experiment_registry.csv

experiment_id,date,claim_id,stage,dataset,split_unit,initializer,candidate_generator,selector,optimizer,contact_loss,config_path,checkpoint,seeds,gpu,runtime,status,primary_metrics,result_summary,log_path,notes
GATE-A,2026-08-24,C00,"oracle headroom pilot","CHI3D 50 main plus 10 noise events","event; one primary view","cached BEV SMPL-X initialization","GT/oracle contact","oracle","frozen paired optimizer","ProsePose hhcmap weight 0.1","configs/gate_a_oracle.yaml","n/a","3 repeats in noise pilot; deterministic main","RTX 5090","mean 56.74s per main run","completed; full gate failed penetration safety","paired interaction gain pose and penetration non-inferiority","Contact headroom supported; unconditional safety unsupported","outputs/gate_a_main_50/analysis.json","Original gate_a_pass=false is frozen; proceed to Gate-A2 rather than changing margins"
GATE-A-SMOKE-HHCMAP,2026-08-24,C00,"contact intervention and convergence smoke","CHI3D val record 5","single event and primary view","cached BEV SMPL-X initialization","O0 none; O2 all GT edges","oracle","official HHCS with early stopping","ProsePose hhcmap weight 0.1","configs/gate_a_oracle.yaml","n/a","seed 42","RTX 5090","44.5s O0; 45.6s O2","completed; scientific smoke only","GT-edge distance MPJPE PA-MPJPE PVE relative translation penetration reprojection","O0/O2 GT-edge 78.11/41.84mm; rel-trans 160.33/105.70mm; pen-max 71.50/65.58mm; reproj 6.206/6.121px; MPJPE 77.27/78.08mm","outputs/gate_a_smoke/eval_record_0005_full.json","heuristic_02_chi3d hhcdistmin was invalidated because it ignores cmap; hhcmap is the frozen intervention"
GATE-A-NOISE-V1,2026-08-24,C00,"repeated optimizer noise pilot","CHI3D val 10 primary events","event","cached BEV SMPL-X initialization","O0 none; O2 all GT edges","oracle","official HHCS max 1000x2 plus early stopping","ProsePose hhcmap weight 0.1","configs/gate_a_oracle.yaml","n/a","42;43;44","RTX 5090","mean 87.78s per run under 3-way concurrency","completed 60/60; advance to main Gate-A","paired metric noise and frozen neutral bands","GT-edge O2-O0 -88.67mm CI[-126.55,-50.59]; rel-trans -91.01mm CI[-153.31,-31.66]; repeat range zero","outputs/gate_a_noise_pilot/analysis.json","Contact F1 CI crosses zero; advancement only, not full Gate-A pass"
GATE-A-MAIN-50,2026-08-24,C00,"held-out main oracle headroom","CHI3D val 50 disjoint action-stratified events","event; one primary view","cached BEV SMPL-X initialization","O0 none; O1 deterministic first GT edge; O2 all GT edges","oracle","official HHCS max 1000x2 plus early stopping","ProsePose hhcmap weight 0.1","configs/gate_a_oracle.yaml","n/a","42; deterministic path","RTX 5090","mean 56.74s per run","completed 150/150; gate_a_pass=false","paired contact effect and all frozen safety noninferiority metrics","O2-O0 GT-edge -93.23mm CI[-113.30,-74.50]; rel-trans -58.10mm; pen-max +13.24mm CI[7.41,19.53]; pen-mean +3.55mm CI[1.53,5.62]","outputs/gate_a_main_50/analysis.json","O1 labels 21 helpful 6 neutral 23 harmful; O2 labels 15 helpful 5 neutral 30 harmful; all current events subject s02"
GATE-A2,2026-08-25,C00,"safety-aware optimizer audit","CHI3D val development20 confirmation20 reserve18","event; one primary view","cached BEV SMPL-X initialization","GT top-1 edge","oracle","official HHCS max 1000x2 plus audited hhcgen","hhcmap 0.10 plus hhcgen 0.03","configs/gate_a2_safety_repair.yaml","n/a","42 deterministic path; 10000 event bootstrap","RTX 5090","weight cells mean 53.24s on confirmation; hhcgen cells mean 63.63s on reserve","completed; final confirmation_pass=true","GT-edge gain and every original pose penetration reprojection noninferiority check","Stage0 confirmation contact -41.58mm but penetration failed; final untouched reserve contact -80.86mm CI[-118.67,-47.46] penetration max/mean upper 6.21/2.60","outputs/gate_a2_penloss_reserve/analysis.json","276 full cells plus 4-step smoke; original gate_a_pass=false unchanged; scope cached BEV and CHI3D s02"
GATE-A2-WEIGHT-CONF,2026-08-25,C00,"weight-only held-out confirmation","CHI3D val confirmation20","event; one primary view","cached BEV SMPL-X initialization","GT top-1 edge","oracle","official HHCS max 1000x2","hhcmap 0.10; no hhcgen","configs/gate_a2_safety_repair.yaml","n/a","42; 10000 bootstrap","RTX 5090","mean 53.24s per cell","completed 40/40; confirmation_pass=false","weight artifact control","GT-edge -41.58mm CI[-70.75,-20.09]; penetration max upper12.79>10 and mean upper4.23>3","outputs/gate_a2_weight_confirmation/analysis.json","Consumed once and never reused for safety-loss selection"
GATE-A2-PENLOSS-RESERVE,2026-08-25,C00,"audited safety-loss held-out confirmation","CHI3D val untouched reserve18","event; one primary view","cached BEV SMPL-X initialization","GT top-1 edge","oracle","official HHCS max 1000x2 plus GeneralContactLoss","hhcmap 0.10 plus hhcgen 0.03","configs/gate_a2_safety_repair.yaml","n/a","42; 10000 bootstrap","RTX 5090","mean 63.63s per cell","completed 36/36; confirmation_pass=true","primary gain and all original safety checks","GT-edge -80.86mm CI[-118.67,-47.46]; rel-trans -44.44mm; pen max/mean upper6.21/2.60","outputs/gate_a2_penloss_reserve/analysis.json","8 helpful 2 neutral 8 harmful; aggregate pass does not establish selector predictability"
CHI3D-CACHED-INIT-V0,2026-08-24,C00,"initializer and metric preflight","CHI3D val 118 eligible events","event; one primary view","cached BEV SMPL-X initialization","none","none","none","none","scripts/audit_chi3d_cached_geometry.py","n/a","deterministic","RTX 5090","~30 seconds including intersection","completed; non-scientific preflight","GT-edge distance relative translation reprojection penetration","mean GT-edge distance 164.96mm; relative translation 180.85mm; reprojection 17.27px; max penetration 35.16mm","logs/chi3d_cached_initializer_audit.log","scientific_result=false; establishes metric plumbing and likely headroom only"
VLM-S00,2026-08-24,C01,"VLM model selection","20 debug plus 80 locked events","event and subject pair","frozen common initializer","three VLMs plus union fixed bank","geometry versus geometry plus VLM logistic","frozen paired optimizer","frozen region contact loss","to create","Qwen InternVL MiniCPM revisions pending","5 proposer seeds","RTX 5090","pending","planned","utility AUPRC AURC matched-coverage HarmRate latency VRAM JSON validity","","","Primary model selection is downstream utility not contact F1"
VLM-S01,2026-08-24,C02,"identity and visual prompt audit","locked VLM audit subset","event","n/a","fixed edge bank","n/a","n/a","n/a","to create","same VLM revision","fixed","RTX 5090","pending","planned","identity error marker bias grounding accuracy","","","raw marked joint crop contact-site crop and color swap"
VLM-S02,2026-08-24,C03,"verifier increment","locked VLM audit subset","event","frozen common initializer","fixed bank with balanced negatives","proposer evidence versus plus verifier","frozen paired optimizer","frozen loss","to create","same VLM revision","fixed","RTX 5090","pending","planned","delta utility AUPRC confirmation bias","","","Verifier sees positives negatives and adversarial substitutions"
VLM-S03,2026-08-24,C02,"sampling and top-K ablation","locked VLM audit subset","event","n/a","VLM K=1/3/5","rule and logistic","n/a","n/a","to create","same VLM revision","N=1/3/5/10","RTX 5090","pending","planned","candidate recall utility gain query cost","","","Measure saturation; do not call agreement utility"
VLM-S04,2026-08-24,C04,"consistency feature ablation","locked VLM audit subset","event","frozen common initializer","fixed bank","geometry plus selected consistency features","frozen paired optimizer","frozen loss","to create","same VLM revision","fixed","RTX 5090","pending","planned","delta AUPRC delta AURC HarmRate","","","Paraphrase flip person swap crop each removed"
VLM-S05,2026-08-24,C04,"consistent-yet-wrong audit","locked VLM audit subset","event","n/a","fixed bank","n/a","n/a","n/a","to create","same VLM revision","fixed","RTX 5090","pending","planned","consistency-contact correlation consistency-utility correlation stable-error rate","","","Separates evidence stability from correctness and utility"
VLM-S06,2026-08-24,C05,"confidence calibration","locked VLM audit subset","event","n/a","fixed bank","raw confidence token margin calibrated evidence","n/a","n/a","to create","same VLM revision","fixed","RTX 5090","pending","planned","ECE Brier AUPRC risk coverage","","","Calibration split only; self confidence is weak baseline"
VLM-S07,2026-08-24,C06,"cross-VLM transfer","held-out events","event and subject pair","frozen common initializer","fixed bank","Qwen-trained selector transferred to InternVL and MiniCPM","frozen paired optimizer","frozen loss","to create","three frozen revisions","3 selector seeds","RTX 5090","pending","planned","zero-shot recalibrated retrained AURC and HarmRate","","","Compare raw standardized and target-recalibrated features"
VLM-S08,2026-08-24,C03,"verification confirmation bias","locked VLM audit subset","event","n/a","balanced true false near-contact adversarial bank","n/a","n/a","n/a","to create","same VLM revision","fixed choice permutations","RTX 5090","pending","planned","TPR TNR near-contact error option-position bias","","","Candidate wording and choice order randomized"
VLM-S09,2026-08-24,C07,"adaptive query budget","held-out events","event","frozen common initializer","fixed bank","always-query versus uncertainty cascade","frozen paired optimizer","frozen loss","to create","selected VLM revision","3 selector seeds","RTX 5090","pending","planned","safe coverage queries latency Pareto","","","Full or stretch after selector calibration"
VLM-S10,2026-08-24,C08,"3D counterfactual visual prompt","held-out subset","event","frozen common initializer","fixed bank","RGB-only versus highlighted-render evidence","frozen paired optimizer","frozen loss","to create","selected VLM revision","fixed","RTX 5090","pending","planned","incremental utility AUPRC AURC token cost","","","Stretch; novelty re-audit before implementation"
TC-MAIN,2026-08-24,C09,"main selective refinement","CHI3D grouped test","event and subject pair","frozen selected initializer","fixed common bank","geometry versus geometry plus VLM utility selector","frozen paired optimizer","frozen loss","to create","selected VLM revision","3 selector seeds","RTX 5090","pending","blocked by Gate-A2 baseline gate Gate-V0 and Gate-B","HarmRate AURC retained oracle gain pose penetration","","","No test execution until safer-optimizer control fixed candidates and calibration are frozen"
CHI3D-INPUT-V1,2026-08-25,C00,"unified baseline input-integrity gate","CHI3D main-50 frozen cohort","event; one primary view","n/a","n/a","n/a","n/a","n/a","configs/chi3d_baseline_manifest.yaml","n/a","deterministic","RTX 5090","sub-second after hashes cached by OS","implemented; awaiting licensed raw images","manifest hash per-image SHA-256 bytes dimensions event order","fail-closed verifier implemented; current frozen manifest absent","audit/chi3d_baseline_input_verification_v1.json","Must pass immediately before every PromptHMR or Multi-HMR batch"
HI4D-AUDIT-V1,2026-08-26,C10,"licensed dataset exact structure audit","Hi4D train and test archives","sequence camera frame person","n/a","n/a","n/a","n/a","n/a","scripts/audit_hi4d_dataset.py","n/a","deterministic","CPU","peak 24.43GB train annotation load","completed; all exact checks pass","sequence disjointness archive coverage cameras two-person schema contact index range","75/25 train/test sequences; 95560 images; 191120 masks; SMPL6890 one-based contact IDs","/root/autodl-tmp/vlmcontact/audit/hi4d/dataset_audit_v1.json","Contact correspondence semantics beyond index range still require geometry/source validation"
HI4D-MANIFEST-TEST100-V1,2026-08-26,C10,"immutable baseline input freeze","Hi4D test 25 sequences Camera04","sequence clustered; four frames each","n/a","n/a","n/a","n/a","n/a","scripts/freeze_hi4d_manifest.py","n/a","deterministic","CPU","hash and extraction preflight","completed 100/100","path dimensions bytes image SHA256 and manifest SHA256","manifest HI4D-BASELINE-TEST100-V1 sha256 32b7af96a1af8c93e7c840b5159d973d194a008cc403059014bac6b11b8cc5e4e","/root/autodl-tmp/vlmcontact/audit/hi4d/manifest_freeze_test100_v1.json","Rows within each sequence are correlated; cluster bootstrap mandatory"
PROMPTHMR-HI4D-GTBOX-V1,2026-08-26,C10,"interaction off/on mechanism control","HI4D-BASELINE-TEST100-V1","sequence clustered","PromptHMR official checkpoint","Hi4D person-id boxes","none","released cross-person interaction off/on","none","scripts/run_prompthmr_manifest.py; box-only GT boxes SDPA-sm120 adapter","checkpoint sha256 a3ef04ef8a12c3068682b03c62c95f8959cd8554424e105c63bf97f6c8e97e99","20260825","RTX 5090","median 48.20ms off 49.52ms on","completed 100/100","shared SMPL24 root PA pair-PA relative pelvis","on-off root -0.378mm CI[-0.709,-0.028]; pair-PA +1.140[0.115,2.208]; relative pelvis +3.692[0.626,6.796]","results/hi4d_shared_joint_eval_v1/summary.json","Mechanism control only; GT boxes are not a detector-native full pipeline"
PROMPTHMR-HI4D-DETECTOR-V1,2026-08-26,C10,"detector-native interaction off/on baseline","HI4D-BASELINE-TEST100-V1","sequence clustered","PromptHMR official checkpoint","YOLOv8x top-two person boxes","none","released cross-person interaction off/on","none","scripts/run_prompthmr_manifest.py; detector-native box identity matching","same pinned official checkpoint","20260825","RTX 5090","median 54.83ms off 56.70ms on","completed inference 97/100 two-person coverage","coverage plus shared SMPL24 root PA pair-PA relative pelvis","off/on root37.53/36.98 PA20.34/20.05 pair50.28/49.96mm; pair delta CI crosses zero","results/hi4d_shared_joint_eval_detector_v1/summary.json","Three one-person detector failures retained and excluded only from conditional geometry mean"
MULTIHMR-HI4D-896L-V1,2026-08-26,C10,"ECCV baseline full-image inference","HI4D-BASELINE-TEST100-V1","sequence clustered","Multi-HMR 896L full image","internal detection","none","released model","none","scripts/run_multihmr_manifest.py; FOV60 det0.3 NMS3 offline DINOv2","multiHMR_896_L sha256 c1b55933eb35df9b27966e1d8d3a69c4bc5e28eee0f0dd7f0cdc41c3f27202ca","20260825","RTX 5090","median 42.84ms forward","completed 100/100 inference; 93/100 two-person coverage","coverage plus shared SMPL24 root PA pair-PA relative pelvis","conditional root70.11 PA54.86 pair85.89 relative pelvis159.43mm","results/hi4d_shared_joint_eval_detector_v1/summary.json","Seven one-person outputs retained; 2026 Anny checkpoint excluded from scientific main row"
HI4D-SHARED-JOINT-EVAL-V1,2026-08-26,C10,"shared baseline evaluator","HI4D-BASELINE-TEST100-V1","25 sequence clusters","method-specific predictions","n/a","n/a","n/a","n/a","scripts/evaluate_hi4d_joint_baselines.py; SMPL24 hip root 48-joint pair PA 10000 cluster bootstrap","SMPLX-to-SMPL joint regressor sha256 49864006d20c5205d7b30697db01744f3a56faf787c1cea7448b25895e4b16ab2","20260825","CPU","under one minute including 4GB annotation load","completed; auditable local protocol","coverage root MPJPE PA-MPJPE pair-PA-MPJPE relative pelvis","Prompt detector outperforms Multi conditionally; GT-box interaction improves individual pose but harms pair layout","results/hi4d_shared_joint_eval_v1/summary.json","Not exact PromptHMR H36M14/PVE paper protocol; no cross-method claim without coverage column"
GATE-V0-QWEN3VL-DEBUG20-V3,2026-08-26,C01,"frozen VLM proposer smoke","HI4D DEBUG20; LOCKED80 untouched","5 complete sequences; 20 images","n/a","Qwen3-VL-8B-Instruct frozen proposer top-3","none","none","none","configs/gate_v0_protocol_v3_2.json","HF revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b; verified 16/16 files","5 samples/image; seed base 2026082600","RTX 5090","288.72s summed inference; median 2.62s/sample","completed 100/100; gate_pass=false; provenance 10/10","JSON identity contact-existence coarse-edge precision laterality sampling provenance","JSON/identity 1.0; frozen contact BA 0.444; coarse precision 0; candidate-only diagnostic BA 0.899 but edge precision 0.10","audit/gate_v0/qwen_debug20_v3_2_audit.json","v3.2 corrects manifest-hash transcription only; post-gate diagnostic does not change failure; no LOCKED80 cross-VLM or selector run"
GATE-V0B-QWEN-FIXED-EDGE-PREFLIGHT,2026-08-26,C03,"balanced fixed-edge verifier preflight","HI4D DEBUG20; 2 candidates only","one true and same-event hard negative","PromptHMR GT-box baseline available but not used in prompt","frozen 49-edge coarse bank","two deterministic verifier prompts","none","none","configs/gate_v0b_protocol_v3.json","Qwen3-VL revision 0c351dd01ed87e9c1b53cbc748cba10e6187ff3b","deterministic seed base 2026082600","RTX 5090","11.69s summed generation","completed preflight; gate_pass=false; full 98-query run not started","JSON prompt agreement polarity conflict truth-conditioned decision","JSON 0.75; support-first supported true and false; falsification-first contradicted true and false","audit/gate_v0b/qwen_preflight_v2_audit.json","Direction-agnostic bank replaced invalid abdomen-front mapping; LOCKED80 untouched"
GATE-V0C-QWEN-PAIRED-LOCALIZATION,2026-08-26,C03,"paired boxes-only versus predicted-joint overlay","HI4D DEBUG20; same 2 candidates","paired visual intervention","PromptHMR interaction-off predicted OpenPose BODY-25 joints","fixed true and hard-negative edges","two byte-identical deterministic prompts per arm","none","none","configs/gate_v0c_protocol_v1.json","same frozen Qwen revision","4 queries per arm","RTX 5090","12.00s control plus 12.02s overlay generation","completed paired preflight; gate_pass=false","JSON identity echo prompt agreement polarity conflict","Both arms agreement 0 and polarity conflict 1.0; overlay changed no decision","audit/gate_v0c/qwen_paired_preflight_v1_audit.json","Corrected prior SMPL-X/OpenPose index mismatch; 49-edge full run prohibited"
GATE-V0D-QWEN-NEUTRAL-EVIDENCE,2026-08-26,C03,"neutral evidence decomposition prompt-development","HI4D basketball28 prompt-dev; talk23 audit untouched","4 candidates; 2 positive and 2 same-event negative","PromptHMR interaction-off predicted joint regions","fixed evidence bank","fixed-rule decision from alignment and surface relation","none","none","configs/gate_v0d_protocol_v1.json","same frozen Qwen revision","2 prompts per edge deterministic","RTX 5090","24.97s summed generation","completed prompt-dev; gate_pass=false; audit not run","validity agreement precision recall specificity coverage","JSON 1.0; agreement 0.5; all 4 consensus uncertain; decision coverage 0","audit/gate_v0d/qwen_prompt_dev_v1_audit.json","Sequence-disjoint audit was never inspected or inferred"
HI4D-CONTACT-VISIBILITY-DEBUG20-V1,2026-08-26,C03,"camera-specific GT contact visibility diagnostic","HI4D DEBUG20 Camera04","20 frames and 87 coarse contact edges","neutral SMPL from Hi4D GT pose betas trans","GT contact correspondences for evaluation only","n/a","n/a","n/a","scripts/build_hi4d_contact_visibility_gt.py","SMPL_NEUTRAL plus Hi4D masks and annotation hashes stored","deterministic z-buffer","RTX 5090","under 30s including rasterization; annotation load excluded","completed diagnostic proxy","joint reprojection visible silhouette IoU endpoint visibility","mean reprojection 11.59px; mean IoU 0.710; all 87 edges at least 5 percent both-visible endpoints","data/gate_v0/hi4d_debug20_contact_visibility_v1.json","Not official Hi4D visibility labels; worst person mean 25.32px and min IoU 0.340"
GATE-V0E-QWEN-SURFACE-GROUNDING,2026-08-26,C03,"predicted mesh-surface grounding prompt-development","HI4D basketball28 prompt-dev; talk23 audit untouched","same 4 candidates as V0d","PromptHMR interaction-off SMPL-X to SMPL visible surface render","fixed evidence bank","same V0d fixed-rule selector","none","none","configs/gate_v0e_protocol_v1.json","same frozen Qwen revision","2 prompts per edge deterministic","RTX 5090","25.30s summed generation","completed prompt-dev; gate_pass=false; audit not run","agreement polarity precision recall specificity coverage","JSON 1.0; agreement 0.5; polarity conflict 0.25; all 4 consensus uncertain; coverage 0","audit/gate_v0e/qwen_prompt_dev_v1_audit.json","Only visual representation changed from V0d; no GT contact or GT masks used in overlay"
GATE-B-PILOT-V1,2026-08-26,C09,"fixed-edge outcome-label and geometry-only development pilot","CHI3D Gate-A2 reserve18 previously inspected","event; leave-one-interaction-family-out","cached BEV SMPL-X","frozen GT top-1 75-region edge","standardized geometry logistic","official ProsePose HHCS","hhcmap0.10 plus hhcgen0.03","configs/gate_b_pilot_v1.yaml","n/a","optimizer42;selector20260826;bootstrap20260826","RTX 5090 interventions plus CPU selector","reuses completed 36 optimizer cells; 10000 group bootstrap","completed development only; semantic residual not evaluable","helpful-neutral-harmful harmful AP ROC-AUC Brier risk-coverage","8 helpful 2 neutral 8 harmful; geometry AP0.714 AUC0.688 Brier0.211 risk-AUC0.284; 25pct coverage HarmRate0","outputs/gate_b_pilot_v1/geometry_evaluation.json","No outcome leakage; candidate hash 99bb5827...; Qwen Hi4D bank has zero exact overlap so geometry+VLM was not run"
GATE-B-HI4D-SMOKE-V1,2026-08-26,C09,"aligned paired-intervention engineering smoke","Hi4D basketball28 prompt-development","2 events; 4 fixed candidates; 8 O0/O1 cells","PromptHMR box-only interaction-off","balanced GT-positive and same-event counterfactual coarse edges","none","official ProsePose HHCS","hhcmap0.10 plus hhcgen0.03","configs/gate_b_hi4d_smoke_v1.yaml","PromptHMR and SMPL-X assets frozen in result provenance","42","RTX 5090","8 cells with 2 iterations per stage","completed; smoke_pass=true","pair completeness initialization equality edge cardinality gradient determinism finiteness","8/8 cells complete; all checks true; authorized full same four candidates","audit/gate_b/hi4d_smoke_v1_audit.json","Engineering gate only; no outcome or selector claim"
GATE-B-HI4D-PROMPTDEV-V1,2026-08-26,C09,"aligned real optimizer outcomes and offline VLM residual diagnostic","Hi4D basketball28 prompt-development","2 events; 4 candidates; event-confounded descriptive diagnostic","PromptHMR box-only interaction-off","same frozen V0d/V0e balanced candidates","no selector trained","official ProsePose HHCS","hhcmap0.10 plus hhcgen0.03","configs/gate_b_hi4d_smoke_v1.yaml","same frozen Qwen revision for offline evidence","optimizer42; deterministic cached VLM","RTX 5090","8 full cells; approximately 99-141s each","outcome mechanism complete; semantic residual gate_pass=false","GT contact correspondence safety deltas fixed-rule coverage alignment","1 helpful 2 neutral 1 harmful; V0d/V0e fixed-rule coverage0; exact candidate event image edge alignment","audit/gate_b/hi4d_promptdev_vlm_residual_v1.json","n4 two events; no learned residual estimate; LOCKED80 untouched"
GATE-B-HI4D-DEBUG20-V1,2026-08-26,C11,"frozen paired contact intervention and flat residual test","Hi4D DEBUG20 development","13 events from only 5 sequences","PromptHMR GT-box box-only interaction-off","35 frozen Qwen coarse candidates","event-grouped standardized logistic controls","ProsePose HHCS","hhcmap 0.10 plus hhcgen 0.03","configs/gate_b_hi4d_debug20_v1.yaml","n/a","42 optimizer; 20260826 selector","RTX 5090","48/48 full cells complete","completed; registered No-Go","harm AP ROC-AUC Brier risk-coverage and matched coverage","28 harmful 5 helpful 2 neutral; all four VLM-residual Go checks false","outputs/gate_b_hi4d_debug20_v1/selector_evaluation.json","Development only; event grouping shares sequence context and LOCKED80 remains frozen"
GATE-B-FAILURE-ATTRIBUTION-V1,2026-08-26,C11,"post-hoc failure attribution without refitting","Hi4D Gate-B DEBUG20","candidate event and sequence audit","frozen","frozen","none","none","none","n/a","n/a","n/a","CPU","seconds","completed; diagnostic only","candidate bias label mechanism effective sample feature rank and target alignment","3/35 semantic positive; 16/28 harmful safety-only; 35 candidates from 5 sequences; start BR0 not locked test","audit/gate_b/hi4d_debug20_failure_attribution_v1.json","Four tests passed; p-values and univariate AUCs are descriptive only"
GATE-BR0-HI4D-DEBUG20-V1,2026-08-26,C12,"candidate-local vector target repair and disagreement audit","Hi4D Gate-B DEBUG20","35 candidates; 13 events; 5 sequences","saved paired O0/O1 SMPL-X parameters","frozen natural Qwen bank","none","mesh reconstruction audit only","none","research/09_gate_br0_vector_outcome_audit.md","n/a","42 inherited","RTX 5090 mesh reconstruction plus CPU audit","48 body-model reconstructions; no optimization","completed; BR0 pass_for_BR1=true","candidate-local distance desired contact induced false contact global geometry safety and compute vector","48/48 reconstructed distances within 0.0238mm; four false-helpful edges remain 127.75-332.85mm apart; zero false edges within 50mm","audit/gate_br0/hi4d_debug20_vector_outcomes_v1.json","Post-hoc development target audit; LOCKED80 untouched; no VLM inference or selector training"
GATE-BR1-HI4D-TRAIN12-V1,2026-08-26,C12,"reversible short optimizer response probe","Hi4D train Camera04; DEBUG20 and LOCKED80 excluded","12 sequences; 12 subject pairs; 36 candidates","PromptHMR GT-box interaction-off","12 GT-positive plus 12 nearest hard-negative plus 12 distance-matched negative","none","official ProsePose HHCS","hhcmap0.10 plus hhcgen0.03","configs/gate_br1_response_probe_v1.yaml; configs/gate_br1_decision_policy_v1.yaml","n/a","42","RTX 5090","240 optimizer cells; 5851s raw matrix plus vector evaluation","completed; frozen decision pause_and_diagnose","probe-to-full candidate-local and global Spearman safety and false-contact violation recall runtime fraction","K100 local/global rho 0.831/0.745 and runtime 0.242x pass; safety recall 4/7 fails; no K eligible","audit/gate_br1/hi4d_train12_probe_response_v1.json","Innovation-one development No-Go; diagnose delayed safety before BR2; no new VLM selector LOCKED80 or innovation two"
GATE-BR1-SAFETY-TRAJECTORY-V1,2026-08-26,C14,"post-hoc delayed-safety trajectory and gradient-conflict audit","Hi4D BR1 train12","36 candidates grouped by 12 sequences","saved K1/5/20/100/1000 SMPL-X states","frozen three-stratum bank","none","read-only trajectory replay","none","configs/gate_br1_safety_diagnosis_v1_1.yaml","n/a","bootstrap 20260826","RTX 5090 gradient recomputation plus CPU audit","72 endpoint gradient states; no optimizer rerun","completed; require_intermediate_checkpoint_pilot","full and late violation recall precision specificity gradient-conflict AUC","best fixed bound 6/7 recall but precision0.429; K100 negative gradient cosine AUC0.918 on 27 finite cases","audit/gate_br1_safety/hi4d_train12_delayed_safety_diagnosis_v1_1.json","Post-hoc same-cohort diagnosis only; V1.1 discloses exact 2/3 floating-point correction"
GATE-BR1-SAFETY-K250-V1,2026-08-26,C14,"intermediate safety checkpoint pilot","Hi4D BR1 train12","36 candidates; 12 sequences; 48 cells","PromptHMR GT-box interaction-off","same frozen three-stratum bank","none","official ProsePose HHCS","hhcmap0.10 plus hhcgen0.03","configs/gate_br1_safety_k250_pilot_v1.yaml","n/a","42","RTX 5090","1326.79s O1 aggregate; 0.5053x K1000","completed; stop_checkpoint_expansion","K250 direct safety recall late recall precision and runtime fraction","recall6/7=0.857; late2/3; precision1.0; runtime0.5053 exceeds frozen0.50; BR1-10-2 remains false negative","audit/gate_br1_safety/hi4d_train12_k250_safety_v1.json","Same-cohort post-hoc pilot; no new confirmation authorized"
GATE-I1-PENETRATION-FILTER-TARGETED-V1,2026-08-27,C15,"preregistered optimizer mechanism smoke","Hi4D BR1 known failures","3 selected candidates","PromptHMR GT-box interaction-off","frozen BR1 edges","none","official HHCS plus post-step penetration filter","hhcmap0.10 plus hhcgen0.03","configs/gate_i1_penetration_filter_targeted_v1.yaml","n/a","42","RTX 5090","three K1000 filtered interventions","completed; rejected","frozen safety vector candidate-local gain proposal acceptance runtime","3/3 harmful; filter freezes base fitting and does not remove full safety failures","outputs/gate_i1_penetration_filter_targeted_v1","Known-case mechanism smoke only; no threshold retuning"
GATE-I1-COUNTERFACTUAL-FILTER-TARGETED-V2,2026-08-27,C15,"preregistered same-step counterfactual filter smoke","Hi4D BR1 known failures","3 selected candidates","PromptHMR GT-box interaction-off","frozen BR1 edges","none","official HHCS plus base-vs-full residual filter","hhcmap0.10 plus hhcgen0.03","configs/gate_i1_counterfactual_filter_targeted_v2.yaml","n/a","42","RTX 5090","175.8--295.3s per intervention","completed; rejected","frozen safety vector candidate-local gain runtime","BR1-11-0 neutral and safe but 10-2/11-2 harmful; about 2--4.5x weighted-sum cost","outputs/gate_i1_counterfactual_filter_targeted_v2","Proxy insufficient; rejected without threshold tuning"
GATE-I1-HI4D-TRAIN15-GRADIENT-COMPAT-CONFIRMATION-V1,2026-08-27,C15,"sequence and subject-pair disjoint Innovation-One confirmation","Hi4D train Camera04","15 sequences; 3 unseen subject pairs; 45 candidates","PromptHMR GT-box interaction-off","15 GT-positive plus 15 nearest hard-negative plus 15 distance-matched negative","fixed zero-threshold gradient-compatibility abstention","official ProsePose HHCS","hhcmap0.10 plus hhcgen0.03","configs/gate_i1_gradient_compat_confirmation_v1.yaml","n/a","42; bootstrap5000","RTX 5090","60 K100 cells plus 60 K1000 cells; expected policy runtime fraction0.379","completed; frozen confirmation failed","coverage accepted full-safety violations harm-rate reduction helpful retention restart compute","accepted5/45 coverage0.111; harm1/5=0.20; safety violations1; helpful retained2/13; only harm improvement and runtime pass","audit/gate_i1/hi4d_train15_gradient_compat_confirmation_v1.json","No threshold tuning; conditional PCGrad and matched-coverage expansion correctly not run; LOCKED80 untouched"
GATE-I1-FACTORIZED-FEATURE-FAMILY-AUDIT-V1,2026-08-27,C15,"post-failure grouped feature-family reconstruction audit","Hi4D BR1 train12","36 candidates; leave-one-sequence-out","saved K100 PromptHMR/HHCS states","same frozen three-stratum bank","dual harmful-risk and helpful-benefit logistic heads","read-only response and gradient replay","none","audit/gate_i1/hi4d_feature_family_audit_v1.json","n/a","20260827","CPU plus RTX 5090 clearance extraction","36 endpoint reconstructions; no new optimizer outcome","completed; all four families rejected","harmful and helpful OOF AUC plus strict selection gate","risk AUC response0.482 gradient0.451 clearance0.409 combined0.455; no family has feasible safety-preserving coverage","audit/gate_i1/hi4d_feature_family_audit_v1.json","Feature families defined after prior failure; former confirmation and LOCKED80 not used"
GATE-I1-COUNTERFACTUAL-OFFTARGET-TARGETED-V3,2026-08-27,C15,"candidate-conditioned off-target counterfactual optimizer smoke","Hi4D BR1 known failures","3 selected candidates","PromptHMR GT-box interaction-off","frozen BR1 edges","none","HHCS same-step base/full residual filter plus 30mm off-target envelope","hhcmap0.10 plus hhcgen0.03","configs/gate_i1_counterfactual_offtarget_filter_targeted_v3.yaml","n/a","42","RTX 5090","820.91s total; 202.36--358.59s per intervention","completed; reject_v3_stop_method_expansion","frozen safety vector candidate-local full-min and RMS gain trace runtime","BR1-10-2 and 11-0 safe with +13.126/+0.037mm gain; BR1-11-2 penetration-max violation remains","audit/gate_i1/hi4d_counterfactual_offtarget_targeted_v3.json","Known-case diagnostic only; sparse envelope misses non-target R75 3/17/72 versus 61 collision; no retuning"
GATE-I1-TWINGUARD-TARGETED-R4,2026-08-27,C16,"evaluator-congruent full-mesh counterfactual certificate","Hi4D BR1 known failures","2+2 smoke then first formal targeted candidate","PromptHMR GT-box interaction-off","frozen BR1 edges","none","synchronized contact-off/contact-on HHCS twins with bidirectional full-mesh penetration certificate","hhcmap0.10 plus hhcgen0.03","configs/gate_i1_twinguard_r4.yaml","n/a","42","RTX 5090","BR1-10-2 R4 341.535s versus paired O0+O1 192.746s","completed; rejected at frozen efficiency stop","candidate-local gain full safety vector certificate trace runtime fraction","runtime fraction1.77195 exceeds frozen1.25 maximum; remaining targeted matrix stopped","audit/gate_i1/hi4d_twinguard_r4_runtime_stop_v1.json","Smoke and metric-congruence pass do not rescue deployability; thresholds/frequency not tuned"
CIU-BENCH-PILOT-V1,2026-08-27,C17,"Contact Intervention Utility problem/evaluation MVP","Hi4D train Camera04; three sequence-disjoint historical cohorts","116 candidates; 40 events; 32 sequences","PromptHMR GT-box interaction-off","35 natural Qwen plus 81 balanced GT-positive/hard-negative/distance-matched","11 rule failed-method and oracle policies","official ProsePose HHCS paired O0/O1","hhcmap0.10 plus hhcgen0.03","configs/ciu_bench_pilot_v1.yaml","n/a","optimizer42;bootstrap42x5000","RTX 5090 outcomes plus CPU build","reuses frozen optimizer cells; artifact build 8s","completed; innovation_one_problem_and_evaluation_mvp","coverage harmful rate safety violation helpful retention local/global gain cost cluster CI","116/116 complete primary/safety records; always harm47.4%; semantic-filter harm30.0%; K100 local rho0.907 but late safety9/38; all MVP checks pass","audit/ciu_bench/ciu_bench_pilot_v1.json","Not a positive selector or full conference benchmark; train15 is archival not a fresh future test"
CIU-F0-SCOPE-FREEZE-V1,2026-08-27,C21,"fail-closed Hi4D exposure and split-feasibility audit","Hi4D full train+test Camera04","100 sequences;20 subject pairs; historical manifests and CIU pilot","n/a","n/a","none","read-only inventory","none","configs/ciu_f0_scope_freeze_v1.yaml","n/a","deterministic","CPU","full annotation inventory","completed; in-dataset subject-pair test blocked","historical sequence/pair exposure fully-unexposed pool split feasibility","48 fully unexposed sequences but 0 fully unexposed subject pairs; construction smoke allowed only","audit/ciu_f0/hi4d_scope_freeze_v1.json","Do not describe any new Hi4D split as unseen-subject-pair generalization"
CIU-F0-INITIALIZER-QWEN-SMOKE20-V1,2026-08-27,C22,"construction-only initializer and natural proposer validation","Hi4D train Camera04","20 outcome-blind sequences;12 subject pairs;12 interactions","PromptHMR GT-box off/on and Multi-HMR 896L","Qwen3-VL-8B-Instruct 20x5 frozen proposer","none","none","none","configs/ciu_f0_construction_smoke20_v1.yaml;configs/gate_v0_qwen_prompt_v3.txt","Qwen revision0c351dd; PromptHMR and Multi-HMR pinned checkpoints","2026082700","RTX 5090","Qwen 100 samples at about2.6-5.5s each","completed; construction diagnostic","two-person coverage JSON validity R75 proposal mapping","PromptHMR coverage1.0;Multi-HMR0.8;Qwen JSON100/100;20 proposals;5 posthoc GT-positive;1 alias","audit/ciu_f0/hi4d_smoke20_natural_proposal_freeze_v1.json","Not a VLM accuracy result or held-out evaluation"
CIU-F0-K20-CONSTRUCTION-SMOKE-V1,2026-08-27,C22,"four-stratum paired intervention pipeline smoke","Hi4D train Camera04","20 events;80 candidates;100 cells","PromptHMR GT-box interaction-off","20 each GT-positive nearest-hard-negative distance-matched-negative natural-Qwen","none","official ProsePose HHCS weighted sum","hhcmap0.10 plus hhcgen0.03","configs/ciu_f0_construction_smoke20_v1.yaml","n/a","42","RTX 5090","923.817s for 100 K20 cells plus evaluator","completed; closure16/16 pass","O0 reuse O1 execution seven-axis safety vector cost logging","100/100 cells;80/80 seven-axis vectors;K20 diagnostic labels18/34/28 helpful-neutral-harmful","audit/ciu_f0/ciu_f0_construction_smoke20_closure_v1.json","K20 labels are not full outcomes DEV VAL TEST or population estimates"
CIU-F1-PROTOCOL-FREEZE-V1,2026-08-27,C23,"pre-outcome exposure split power and access-control freeze","Hi4D full train+test Camera04","100 sequence ledger; DEV32 VAL14 TEST-ID28 ACTION-OOD6 excluded20","PromptHMR main; frozen20-event Multi-HMR cross-context plan","four planned strata per event; no edges generated","none","ProsePose HHCS protocol frozen; no outcome run","hhcmap0.10 plus hhcgen0.03","configs/ciu_f1_protocol_freeze_v1.yaml","n/a","optimizer42;bootstrap42x10000","CPU","about90s including 24GB annotations and canonical ZIP member reads","completed; 17/17 checks pass; DEV candidate freeze only","sequence exposure overlap event and candidate counts action scope power access control manifest hashes","80 evaluation sequences;136 events;544 planned candidates;TEST-ID336 candidates;strong-effect power pass medium-effect fail","audit/ciu_f1/ciu_f1_protocol_freeze_v1.json","No VLM optimizer outcome selector or LOCKED80 inference; TEST remains sealed and external unseen-subject OOD remains required"
CIU-F2-DEV-CANDIDATE-FEATURE-FREEZE-V2,2026-08-27,C24,"DEV pre-outcome candidate and feature contract freeze","Hi4D CIU-DEV Camera04","32 sequences;32 events;128 decision slots","PromptHMR GT-box interaction-off/on","32 GT-positive;32 nearest-negative;32 matched-negative;28 Qwen proposals;4 explicit Qwen abstentions","none","no full optimizer outcome","none","configs/ciu_f2_dev_freeze_v2.yaml","Qwen revision0c351dd;PromptHMR pinned checkpoint","Qwen seed base2026082800","RTX 5090","Qwen160 samples plus candidate geometry construction","completed; closure15/15 pass; DEV K1000 authorized","initializer coverage JSON validity executable edge count abstention count leakage-safe feature schema","PromptHMR32/32;Qwen160/160 valid;124 executable edges plus4 abstentions;features124/124 with0 forbidden/outcome fields","audit/ciu_f2/ciu_f2_dev_closure_v2.json","V1 assumption of32 natural edges correctly blocked; V2 amendment occurred before optimizer outcomes; VAL TEST LOCKED80 sealed"
CIU-F2-DEV-K100-PREFLIGHT-V1,2026-08-27,C25,"bounded DEV response and safety-vector engineering preflight","Hi4D CIU-DEV Camera04","2 events;8 candidates;10 cells","PromptHMR GT-box interaction-off","two events with all four executable strata","none","official ProsePose HHCS weighted sum K100","hhcmap0.10 plus hhcgen0.03","configs/ciu_f2_dev_freeze_v2.yaml","n/a","42","RTX 5090","238.656s","completed;10/10 cells;0 failure","O0 reuse O1 execution seven-axis safety completeness runtime","2 helpful;5 neutral;1 harmful engineering labels;all10 cells have seven-axis vector","outputs/ciu_f2_dev_k0100_preflight_v1/evaluation.json","Not a full outcome selector result VLM accuracy estimate VAL TEST or population estimate"
CIU-F3-DEV-K1000-V1,2026-08-28,C26,"DEV full contact-intervention outcome construction","Hi4D CIU-DEV Camera04","32 events;124 candidates;156 cells","PromptHMR GT-box interaction-off","32 GT-positive;32 nearest-negative;32 matched-negative;28 Qwen proposals","none","official ProsePose HHCS weighted sum K1000","hhcmap0.10 plus hhcgen0.03","configs/ciu_f3_dev_k1000_v1.yaml","n/a","42","RTX 5090","11825.0s","completed;closure10/10 pass","paired utility label continuous gain seven-axis safety","43 helpful;30 neutral;51 harmful;Qwen proposals3/5/20 helpful-neutral-harmful","audit/ciu_f3/ciu_f3_dev_k1000_closure_v1.json","DEV construction outcomes only;not held-out method performance"
CIU-CORE-SELECTOR-V1,2026-08-28,C27,"sequence-grouped core selector signal experiment","Hi4D CIU-DEV Camera04","124 candidates;32 event groups;28 Qwen-proposal evaluation subset","PromptHMR GT-box interaction-off","frozen Qwen proposer","five-fold OOF geometry harmful-risk logistic filter","uses frozen K1000 labels only for training and evaluation","none","configs/ciu_core_selector_v1.yaml","n/a","42;tie20260828;bootstrap20260829","CPU","about30s plus tests","completed;authorize small held-out confirmation","coverage harmful rate helpful retention ordinal net utility cluster bootstrap","Qwen always-use harm20/28=71.4%;geometry risk at25% accepts7 with1 harm and2/3 helpful retained;early Geometry+VLM fusion rejected","audit/ciu_core_selector/ciu_core_selector_v1_audit.json","DEV OOF only;accepted n=7;CI includes zero;held-out required"
CIU-VAL-HELDOUT-CONFIRMATION-V1,2026-08-28,C28,"small sequence-disjoint frozen selector confirmation","Hi4D CIU-VAL-ID Camera04","14 sequences;11 executable Qwen proposals;3 abstentions;22 cells","PromptHMR GT-box interaction-off","frozen Qwen3-VL five-sample proposer","DEV-trained geometry harmful-risk logistic;frozen floor-25% policy","official ProsePose HHCS weighted sum K1000","hhcmap0.10 plus hhcgen0.03","configs/ciu_heldout_val_confirmation_v1.yaml","Qwen revision0c351dd;PromptHMR pinned;selector model SHA99ab4d9","optimizer42;Qwen2026083000;bootstrap20260830","RTX 5090","2246.8s cell runtime;about21min wall with3-worker tail","completed;strict held-out gate failed","coverage harmful rate helpful retention safety violation ordinal net utility harmful-risk AUC","Always harm6/11=54.5%;frozen accepts2 with0 harm 1 helpful 1 neutral but retention1/4=25%<40%;posthoc risk AUC0.933","audit/ciu_val_confirmation/closure_v1.json","TEST-ID and ACTION-OOD untouched;posthoc top3 diagnostic cannot relabel the strict failure"