failure_case_registry.csv

case_id,dataset,event_id,view_id,interaction,occlusion_level,failure_type,candidate_source,edge,selector_score,decision,outcome_delta,responsible_module,diagnosis,fixability,artifact_path,notes
F-V01,"","","","","","person identity swap","VLM","","","","","VLM identity packet","marker or A/B grounding error","testable","",""
F-V02,"","","","","","left-right confusion","VLM","","","","","VLM grounding","anatomical versus image-frame ambiguity","testable","",""
F-V03,"","","","","","stable but wrong consistency","VLM","","","","","VLM evidence","prior-driven answer collapse","partly testable","",""
F-V04,"","","","","","verification confirmation bias","VLM","","","","","VLM verifier","candidate wording or acquiescence","testable","",""
F-V05,"","","","","","JSON invalid or schema retry","VLM","","","","","VLM interface","decoding and parser mismatch","fixable","",""
F-V06,"","","","","","semantically correct but optimizer harmful","mixed","","","","","optimizer-selector interface","context-dependent constraint utility","core target","",""
F-V07,"","","","","","semantically wrong but geometrically harmless","mixed","","","","","optimizer context","inactive gradient or already-satisfied edge","analyzable","",""
F-V08,"","","","","","outcome leakage","pipeline","","","","","data pipeline","post-intervention feature entered selector","must prevent","",""
F-V09,"","","","","","cross-VLM calibration shift","VLM","","","","","evidence normalization","raw score scale and output-style shift","partly fixable","",""
F-V10,"","","","","","no-contact hallucination","VLM","","","","","VLM proposer","action prior overrides visible evidence","partly fixable","",""
F-B01,"Hi4D","cheers32/000096;dance27/000061;dance28/000049","Camera04","cheers/dance","high or ambiguous","two-person detector miss","PromptHMR YOLOv8x","n/a","n/a","failed before off/on","no geometry output","person detector","Only one person exceeded detector threshold in 3/100 frozen frames","report coverage; detector control possible","/root/autodl-tmp/vlmcontact/results/prompthmr_hi4d_test100_box_v1","Preserved in both off/on conditions; GT-box control succeeds 100/100"
F-B02,"Hi4D","basketball28/000047;hug27/000061;hug27/000090;pose32/000060;pose32/000080;sidehug32/000061;sidehug32/000090","Camera04","multiple","high interaction overlap","one-person output","Multi-HMR 896L","n/a","n/a","conditional geometry skipped","missing second mesh","internal detection/reconstruction","Model returns one human in 7/100 frozen frames","report coverage; threshold ablation later","/root/autodl-tmp/vlmcontact/results/multihmr_hi4d_test100_896l_v1","Never imputed or silently removed from denominator"
F-B03,"Hi4D","aggregate GT-box cohort","Camera04","multiple","mixed","interaction improves individuals but harms pair layout","PromptHMR interaction-on","n/a","n/a","always on","pair-PA +1.140mm; relative pelvis +3.692mm","cross-person interaction module","Paired sequence-cluster intervals exclude zero despite root/PA improvement","core motivation for outcome-conditioned selection","results/hi4d_shared_joint_eval_v1/summary.json","Does not show a VLM selector can predict the harmful events"
F-V10-G0,"Hi4D","talk23/000039","Camera04","talk","visually separated","stable no-contact hallucination","Qwen3-VL v3 proposer","A.left_hand-B.left_hand contact","n/a","candidate emitted in 5/5 samples","GT has no mesh contact","VLM proposer","Prompt safeguards removed most negative hallucinations but this error is perfectly sampling-consistent","fixed-edge verification and visual grounding probe required","results/gate_v0_qwen_debug20_v3/talk23__000039","Consistency is not correctness and must remain only a selector feature"
F-V11-G0,"Hi4D","11 positive DEBUG20 events","Camera04","multiple","mixed/occluded","contact-to-near-contact state collapse","Qwen3-VL v3 proposer","aggregate","n/a","no majority contact prediction","contact TPR 0/11; 112 near_contact vs 5 contact edges","VLM state assignment","Conservative v3 fixed negative hallucination at the cost of never committing to contact on positive events","balanced verifier with explicit abstention","audit/gate_v0/qwen_debug20_v3_2_audit.json","Candidate-level existence is exploratory and cannot overwrite the frozen Gate"
F-V12-G0,"Hi4D","DEBUG20 aggregate","Camera04","multiple","mixed","fine body-part grounding mismatch","Qwen3-VL v3 proposer","20 consensus candidate edges","n/a","candidate diagnostic only","2 exact coarse TPs and 18 FPs; precision 0.10","VLM region grounding and RGB-to-mesh ontology interface","Model often detects interaction proximity but chooses the wrong side or limb/torso region","test fixed-edge verifier and freeze semantic-to-75 mapping only after audit","audit/gate_v0/qwen_debug20_v3_2_audit.json","No optimizer-ready 31-to-75 mapping exists yet"
F-V13-G0,"Hi4D","fever37/000073","Camera04","fever","mixed","prompt-polarity lock","Qwen3-VL fixed-edge verifier","true right_hand-left_arm and false left_hand-left_arm","n/a","support prompt always supported; falsification prompt always contradicted","agreement 0; polarity conflict 1.0 in both boxes-only and overlay arms","VLM verifier prompting","Model follows the requested verification stance more strongly than the fixed image/candidate evidence","not repaired by OpenPose joint overlay; stop direct verifier","audit/gate_v0c/qwen_paired_preflight_v1_audit.json","JSON identity echo and provenance all pass, isolating semantic decision failure"
F-E01-G0,"Hi4D","fever37/000073","Camera04","fever","n/a","wrong joint-convention visualization","PromptHMR body_joints","right_hand and left_arm overlays","n/a","caught before VLM inference","right hand rendered at foot and left arm at face","overlay engineering","PromptHMR body_joints is OpenPose BODY-25; native SMPL-X indices mapped index21 to left heel and index16 to left eye","fixed and regression-tested","data/gate_v0/gate_v0c_predicted_overlays_preflight_v2.json","Wrong v1 artifact retained; no scientific result used it"
F-V14-G0,"Hi4D","basketball28/000047;basketball28/000067","Camera04","basketball","mixed","neutral-evidence over-abstention","Qwen3-VL evidence decomposition","4 balanced dev edges","n/a","all consensus uncertain","decision coverage 0; prompt agreement 0.5","VLM evidence interface","Removing final decision reduces polarity bias but model collapses to depth ambiguity or prompt-dependent gap","do not lower thresholds; require outcome-level test","audit/gate_v0d/qwen_prompt_dev_v1_audit.json","talk23 audit remained untouched"
F-V15-G0,"Hi4D","basketball28/000067","Camera04","basketball","candidate surfaces occluded","empty-pointer hallucination","Qwen3-VL with predicted surface render","right_shoulder-right_forearm and left_shoulder-right_forearm","n/a","opposite deterministic decisions across prompts","surface_meeting versus visible_gap while queried visible-surface pixel count is zero","VLM visual grounding","Model reports confident relation even when z-buffer exposes no queried candidate surface","current verifier branch closed","audit/gate_v0e/qwen_prompt_dev_v1_audit.json","Surface rendering used predicted PromptHMR only; no GT contact or masks"
F-B04-G0,"CHI3D versus Hi4D","Gate-A2 reserve18 versus V0d four-event bank","mixed","multiple","n/a","utility-evidence cohort misalignment","data pipeline","no shared candidate edge","n/a","geometry+VLM prohibited","exact event overlap 0","Gate-B data alignment","Optimizer outcomes and offline VLM evidence belong to different datasets and cannot be paired by action semantics","generate Hi4D optimizer outcomes for the exact frozen VLM candidates","outputs/gate_b_pilot_v1/geometry_evaluation.json","Fail-closed audit prevented a spurious semantic-residual result"
F-B05-G0,"Hi4D","basketball28/000067","Camera04","basketball","mixed","harmful counterfactual contact intervention","balanced fixed candidate bank","A.left_shoulder-B.right_forearm","n/a","O1 applied","GT contact distance +47.34mm; pair-PA +19.95mm; relative pelvis +96.42mm","contact optimizer input","A semantically false coarse edge can strongly distort pair layout despite a safety-aware optimizer","outcome-supervised use/abstain required","outputs/gate_b_hi4d_pilot_v1/evaluation.json","Single prompt-development event; mechanism case not a population estimate"
F-B06-G0,"Hi4D","basketball28/000047;basketball28/000067","Camera04","basketball","mixed","aligned VLM evidence has zero useful coverage","Qwen V0d and V0e fixed verifier","4 balanced candidates","n/a","all abstain","1 helpful 2 neutral 1 harmful outcomes all receive uncertain consensus","VLM evidence interface","Valid identity-compliant JSON does not separate downstream utility under the frozen fixed rule","expand event-balanced outcome cohort before any learned residual test","audit/gate_b/hi4d_promptdev_vlm_residual_v1.json","Prompt disagreement is event-confounded and cannot be claimed as harm prediction"
F-B07-G0,"Hi4D","DEBUG20 aggregate","Camera04","mixed","mixed","candidate bank semantic collapse","Qwen","32/35 non-GT","n/a","No-Go","28 false harmful and 4 false helpful","candidate generator","Natural proposer bank is unsuitable as the only mechanism cohort","repair with dual natural and balanced diagnostic banks","audit/gate_b/hi4d_debug20_failure_attribution_v1.json","Do not alter frozen natural-bank result"
F-B08-G0,"Hi4D","DEBUG20 aggregate","Camera04","mixed","mixed","utility target dominated by optimizer safety","mixed","35 edges","n/a","No-Go","16 safety-only plus 11 contact-and-safety harmful","outcome definition","Flat label asks VLM semantics to predict optimizer dynamics","repair with vector outcomes and factorized heads","audit/gate_b/hi4d_debug20_failure_attribution_v1.json","Only one harmful case is contact-only"
F-B09-G0,"Hi4D","DEBUG20 aggregate","Camera04","mixed","mixed","sequence-level effective sample collapse","mixed","35 edges","n/a","No-Go","13 events from 5 sequences; largest sequence 42.9%","evaluation split","Event-grouped folds share sequence context","future split at sequence or subject-pair level","audit/gate_b/hi4d_debug20_failure_attribution_v1.json","Current No-Go remains valid; point estimates may be optimistic"
F-B10-G0,"Hi4D","DEBUG20 aggregate","Camera04","mixed","mixed","VLM feature redundancy","Qwen","35 edges","n/a","No-Go","10 declared VLM features have rank 7","feature interface","Visibility equals frequency and uncertainty rate is constant","compress evidence and test semantic axis only","audit/gate_b/hi4d_debug20_failure_attribution_v1.json","No post-hoc feature selection on DEBUG20"
F-B11-G0,"Hi4D","dance27/000061;dance27/000120","Camera04","dance","mixed","semantic false but global proxy helpful","mixed","4 edges","n/a","audit","large global contact and pair-geometry improvements","utility target","instrumental constraint versus candidate-local proxy exploitation unresolved","compute local desired and induced-false contact outcomes","audit/gate_b/hi4d_debug20_failure_attribution_v1.json","Must be resolved before innovation-one selector training"
F-B12-BR0,"Hi4D","dance27/000061;dance27/000120","Camera04","dance","mixed","semantic false instrumental constraints","frozen Qwen","4 edges","n/a","vector audit","local gains 50.43-182.34mm and global GT-contact gains 198.75-244.55mm without any edge entering 50mm","flat outcome definition","Old helpful label reflects safe global utility but not contact truth; no evidence of induced false contact in these four cases","factorize semantic local global and safety axes; add pre-outcome nearest hard-negative stratum in BR1","audit/gate_br0/hi4d_debug20_vector_outcomes_v1.json","All four cases are development-only and come from dance sequences"
F-B13-BR1,"Hi4D","football18/000090;game22/000044","Camera04","football;game","mixed","delayed optimizer safety violation","frozen three-stratum bank","3 edges","n/a","pause_and_diagnose","K100 misses 3 of 7 K1000 safety violations","short response probe","Candidate-local and global gains are predictable at K100 but penetration risk can emerge after K100","audit intermediate safety trajectories and conservative bounds before any selector","audit/gate_br1/hi4d_train12_probe_response_v1.json","K100 safety recall 0.571 below frozen 0.80; two misses are game22 penetration max"
F-B14-BR1S,"Hi4D","train12 aggregate","Camera04","mixed","mixed","trajectory-bound false alarms","frozen three-stratum bank","36 edges","n/a","require intermediate checkpoint","last-segment bound catches 6/7 but flags 14 total","sparse response probe","Log-slope improves recall but extrapolation is too imprecise for selective safety","run one frozen all-candidate K250 direct checkpoint; stop expansion if it fails","audit/gate_br1_safety/hi4d_train12_delayed_safety_diagnosis_v1_1.json","Gradient conflict is promising descriptive evidence but incomplete and not independently validated"
F-I1-15-CONF,"Hi4D","train15 independent confirmation aggregate","Camera04","mixed","mixed","gradient-compatibility confirmation failure","frozen three-stratum bank","45 edges","fixed finite cosine>=0","reject","accepted5/45; harmful1; safety violation1; helpful retained2/13","gradient abstention policy","Development zero-threshold mechanism does not transfer across sequence-disjoint unseen subject pairs","reject rule; no confirmation threshold tuning","audit/gate_i1/hi4d_train15_gradient_compat_confirmation_v1.json","Conditional PCGrad and matched-coverage baselines correctly remain unrun"
F-I1-16-V3,"Hi4D","game22/000044","Camera04","game","mixed","off-target collision witness missed by sparse envelope","distance-matched negative","A49-B8","n/a","reject V3","penetration max increases by more than frozen10mm margin; maximum depth86.502mm","counterfactual off-target envelope","Maximum B vertex4807 belongs R75 3/17/72 and nearest A vertex3289 belongs R75 61; neither is a target patch; 10 samples per region missed it","requires a newly preregistered coverage-complete collision certificate; do not tune V3 sampling","audit/gate_i1/hi4d_br1_11_2_v3_max_intersection_patch_v1.json","V3 repaired BR1-10-2 and kept BR1-11-0 safe but failed the all-cases gate"
F-I1-17-R4,"Hi4D","football18/000090","Camera04","football","mixed","evaluator-congruent safety certificate exceeds action cost","distance-matched negative","BR1-10-2","n/a","reject R4 efficiency","TwinGuard341.535s versus paired O0+O1 192.746s; fraction1.77195","full-mesh per-step certificate","Exact bidirectional winding checks dominate runtime and are more expensive than observing both endpoints","future efficient BVH/SDF work may change cost but cannot retune current frozen R4","audit/gate_i1/hi4d_twinguard_r4_runtime_stop_v1.json","Formal matrix stopped after first candidate per preregistration"
F-I1-18-CIU,"Hi4D","CIU pilot aggregate","Camera04","mixed","mixed","semantic truth and intervention utility mismatch","natural plus balanced","116 edges","n/a","benchmark finding","semantic true:13 helpful 8 neutral 9 harmful; semantic false:27 helpful 13 neutral 46 harmful","evaluation target","Contact correctness does not determine the causal outcome of applying the constraint to a fixed optimizer","requires utility-aware evaluation and a new untouched test for any future method","audit/ciu_bench/ciu_bench_pilot_v1.json","Composition-dependent pilot proportions; not a prevalence claim"
F-CIU-F2-ABSTAIN,"Hi4D","basketball28/000072;fight12/000064;football18/000090;highfive13/000126","Camera04","basketball;fight;football;highfive","mixed","natural proposer stable empty output","Qwen3-VL-8B-Instruct five-sample proposer","no edge","n/a","abstain","5/5 empty edges for each of4 events","VLM semantic contact sensor","A fixed per-event four-edge budget is incompatible with an honest VLM that may report no supported contact","preserve explicit abstention slot and report proposal coverage; never substitute a geometry-selected edge","audit/ciu_f2/ciu_f2_dev_natural_proposal_freeze_v2.json","Construction finding only; not a population VLM recall or abstention-rate estimate"
F-CIU-VAL-OP,"Hi4D","CIU-VAL-ID aggregate","Camera04","dance;fight;highfive;hug;kiss;pose","mixed","safe ranking but frozen operating-point retention failure","Qwen proposer plus DEV geometry risk head","11 proposals","two lowest predicted risks","strict fail","accepted2:0 harmful 1 helpful 1 neutral;helpful retention25% below40% although all-set harm54.5%","selector operating policy","floor of nominal25% on11 proposals yields only2 accepts;ranking AUC0.933 and posthoc top3 remains harm-free but is not confirmatory","freeze a minimum-count or calibrated-risk policy before TEST using VAL only;never relabel this Gate","audit/ciu_val_confirmation/closure_v1.json","TEST untouched;strict failure is preserved"