision",
+    "feature_encoder": "frozen base OpenVLA revision 47a0ec7fc4ec123775a391911046cf33cf9ed83f"
+  },
+  "evaluation": {
+    "training_allowed": false,
+    "checkpoint_selection_allowed": false,
+    "calibration_allowed": false,
+    "threshold_selection_allowed": false,
+    "bootstrap_replicates": 5000,
+    "bootstrap_unit": "source rollout/root",
+    "metrics": ["proposal_recall", "proposal_false_positive_rate", "admission_macro_f1", "observable_invalidation_recall", "observable_validity_macro_f1", "irrelevant_change_false_repair", "native_route_macro_f1"]
+  },
+  "confirmatory_gate": {
+    "proposal_recall_minimum": 0.9,
+    "proposal_false_positive_rate_maximum": 0.25,
+    "admission_macro_f1_minimum": 0.9,
+    "observable_invalidation_recall_minimum": 0.75,
+    "observable_validity_macro_f1_minimum": 0.9,
+    "irrelevant_change_false_repair_maximum": 0.15,
+    "native_route_macro_f1_minimum": 0.85,
+    "seed_rule": "Every metric threshold must pass in the three-seed mean; proposal recall and observable invalidation recall must also pass in every individual seed."
+  },
+  "forbidden": [
+    "changing checkpoints, architecture, weights, or thresholds after this registration",
+    "training on or calibrating from confirmatory artifacts",
+    "dropping a root because an intervention, metric, or model score is unfavorable",
+    "calling controlled target-break recall natural-invalidation recall",
+    "selectively adding confirmatory rollouts or examples",
+    "using confirmatory outcomes to revise the current U13 claim"
+  ],
+  "claim_boundary": "Passing the controlled confirmatory panel supports frozen monitor generalization to new rollouts and initial states under registered perturbations. It does not establish naturally occurring invalidation recall, task-OOD robustness, recovery, or task-success improvement."
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_model_v1.json

@@ -0,0 +1,68 @@
+{
+  "schema_version": 1,
+  "update_id": "U13",
+  "status": "REGISTERED-BEFORE-TWO-STAGE-TRAINING",
+  "registered_date": "2026-10-06",
+  "purpose": "Evaluate whether a temporal proposal plus semantic verifier separates controlled invalidation from natural-valid continuation and registered nuisance controls.",
+  "inputs": {
+    "training_contract": {
+      "path": "configs/updates/natural_error_v4_two_stage_training_v1.json",
+      "sha256": "d3dbf428fbe4750eacc34fc1c45a802df07b2ff30f463d1e019db6313f83bd78"
+    },
+    "feature_result": {
+      "path": "configs/updates/natural_error_v4_two_stage_features_result_v1.json",
+      "sha256": "8e774db4b680cc747ec6c47fbc79a337c28774f35f8aff97d9dc3a36fb815508",
+      "required_status": "PASSED-V4-TWO-STAGE-FEATURE-EXTRACTION"
+    },
+    "confirmatory_opened": false
+  },
+  "architecture": {
+    "name": "TemporalTwoStageVerifier",
+    "hidden_dim": 256,
+    "depth_width": 32,
+    "attention_heads": 8,
+    "fusion_layers": 2,
+    "dropout": 0.1,
+    "proposal": "MLP over predicate, pooled reference/current/delta visual tokens, and paired-depth token",
+    "verifier": "Transformer over predicate, paired-depth, reference, current, and signed-delta tokens",
+    "outputs": ["proposal", "admission", "invalidity"],
+    "native_route": "proposal<threshold => CONTINUE; otherwise admission<threshold => REOBSERVE; otherwise invalidity>=threshold => REPAIR; else CONTINUE"
+  },
+  "loss": {
+    "proposal_coefficient": 1.5,
+    "admission_coefficient": 1.0,
+    "invalidity_coefficient": 1.5,
+    "invalidity_scope": "ground-truth admitted observations only",
+    "class_weighting": "inverse square root of train-only class frequencies",
+    "gradient_clip_norm": 1.0
+  },
+  "checkpoint_selection": {
+    "objective": "minimum validation joint loss",
+    "minimum_epochs": 10,
+    "early_stopping_patience": 10,
+    "maximum_epochs": 50
+  },
+  "threshold_selection": {
+    "split": "validation only after checkpoint selection",
+    "proposal": "highest threshold satisfying recall >= 0.90; if infeasible maximize recall, then minimize false-positive rate",
+    "admission": "maximum binary macro-F1, tie-broken by lower false admission and higher threshold",
+    "invalidity": "maximum binary macro-F1 on ground-truth admitted observations, tie-broken by higher invalidation recall and higher threshold"
+  },
+  "development_gate": {
+    "proposal_recall_minimum": 0.9,
+    "proposal_false_positive_rate_maximum": 0.25,
+    "admission_macro_f1_minimum": 0.9,
+    "observable_invalidation_recall_minimum": 0.75,
+    "observable_validity_macro_f1_minimum": 0.9,
+    "irrelevant_change_false_repair_maximum": 0.15,
+    "native_route_macro_f1_minimum": 0.85,
+    "seed_rule": "All thresholds pass in the three-seed mean; proposal recall and observable invalidation recall also pass for every seed."
+  },
+  "forbidden": [
+    "opening confirmatory labels or scores",
+    "changing this architecture, loss, threshold policy, seeds, or gates after validation metrics are read",
+    "calling controlled target-break performance natural-invalidation recall",
+    "using natural-valid continuation frames as invalid positives"
+  ],
+  "claim_boundary": "This is development evidence on controlled target breaks, natural-valid continuation hard negatives, and registered nuisance controls. It is not confirmatory natural-invalidation or recovery evidence."
+}

releases/natural-error-v4-two-stage-confirmatory-2026-10-07/updates/natural-error-v4/two-stage-confirmatory-v1/protocol/natural_error_v4_two_stage_result_v1.json

@@ -0,0 +1,248 @@
+{
+  "aggregate_metrics": {
+    "admission_macro_f1": {
+      "all_seeds_passed": true,
+      "maximum": 0.9768407346930064,
+      "mean": 0.9735322682205787,
+      "mean_passed": true,
+      "minimum": 0.970223801748151,
+      "passed": true,
+      "sample_sd": 0.0033084664724276625,
+      "threshold": 0.9
+    },
+    "irrelevant_change_false_repair": {
+      "all_seeds_passed": true,
+      "maximum": 0.0,
+      "mean": 0.0,
+      "mean_passed": true,
+      "minimum": 0.0,
+      "passed": true,
+      "sample_sd": 0.0,
+      "threshold": 0.15
+    },
+    "native_route_macro_f1": {
+      "all_seeds_passed": true,
+      "maximum": 0.9650876417481715,
+      "mean": 0.9544686326035053,
+      "mean_passed": true,
+      "minimum": 0.9485538852498214,
+      "passed": true,
+      "sample_sd": 0.009216226735234721,
+      "threshold": 0.85
+    },
+    "observable_invalidation_recall": {
+      "all_seeds_passed": true,
+      "maximum": 0.9354838728904724,
+      "mean": 0.9139784773190817,
+      "mean_passed": true,
+      "minimum": 0.9032257795333862,
+      "passed": true,
+      "sample_sd": 0.018624218883257784,
+      "threshold": 0.75
+    },
+    "observable_validity_macro_f1": {
+      "all_seeds_passed": true,
+      "maximum": 1.0,
+      "mean": 0.9967297488100931,
+      "mean_passed": true,
+      "minimum": 0.9950946232151396,
+      "passed": true,
+      "sample_sd": 0.0028321206072156697,
+      "threshold": 0.9
+    },
+    "proposal_false_positive_rate": {
+      "all_seeds_passed": true,
+      "maximum": 0.003367003367003367,
+      "mean": 0.003367003367003367,
+      "mean_passed": true,
+      "minimum": 0.003367003367003367,
+      "passed": true,
+      "sample_sd": 0.0,
+      "threshold": 0.25
+    },
+    "proposal_recall": {
+      "all_seeds_passed": true,
+      "maximum": 0.934984520123839,
+      "mean": 0.9164086687306501,
+      "mean_passed": true,
+      "minimum": 0.9040247678018576,
+      "passed": true,
+      "sample_sd": 0.016382361059223492,
+      "threshold": 0.9
+    }
+  },
+  "artifacts": [
+    {
+      "bytes": 52318007,
+      "name": "best.pt",
+      "seed": 17,
+      "sha256": "074ef6bc997dfa39310a3ad9056ed6d1f0faf2c7000f8838404b209ef875b1a2"
+    },
+    {
+      "bytes": 6914,
+      "name": "history.json",
+      "seed": 17,
+      "sha256": "4b0760f7c408e21089b696973abf5796649bc3fee21dc2111c07154fb64f0c48"
+    },
+    {
+      "bytes": 52318007,
+      "name": "last.pt",
+      "seed": 17,
+      "sha256": "f0af2eb994e000a512fff78e40fff821048564e768d7b1a6648a825482a1c440"
+    },
+    {
+      "bytes": 1572,
+      "name": "result.json",
+      "seed": 17,
+      "sha256": "91d774393c370f4a9eb366a2a4e57675af316068454c16246f289d98228d3b92"
+    },
+    {
+      "bytes": 6707,
+      "name": "validation_predictions.npz",
+      "seed": 17,
+      "sha256": "812dde93a13d341fe719ae28ed01a32db696d19e91f1ef82667cdacd153863b6"
+    },
+    {
+      "bytes": 52318007,
+      "name": "best.pt",
+      "seed": 29,
+      "sha256": "dbdfe0a8da046925b5891a13921bc477a475cace25d1f1d84da0ef98670ab34a"
+    },
+    {
+      "bytes": 6905,
+      "name": "history.json",
+      "seed": 29,
+      "sha256": "925788b8a603a2082e3f2151dea82906f0bf31cd922313013e7dbbb079fb3570"
+    },
+    {
+      "bytes": 52318007,
+      "name": "last.pt",
+      "seed": 29,
+      "sha256": "ceacdfec45ad785febbe17d7c8ff496e86112bc1e8dc107b3831be0b61e20ca5"
+    },
+    {
+      "bytes": 1574,
+      "name": "result.json",
+      "seed": 29,
+      "sha256": "5a1bb6fd8a023de64fe010142ac3d026713fc7c9f4b94547df1d7793a1067508"
+    },
+    {
+      "bytes": 6658,
+      "name": "validation_predictions.npz",
+      "seed": 29,
+      "sha256": "5e29d2796de4da9a4b7e510f235b4ae8d69a0a2f6f3edbd027955b3c323c81a5"
+    },
+    {
+      "bytes": 52318007,
+      "name": "best.pt",
+      "seed": 41,
+      "sha256": "61954e149c007b9a75a9b2010e8c749768a8c2386484b62e1ec2d05b3fc7c7c3"
+    },
+    {
+      "bytes": 8640,
+      "name": "history.json",
+      "seed": 41,
+      "sha256": "4b31dc03c1807678d185dcabc4fe9b1e4d4c96b0b8ac099d04e8093f5d5a5525"
+    },
+    {
+      "bytes": 52318007,
+      "name": "last.pt",
+      "seed": 41,
+      "sha256": "c517952231de8fbb39cef938332ff31b375521ff309302c1569d9b206ec1dd03"
+    },
+    {
+      "bytes": 1549,
+      "name": "result.json",
+      "seed": 41,
+      "sha256": "014cab2627a22924b9059513c616bd4974710f47322544f4129326b1e56a3f1b"
+    },
+    {
+      "bytes": 6692,
+      "name": "validation_predictions.npz",
+      "seed": 41,
+      "sha256": "34bfc39e6628ea0a92017a11649242c374a1c39f1fd3a50eb8da09e3db30fc66"
+    }
+  ],
+  "claim_boundary": "This is development evidence on controlled target breaks, natural-valid continuation hard negatives, and registered nuisance controls. It is not confirmatory natural-invalidation or recovery evidence.",
+  "confirmatory_opened": false,
+  "contract": {
+    "path": "/HELIOS/Robotics/SCVLA_Cover/SCVLA-update-two-stage-verifier/configs/updates/natural_error_v4_two_stage_model_v1.json",
+    "sha256": "b6abfbd7d301de2edb63690b4345e01a104532df8240eb94b8b79fedad3091fe"
+  },
+  "feature_sha256": "c64924957ff0d616e1a28948bc6aaa76848ac083d418d8390a60f0ac59a1fbdc",
+  "gate_passed": true,
+  "git_dirty": false,
+  "git_revision": "17d7924a2c2616ff053288ba559598e3c8f7063a",
+  "schema_version": 1,
+  "seed_results": [
+    {
+      "best_epoch": 6,
+      "epochs_completed": 16,
+      "metrics": {
+        "admission_macro_f1": 0.970223801748151,
+        "irrelevant_change_false_repair": 0.0,
+        "native_route_macro_f1": 0.9497643708125231,
+        "observable_invalidation_recall": 0.9032257795333862,
+        "observable_validity_macro_f1": 0.9950946232151396,
+        "proposal_false_positive_rate": 0.003367003367003367,
+        "proposal_recall": 0.9102167182662538
+      },
+      "peak_gpu_memory_bytes": 339286528,
+      "seconds": 3.800580130016897,
+      "seed": 17,
+      "thresholds": {
+        "admission": 0.52734375,
+        "invalidity": 0.984375,
+        "proposal": 0.9921875
+      }
+    },
+    {
+      "best_epoch": 6,
+      "epochs_completed": 16,
+      "metrics": {
+        "admission_macro_f1": 0.9768407346930064,
+        "irrelevant_change_false_repair": 0.0,
+        "native_route_macro_f1": 0.9485538852498214,
+        "observable_invalidation_recall": 0.9032257795333862,
+        "observable_validity_macro_f1": 0.9950946232151396,
+        "proposal_false_positive_rate": 0.003367003367003367,
+        "proposal_recall": 0.9040247678018576
+      },
+      "peak_gpu_memory_bytes": 339286528,
+      "seconds": 3.9219521440099925,
+      "seed": 29,
+      "thresholds": {
+        "admission": 0.85546875,
+        "invalidity": 0.96875,
+        "proposal": 0.98828125
+      }
+    },
+    {
+      "best_epoch": 10,
+      "epochs_completed": 20,
+      "metrics": {
+        "admission_macro_f1": 0.9735322682205787,
+        "irrelevant_change_false_repair": 0.0,
+        "native_route_macro_f1": 0.9650876417481715,
+        "observable_invalidation_recall": 0.9354838728904724,
+        "observable_validity_macro_f1": 1.0,
+        "proposal_false_positive_rate": 0.003367003367003367,
+        "proposal_recall": 0.934984520123839
+      },
+      "peak_gpu_memory_bytes": 339286528,
+      "seconds": 4.589294224017067,
+      "seed": 41,
+      "thresholds": {
+        "admission": 0.75,
+        "invalidity": 0.71484375,
+        "proposal": 1.0
+      }
+    }
+  ],
+  "seeds": [
+    17,
+    29,
+    41
+  ],
+  "status": "PASSED-V4-TWO-STAGE-DEVELOPMENT-GATE"
+}