状态:B1_A0_PERMANENT_NEGATIVE_TERMINAL / R2_F0_SYNTHETIC_REDUCER_PASS / F1_SUPERVISION_FRONTDOOR_SATISFIED / AG_ST_DIRECT_TEACHER_TO_AG_REAL_SEAM_PASS / F1_STUDENT_ATTEMPT17_FAIL_NO_PROMOTION / AG_R2_SUPERTEACHER_TO_AG_FINAL_V2_SEAM_PASS / AG_R2_CROSS_SENSOR_CALIBRATION_CONTROL_R0_AND_R1_FAIL_CLOSED_CONSUMED / R1_INDEPENDENT_REPLAY_CONFIRMED_PRODUCER_FAILURE / CORRECTION_GAIN_LOPO_FAIL_STOP / ANGULAR_BOUNDARY_FAIL_CLOSED_SAFE_BUT_TASK_INERT / SUPPORT_VALIDITY_FAIL_OPEN / OBSTACLE_RGB_INTERACTION_FAIL_STOP / POSE_ANALYTIC_FAIL_STOP / CURRENT_OBSTACLE_TASK_ROUTE_CLOSED / SCIENTIFIC_NOT_RUN / CONFIRMATION_OUTCOMES_UNOPENED
本目录包含 BlindAssist Assistive Geometry B0 的冻结合同、shape/export、metadata roster、 可恢复媒体物化与 label-blind integrity 工具:
2026-08-12 当前交付分两层:run_ag_st_direct_teacher_to_ag_real_seam.py 先把 source-anchored
SuperTeacher factors 直接接入冻结 adapter/reducer;随后 learned metric/factor recipe 加冻结的
session-height scale anchor,经 FactorTensorAdapterV2 在 checkpoint-unseen sitting_rpy 上完成最终
12-frame seam。最终为 12/12 valid、12/12 deterministic、11/11 gates、CLEAR=18 / UNKNOWN=90;
推理 targets_loaded=false,UNKNOWN 未转 negative,也没有任意 baseline 胜负门。Attempt17 与无锚
walking_xyz 负结果保持冻结;当前只支持 research-pipeline mechanics,不是跨传感器或移动部署结论。
ag_r2_cross_sensor_confirmation/ 的 R0/R1 calibration-control one-shot 均已 fail closed consumed。R1 producer
读取 2 个 YAML 后以 F2_R1_KALIBR_ROSTOPIC 停止;matrix discovery 与 target-match count 保持
null/UNKNOWN。producer-free validator 先消费独占 replay receipt,再一次性复现同一失败并封存完整 chain;
离线验证 PASS。session archive、checkpoint、source truth、factor scoring 和 Confirmation 均未运行,R1 不得
rerun/resume/replace;该 Formal calibration 路径没有 active successor。当前算法 successor 仅为下述隔离的
factor-wise no-regret Development。
validate_b0_task_contract.py:对 B0 JSON 合同执行 fail-closed schema/不变量检查;test_validate_b0_task_contract.py:覆盖有效合同和关键违规合同;preflight_depthart_rectangular_shape.py:用真实 DepthART-S metric checkpoint 验证1×3×608×448PyTorch shape、dynamic camera prompt 与 ONNX graph/checker。audit_b0_data_capability.py:只读 master ledger,区分结构候选与研究角色 authority;plan_b0_arkitscenes_rosters.py:按冻结 identity 排除快照生成 visit/video-disjoint16/8/8roster;preflight_b0_arkitscenes_assets.py:对五类冻结源资产执行 label-blind HEAD preflight;download_b0_arkitscenes_assets.py:历史 earliest-common materializer;其 Attempt 3 因 pose 覆盖失败,禁止复用;audit_b0_arkitscenes_pose_coverage.py:重算冻结窗口与 trajectory 时间域关系;download_b0_arkitscenes_pose_covered_assets.py:可恢复地物化 trajectory 域内连续 300 帧;audit_b0_arkitscenes_integrity.py:逐文件 SHA、实际图像解码、内参和 pose 包络审计。arkitscenes_truth_reader.py:按官方 inverse trajectory convention 将注册模态旋转到逐帧 upright metric frame,并派生 gravity ground、三通道 body-swept clearance 与 UNKNOWN;materialize_b0_arkitscenes_upsampling_train.py:仅物化冻结 TRAIN role 的 exact-timestamp AppleDepth/FARO/RGB/confidence/intrinsics 对照;validate_b0_arkitscenes_truth_reader.py:运行 TRAIN-only scale/registration/ground/clearance 双层门,并写入逐帧 evidence receipt。validate_b1_training_protocol.py:冻结并校验 B1 target/loss/confidence、A0–A4 additive arms、 optimizer、数据角色和 implementation-before-training 防火墙。audit_b1_orientation_geometry.py:只读 pose/identity,审计 full-FOV portrait/landscape frame capacity,不打开 image/depth/task outcome;validate_b1_training_protocol_attempt_02.py:校验当前 dual-orientation overlay、orientation buckets、full-FOV K 传播、Development split 与 portrait claim ceiling。materialize_b1_train_targets.py:只为冻结 TRAIN identity 写入 compact source-upright target, 不物化 prediction-dependent confidence truth;validate_b1_train_targets.py:逐 SHA 和 NPZ 语义验证 4,800 个 TRAIN target,并 fail-closed 检查 UNKNOWN、方向、K、ground、clearance 与 occupancy;assistive_geometry_model.py:复用 DepthART-S shared decoder feature,提供 Ground、Clearance、 Occupancy、Confidence heads 与 A0–A4 frozen losses;depthart_training_scan.py:训练时直接进入部署包内显式 custom Autograd Function,绕过没有 Autograd-key registration 的外层 inference/export dispatcher;smoke_b1_dual_orientation_training_model.py:用冻结 checkpoint 在 portrait/landscape 全尺寸上 验证 forward、loss、encoder/head backward 与 SelectiveScan dispatch boundary。assistive_geometry_training.py:提供 deterministic parent-balanced/orientation-bucket loader、 same-orientation carry、augmentation、A0 cosine scheduler 与 collate 合同;smoke_b1_a0_train_execution.py:以真实 TRAIN 数据执行受限 optimizer step,写出并精确恢复 model/optimizer/scheduler/scaler/sampler/RNG checkpoint;smoke_b1_a0_train_execution_attempt_02.py:保留 Attempt 1 RNG-device negative 后,将 checkpoint 首次加载固定在 CPU;必须以-m scripts.research.assistive_geometry.smoke_b1_a0_train_execution_attempt_02运行。train_b1_a0_formal.py:运行冻结的 A0 TRAIN-only 性能 pilot 与三 seed 正式训练;发布 guarded progress,按 epoch 原子保存可恢复状态,并保留5/10/15/20checkpoint。train_b1_additive_arm.py:A1–A4 共用的 outcome-blind 训练 mechanics;所有 arm 从同一 DepthART 初始化独立训练,只逐项开放冻结 head/loss,等待 A0 Development 结果后另立协议激活。evaluate_b1_a0_synthetic.py:验证三 seed × 四 retained checkpoint 的 bytes/SHA、内部状态、 协议与步数完整性,并计算 pooled、九格、parent 与 orientation task metrics;不选择 seed。run_b1_a0_evaluation_dry_run.py:只用合成 fixture 演练通过路径与 checkpoint 缺失、协议漂移、 缺 horizon、全局零分母、coverage 塌缩、best-seed 企图等失败终态,并生成 JSON、短报告和 failure-adjacent log。materialize_b1_development_targets.py:只有三 seed 正式结果完整时才物化冻结的四个DEVELOPMENT_SELECTIONparent;Calibration 与 Confirmation fail closed。observe_b1_a0_development.py:用各 seed epoch-20 dense-depth checkpoint 与冻结 gravity/geometry 后处理生成独立 truth/pred validity 和三态 observation;不读取未训练 task heads。evaluate_b1_a0_development.py:执行三 seed 无选择聚合,并同时检查 coverage、ground、clearance、 false-clear/false-block、temporal delta 与 geometry transition 门。analyze_b1_a0_failure_anatomy.py:只读已消费、SHA-bound 的 A0 Development observations,分解 tri-state 分布、clearance residual、false-block 阈值一致性、transition failure 和跨 seed failure-mask similarity;结果永久NOT_ELIGIBLE_FOR_PROMOTION。geometry_r2_reducer.py:F0 冻结的零参数 interval reducer;只有 positive lower-bound evidence、 guaranteed lateral overlap 与 horizon 内 upper-bound distance 同时成立才输出 occupied,歧义或缺失为 UNKNOWN。fixtures/geometry_r2_f0_cases.json:23 个 SHA-bound analytic factor case,覆盖 depth/scale、support、 boundary、orientation、uncertainty monotonicity、反 A0 场景和 final-task shortcut 负控。run_geometry_r2_f0_canary.py:校验协议/实现/fixture SHA 后执行 10 项 conjunctive F0 kill gate, 只写新 evidence root;不训练、不读真实数据、不自动授予 F1。validate_geometry_r2_f1_protocol.py:只做 F1-P schema、DCA capability、loss/checkpoint、Kill Gate、 successor 与 execution-authority 的静态 SHA/语义校验,并断言 F1 trainer/model/materializer 路径不存在。test_validate_geometry_r2_f1_protocol.py:9 个 mutation tests,覆盖执行扩权、final-task shortcut、 UNKNOWN-as-negative、能力计数漂移、aggregate checkpoint loss、reducer rescue 与 parent-role overlap。audit_geometry_r2_f1_adapter_gap.py:静态核对 byte-frozen F1 factor schema 与 F0 reducer input, 显式列出 scale/support uncertainty、dense→obstacle list 和 camera/frame binding 的 17 个 adapter 操作; 不实现 adapter、不运行 reducer/canary、不授予执行权限。test_audit_geometry_r2_f1_adapter_gap.py:7 个静态/mutation tests,验证缺 adapter 必须 fail closed, 完整静态合同也最多到CANARY_NOT_RUN,并拒绝 learned-graph、可训练参数或 execution 扩权。validate_geometry_r2_f1_adapter_protocol.py:验证14/14F1 field consumers、全部 F0 field producers、17 个 operation、8-case fixture、A01–A10、authority/successor 与 exact SHA bindings; 不实现或执行 adapter。test_validate_geometry_r2_f1_adapter_protocol.py:13 个 mutation tests,拒绝字段/operation 缺失、 task shortcut、receipt/support/missing-depth fail-open、uncertainty strengthening、扩权与 binding drift。factor_tensor_adapter.py:按冻结 17-operation 合同将完整 F1 factor tensors 确定性转换为 F0 frame; 零参数、learned-graph 外、无 task outcome,receipt 或局部证据无效时 fail closed。test_factor_tensor_adapter.py:10 个 focused tests,覆盖 8-case fixture、独立 UNKNOWN 语义、 orientation parity、component split/merge、uncertainty monotonicity 与 final-task shortcut 拒绝。run_factor_tensor_adapter_canary.py:只接受 SHA-bound implementation lock,用两个独立进程重放 每个 frozen case,并只写未存在的版本化 synthetic evidence root。factor_tensor_adapter_v2.py:把 global metric-scale sigma 与 local depth-shape sigma 分开传递, 保留旧 adapter 的 fail-closed reducer ABI,不让任一 uncertainty 分量被另一分量覆盖。ag_r2_cross_sensor_confirmation/:冻结 ETH3D opaque binding/preflight、12+12 roster、RGB+K model-only prediction、两次 seal/reload firewall、session source geometry、27-gate factor-only metrics、 exclusive evidence 与独立重算 validator;包导入不访问 archive/checkpoint,也不创建 evidence root。run_ag_r2_cross_sensor_factor_accuracy_confirmation.py:只接受另行冻结的 exact one-shot execution lock;当前 implementation lock 的全部 execution authority 为 false,不能据此启动真实执行。ag_r2_cross_sensor_confirmation/validate_implementation_lock.py:只读 tracked control files,复核 历史 executor implementation predecessor/code/test byte+SHA、45-test receipt 与零 payload access。ag_r2_cross_sensor_confirmation/control_format.py:只接受 exact camera-node-scoped Kalibr camchain YAML nestedT_cam_imu,拒绝 inline matrix、非正交旋转、重复 path 与无法唯一绑定的 camera node。ag_r2_cross_sensor_confirmation/calibration_control.py:未来只在独立 hash-bound one-shot 下消费独立 control root,并且只有 camera-IMU calibration archive 一个 payload 输入;session archives 与模型没有 API。ag_r2_cross_sensor_confirmation/validate_calibration_control.py:不导入 producer/source adapter/control parser,独立重哈希 calibration archive、枚举 YAML、重算 matrix selection 与 evidence manifest。ag_r2_cross_sensor_confirmation/control_format_r1.py:纯解析每个 Kalibr camera node 的同节点rostopic + T_cam_imu,不以cam0/cam1顺序推断目标相机。ag_r2_cross_sensor_confirmation/calibration_control_r1.py:未来只在另行 hash-bound R1 one-shot 下按/uvc_camera/cam_2namespace 唯一匹配;失败时保存已知/未知计数、摘要与零 first/best selection receipt。ag_r2_cross_sensor_confirmation/validate_calibration_control_r1.py:不导入 R1 producer/source/parser, 先写独占 start receipt,再对 PASS 或任意 fail-closed evidence 独立重放一次 archive、rostopic namespace selection 与完整计数;完成后可纯本地验签且不会重开 archive。ag_r2_cross_sensor_confirmation/validate_calibration_control_r1_repair_lock.py:复核 R0 sealed audit、 preserved legacy runtime、official selection evidence、R1 amendment、9 个 implementation binding、 69-test receipt 与零 archive access。ag_r2_cross_sensor_confirmation/validate_depthart_source_manifest.py:独立复核 29 个实际可导入的 metric/selective-scan Python 文件及 bytes/SHA,不加载 checkpoint 或模型。ag_r2_cross_sensor_confirmation/validate_repair_implementation_lock.py:复核 schema v2、official control、 source manifest、18 个 implementation binding、51-test receipt、零真实 payload 与不可自授权 successor。train_ag_r2_multisource_metric_depth_student.py与train_ag_r2_f1_attempt18_consumed_cross_domain_adaptation.py: 只从分级 factor supervision 学 metric depth 与 support/obstacle/boundary/validity,不训练 final task state。calibrate_ag_r2_session_metric_scale_anchor.py:只在已消费 factor depth 上从六个预声明物理候选选择 camera-height quantile anchor;选择过程不读取 CLEAR/OCCUPIED/UNKNOWN 或 reducer output。run_ag_r2_hybrid_factor_student_to_ag_seam.py:组合冻结 metric/factor checkpoints、可选 session anchor、 分解 uncertainty 与 deterministic adapter/reducer;factor-only inference 可显式禁止加载目标。materialize_ag_r2_tum_sitting_rpy_final_confirmation_labels.py与run_ag_r2_tum_walking_xyz_final_v2_seam.py:前者物化 source-native/geometry-anchored 12-frame labels, 后者是参数化的一次性最终 seam runner;文件名保留 walking_xyz 历史,但 final parent/receipt 由参数锁定。run_ag_st_stage0a.py:独立WILD_LABfactor-only runner;source role 参数化,可从 B0 raw manifest 或 scoped-media manifest 恢复 RGB/K/pose/partial depth,以确定性连续块隐藏 reference,执行 source-anchored MapAnything,并输出 baseline、校准前后 depth residual 与 confidence risk-coverage。 它不读取 clearance/occupancy,不物化 canonical label,不改变 F1 execution authority。test_run_ag_st_stage0a.py:7 个 focused tests,覆盖隐藏 reference 不回流 Teacher input、mask 可重放、 source-role 选择、scoped trajectory receipt、observed-anchor scale 与 confidence selective metrics。build_ag_st_factor_labels.py:把 Stage 0A source-first metric depth、anchor residual 与 multi-view reprojection residual 变成 A/B/C/UNKNOWN 分级 pseudo-label;输出 per-factor validity/provenance、 support-plane、physical-boundary distance 和 uncertainty proxy,可直接供 masked student 读取。test_build_ag_st_factor_labels.py:8 个 focused tests,覆盖 hidden-reference 隔离、reprojection、 source priority、uncertainty、派生 provenance、连续 80° 斜面负控与真实 depth-step 正控。build_ag_st_multiteacher_factor_labels.py:保留 source-native depth 与 MapAnything primary geometry, 只用独立 Depth Anything V2 的 source-anchored 分歧重标 quality/uncertainty/UNKNOWN;分歧不会被当成 truth,也不要求第二 Teacher 的全局误差低于主 Teacher。test_build_ag_st_multiteacher_factor_labels.py:3 个 CPU focused tests,覆盖 observed-anchor scale、 pair disagreement 对称性、source tier 保留及强分歧转 UNKNOWN。ag_st_tum_rgbd.py:从 7 个 TUM RGB-D sequence 的目录或 TGZ 精确恢复 RGB、registered depth、K, 并把官方 groundtruth trajectory 插值为 RGB 时刻的 camera-to-world pose。run_ag_st_tum_cross_source.py:在 4 FIT / 3 held-out TUM sequence 上原样复用 R0 multi-Teacher quality threshold;只在推理后打开隐藏 source depth,PASS 后物化 depth/uncertainty/UNKNOWN, 未验证 gravity 前 support/boundary 恒为 UNKNOWN。test_ag_st_tum_rgbd.py与test_run_ag_st_tum_cross_source.py:6 个 focused tests,覆盖 cohort disjointness、RGB-depth pairing、pose interpolation、真实 payload receipts、跨源门与 source-first 标签不变量。plan_ag_st_tum_third_teacher_cohort.py与run_ag_st_tum_third_teacher.py:冻结另 7 个未引用 TUM parent,在 4 FIT 上选择 DepthART union/consensus witness,随后一次性评 3 held-out;若第三 Teacher 无 no-regret 增益则保留原两教师配方,不回调阈值。ag_st_depthart_teacher.py与test_run_ag_st_tum_third_teacher.py:只读 RGB+K 的冻结 DepthART-S metric Teacher,以及 4 个 focused gate/role tests;DepthART 只作独立 witness,不要求击败主 Teacher。diagnose_ag_st_tum_gravity.py:把 TUM accelerometer 与 mocap pose 联合到 24 个 proper signed-axis candidates,验证 Freiburg1/2 的 IMU→RGB optical 映射与 world+Zgravity;无 accelerometer 的 Freiburg3/Xtion 不作推断。materialize_ag_st_tum_gravity_factors.py与对应 test:对 gravity-eligible TUM 标签物化 continuous normal/support/boundary/obstacle evidence;不具 gravity 的 parent fail-closed 为 UNKNOWN。dominant gravity-aligned plane 仍可能是桌面,不冒充 walkable-ground truth。diagnose_ag_st_tum_support_identity.py:把 source-native depth 通过 pose 投到 parent world frame, 恢复跨帧持续水平高度模式,并用更低持续面识别 per-frame dominant plane 的桌面/高架误标。materialize_ag_st_tum_support_identity_factors.py与validate_ag_st_tum_support_identity_factors.py: 用通过 identity 的 sequence height 重物化 TUM support/boundary pseudo-label,并验证 UNKNOWN、gravity alignment、camera-height binding 与旧/新 support-positive correction。run_ag_st_analytic_support_boundary_canary.py:解析 floor+dominant-table exact renderer;同时检验 support identity、table false-positive 和 2px level-change boundary,全部为 deterministic CPU mechanics。run_ag_st_icl_mesh_support_identity.py:用 ICL-NUIM 官方 living-room OBJ 的room_floor与 global poses 检验最低持续高度;只接受 upward-facing exact mesh surfaces,过低相机或稀疏视角保持 UNKNOWN。materialize_ag_st_sequence_identity_labels.py与validate_ag_st_sequence_identity_labels.py:把 sequence identity 推广到 16-parent multi-Teacher TRAIN 标签;parent 至少 2/3 帧 camera-height plausible, 每帧仍守0.45–2.20 m,否则 factor denominator 为零。train_ag_st_masked_student.py:冻结 DepthART-S 或 MobileNetV3 encoder,只训练小型 dense factor head; 可直接拼接多个互不重叠的 Stage0A/label batch,每个 orientation 留 1 selection + 1 canary parent; 也可在已有独立 confirmation 时把全部 consumed parents 纳入 fit。支持 multifactor、depth/support-only、 metric-precision + calibrated-support 与 boundary-only 目标;train-only scalar temperature/bias 可折叠回 support head。也支持 DepthART 四层 decoder pyramid、dilated pyramid head 与显式 base-depth guidance。 A/B/C tier weights 保留,UNKNOWN 权重恒为零,不调用 reducer 或 task outcome。test_train_ag_st_masked_student.py:12 个 focused tests,覆盖 16/32-parent split 重放、非对称 orientation roster、零残差与 identity-gate 初始化、multi-scale/base-depth head、objective UNKNOWN/NaN 隔离、tier 权重和 scalar calibration。train_ag_st_bonn_anchored_student.py:把三批 40-parent ARKit factor labels 与冻结 Bonn FIT 的 registered source depth 合并;Bonn 只提供 A-tier depth,其他 factor 全 UNKNOWN。使用 5x Bonn 重放形成 domain-balanced optimizer visits,并训练初始回退 DepthART base 的identity_sigmoidcorrection gate。test_train_ag_st_bonn_anchored_student.py:2 个 focused tests,锁定 8/8 cohort disjointness、排除旧 fixed-8,并验证 Bonn adapter 只开放 source depth、其余 factor 分母恒为零。train_ag_st_no_regret_selector.py:冻结 base 与 correction expert,只训练/评价 base-vs-correction selective router,并报告 perfect signed-advantage oracle 的安全 coverage/headroom。threshold admission 要求每个 calibration parent 的 MAE 与>0.10 merror 都 no-regret,且至少一半 parent 有非零 correction coverage;没有 admissible threshold 时确定性回退 base。test_train_ag_st_no_regret_selector.py:9 个 focused tests,覆盖 deterministic split、可扩展 TUM calibration parent 数、oracle headroom、 fallback,以及“macro 改善但单 parent 受伤”必须拒绝的回归测试。run_ag_factorwise_no_regret_oracle_parent_gate_canary.py:重放冻结 prior/expert/selector,并显式加入 perfect signed-advantage oracle;逐 parent 同时约束 MAE 与>0.10 merror,至少一半 parent 必须有 非零 coverage,R21 boundary 只作 SHA/结果重放。train_ag_st_frame_advantage_lcb_router.py:冻结 pixel selector/correction expert,以 neural quantile ensemble 和跨 parent kNN lower bound 形成只会 veto 的 frame gate;可把 checkpoint fallback 与显式 Development pixel candidate threshold 分开记账,也可限制为 TUM-only 并纳入已消费 evaluation parent。evaluate_ag_st_frame_advantage_lcb_router_tum.py:对冻结 frame gate 执行 parent-disjoint TUM 评估, 检查 fit/calibration firewall,并报告 neural/kNN 分数、逐 frame 真值 advantage 与严格 parent gate; consumed 诊断不能重新包装成 fresh evidence。- 对应 8 个 focused tests 覆盖 gate、fallback、pinball asymmetry、veto-only、kNN parent exclusion 与 selector threshold provenance;当前 TUM14 结果为 runtime-observability fail-stop,不授权继续用同一 observable 重训。
run_ag_runtime_correction_gain_observability_canary.py:在 14 个 consumed TUM parent 上以 leave-one-parent-out 检验 flip equivariance、temporal reprojection 与合取 observable;三候选均为全 fallback,按停止条件关闭当前 correction expert/router,未打开 fresh3。run_ag_angular_boundary_body_swept_task_canary.py:固定 ICL source-exact depth/support/obstacle,仅替换 R20/R21 boundary probability;R21 虽提高 task-reference agreement,但 naive conjunction 会产生 unsupported CLEAR,不能晋级。run_ag_angular_boundary_fail_closed_task_canary.py:把 boundary absence 改为 UNKNOWN,只允许通过 component-edge localization sigma 增加不确定性;危险放行归零,但 R20/R21 task state 完全相同,故 boundary-to-task mapping 停止。对应 focused tests 锁定 one-sided evidence 与 UNKNOWN 语义。run_ag_positive_obstacle_support_task_effect_audit.py:固定 source-exact depth/boundary,prediction-first 执行 learned support、learned obstacle 与合取替换臂。结果显示 support 可越权打开 reference-UNKNOWN, obstacle 则安全但丢失23/33known cells 且不产生 OCCUPIED;下一步只允许 obstacle 三态校准,support 保持 veto-only。对应 focused tests 锁定单 factor 替换和 completion 保留。run_ag_obstacle_evidence_tristate_calibration_canary.py:冻结 R21 obstacle logit,在六个 checkpoint-held ARKit/TUM consumed parent 上完成 RGB+K prediction 后才打开 source-valid obstacle truth;嵌套前的首轮 leave-one-parent-out 双阈值校准为0/6可评 fold,因此 current scalar score fail-stop,禁止 reducer seam。run_ag_obstacle_selective_interaction_head_canary.py:只组合 frozen obstacle/support/boundary/depth 与 image-row observable;inner calibration parent 始终由排除自身的模型预测,outer parent 也只评一次。 嵌套结果仍为0/6可评 fold,关闭同一 RGB factor observable 家族上的后续 selector。run_ag_depth_pose_analytic_obstacle_canary.py:只从 RGB+K 预测 DepthART metric depth,并把camera_to_world作为 runtime-equivalent VIO/IMU pose-gravity;跨帧恢复最低水平高度后用冻结 factor geometry 解析 obstacle。3/6parent 几何完整但0/6fold 获得安全双阈值,当前 obstacle task route fail-stop;对应 tests 锁定 persistent mode、world-z pose 变换与最小帧数。evaluate_ag_st_student_checkpoint.py:在 parent-disjoint cohort 上零样本评估冻结 checkpoint;默认 fresh 模式,也可显式签署consumed_development_comparison,防止把已看过的 cohort 再包装成新证据。test_evaluate_ag_st_student_checkpoint.py:6 个 focused tests,覆盖 objective-specific core factors、 all-consumed fit parent firewall、macro improvement、diagnostic split 与 consumed-mode fresh-claim 禁用。evaluate_ag_st_student_bonn_depth.py:默认固定 Bonn RGB-D Dynamic 8 sequence × 3 帧,也可读取 SHA-frozen cohort manifest;以 used-set 形成唯一 RGB-depth 对。模型只读 RGB+K,推理后才打开 registered source depth,比较 initialized DepthART 与 frozen student 的 parent-macro MAE/>0.10 merror。输出仅为 cross-dataset depth Development,不评价 support/boundary/task,也不作许可结论。test_evaluate_ag_st_student_bonn_depth.py:6 个 focused tests,覆盖固定 cohort、唯一 pairing、缺失 depth member、uint16/5000、parent-macro 与非单调 index fail-closed。export_assistive_geometry_onnx.py:把未来选定 checkpoint 导出为 portrait/landscape 静态 ONNX, 保留五个 raw GeometryState tensor 与 host camera prompts;gravity/UNKNOWN 后处理不塞入图内。evaluate_teacher_complementarity.py:在未来另行授权的 truth-bound cohort 上比较 metric 与 temporal geometry 教师的单体、oracle、独占正确 parent、分歧错误浓度和时序优势;任一 kill gate 失败即停止 C1。temporal_geometry_ablation.py:为未来 phase D 提供同一 8-frame GeometryState 下的因果 GRU/TCN/diagonal-SSM 候选,统一 future-clearance/TTC/compute-gate 输出和 50k 参数上限;不决定最终三态。run_hypothesis_canary_lite.py:只用 deterministic synthetic CPU geometry 审查 censored survival、profile-conditioned clearance、widest-path bottleneck 与 one-sided conformal uncertainty 的数学不变量和反例;不读取任何数据 role outcome、模型或 checkpoint。
大体积输出只允许写入 artifacts.local/datasets/、artifacts.local/experiments/、
artifacts.local/evidence/hftf/ 或 artifacts.local/evidence/assistive-geometry/。
roster 选择只依据冻结 metadata/hash,不读取模型输出或 task outcome。当前合同和结果真源位于
docs/research/assistive-geometry/。
本模块的 B1-A0 及 A1–A4 已永久关闭;teacher 只有未激活的历史 C0 complementarity mechanics,
历史 C0 teacher 路线当前不读取 teacher output,也不授权 C1、QNN/HTP、默认 App、产品或 safety。
独立 AG-ST 已在 WILD_LAB 中完成 MapAnything Stage 0A、factor-label 物化和冻结 DepthART encoder
masked-head 训练。两条独立的 train-parent 到 fresh-parent 零样本链均复现 depth/support 学习信号;
obstacle 一条改善、一条退化,只是 diagnostic;boundary 在 multifactor、boundary-only 和两条 fresh
zero-shot 评价中均未通过。后续 combined-32 depth/support-only checkpoint 又在 8 个全新 parent 上取得
depth MAE 85.8% 和 support BCE 47.5% 的 parent-macro 相对下降;累计 44 个互异 ARKitScenes parent
被消费后,combined-40 precision checkpoint 在最终 CONFIRMATION 8 上取得 depth MAE 85.4%、support
BCE 83.2% 的相对下降。累计 52 个互异 ARKitScenes parent 已消费,这仍不是跨数据源泛化。
后续 multi-scale head 在已消费 confirmation 上结果混合且不晋级;两个 ARKit-trained depth residual head
在 Bonn registered depth 上又分别把 DepthART MAE 0.2533 m 恶化到 1.0146 / 1.1755 m,均为
0/8 parent 改善。因此当前 depth transfer 明确不支持,下一算法必须保留域外 identity fallback 或加入
source-diverse metric anchors,不能继续用同域 head-capacity scaling 代替跨数据源监督。
后续 mixed-domain identity-gated student 已把新 Bonn EVAL MAE 收回到 0.2713 m,大幅消除上述
catastrophic collapse;但仍差于冻结 DepthART baseline 0.2517 m,且只有 1/8 parent 改善,故不晋级。
当前 factor-wise Development 先比较 perfect signed-advantage oracle 与现有 selector,回答 correction 是否
存在安全覆盖;只有 oracle 有 headroom 而 selector 失败时,才训练 one-sided advantage-LCB router。selector
准入已从 domain/macro no-regret 收紧到逐 parent 双指标 no-regret 与非集中 coverage,不在已看 EVAL 调 threshold。
support/boundary 仍是 conservative pseudo-label、
sigma 是 proxy,不产生完整 truth、正式 F1、产品或 safety authority。
时序模块同样只有未激活 mechanics;没有新 temporal cohort、训练、任务收益或设备性能 authority。
移动导出受历史 M0 质量先于性能协议约束;现有 DepthART D1 cohort 不得复用为 Assistive Geometry
选模证据。新 R2 已完成 F0 reducer mechanics、F1-P schema/loss/selection/Kill Gate、正式 source-native
supervision frontdoor 和 deterministic FactorTensorAdapter;adapter synthetic evidence 仍为
8/8 cases、10/10 gates、8/8 双进程 replay 与 7/7 sigma mutation PASS。其后 direct SuperTeacher
real seam 又在 12 帧上完成全部 tensor/adapter/reducer receipt。该结果证明 reference factor mechanics,
不证明 learned student、移动推理或真实任务收益。
UNKNOWN 不得当作负例;synthetic shape 与 benchmark geometry 不得冒充任务质量。
合同违规、checkpoint/shape 不匹配、非 finite 输出、camera prompt drift 或 ONNX checker
失败均立即 fail closed。当前 roster、source integrity、truth reader 与 registration 已关闭,
且 B1 target/loss/confidence、dual-orientation overlay、4,800-frame target cache 与模型/loss
implementation lock 与 A0 execution lock 已关闭,三个正式 seed 均完成。合成 evaluator dry-run
与真实 Development Selection 评价均已执行;A0 虽通过前门,但 clearance MAE、false-block 和
geometry transition agreement 均为 0/3 seed 通过,终态为
B1_A0_DEVELOPMENT_EVALUATION_FAIL_TASK_GATES。旧 A1 条件 successor 未激活,A1–A4、teacher、
移动和时序执行继续禁止。只读 failure anatomy 已完成且不可晋级;Selection 已消费且不得复用,
Calibration 与 Confirmation 保持封存。R2 F0 已签署
BLINDASSIST_ASSISTIVE_GEOMETRY_R2_F0_SYNTHETIC_FACTOR_GEOMETRY_CANARY_PASS;历史 F1-P frontdoor 和
adapter blocker 后续均已关闭。当前冻结终态是
AG_R2_SUPERTEACHER_TO_AG_FINAL_V2_SEAM_PASS / AG_R2_CROSS_SENSOR_FACTOR_ACCURACY_CONFIRMATION_F2_FROZEN / CONFIRMATION_OUTCOMES_UNOPENED。
唯一后续是 BLINDASSIST_ASSISTIVE_GEOMETRY_R2_CROSS_SENSOR_FACTOR_ACCURACY_CONFIRMATION_EXECUTOR_IMPLEMENTATION_LOCK:
只实现并以 synthetic/metadata fixture 验证 ETH3D source adapter、deterministic roster、prediction-before-truth
firewall、factor-only scorer 和独立 validator。当前不得枚举/解压七个 opaque archive、运行模型或
Confirmation,也不得用 reducer state 训练、重跑已消费 fresh canary,或把 F2 lock 写成 HTP、默认 App、
产品与 safety 结论。
验证:
E:\codex-tools\bin\blindassist-python.cmd -m unittest discover `
-s scripts/research/assistive_geometry -p "test_*.py"