| { |
| "schema_version": 1, |
| "scope": "14 answer-correct rollouts from ViSTR skill_val (baseline / perception mode, qwen3-vl-plus)", |
| "annotator": "gaozhe (manual notes 2026-08-19) + reviewer cross-check", |
| "label_definitions": { |
| "process_false_positive": "answer correct but the reasoning depends on a false or fabricated premise", |
| "process_ok": "answer correct and the cited evidence supports the conclusion", |
| "gray_zone": "answer correct, no outright false premise, but the rubric is unverifiable / not generalisable, or annotators disagree" |
| }, |
| "confidence_definitions": { |
| "high": "explicit manual note for this case", |
| "derived": "inferred from a manual note covering a cluster of cases", |
| "reviewer_only": "no manual note; reviewer judgement, needs a second annotator" |
| }, |
| "counts": { |
| "process_false_positive": 5, |
| "gray_zone": 3, |
| "process_ok": 6 |
| }, |
| "caveats": [ |
| "Single primary annotator; inter-annotator agreement not yet measured.", |
| "n=14 supports existence and taxonomy claims, not stable rates.", |
| "Reviewer and annotator initially disagreed on #148 (reviewer wrong) and #57 (reviewer stricter), so labels are revisable." |
| ], |
| "annotations": [ |
| { |
| "sample_id": "148", |
| "label": "process_false_positive", |
| "severity": "severe", |
| "confidence": "high", |
| "error_layer": "perception_structural", |
| "human_note_zh": "假阳性：“The block being removed is in the middle where it likely supports significant weight above it”这句话纯扯淡，Jenga稳定性取决于目标块抽走之后，剩余结构能不能稳定。首先这个块在高度的middle 但不在横层意义上的middle，其次抽走后，该层只剩一个边上的块了，所以无法支撑。", |
| "false_claim": "The block being removed is in the middle where it likely supports significant weight above it", |
| "why_wrong": "Jenga stability depends on whether the REMAINING blocks in that layer can support the load. The target is mid-height but off-centre within its layer; after removal only one edge block remains in that layer.", |
| "reviewer_crosscheck": "Frames confirm the hand pulls an off-centre block. The solver applied a generic 'tall tower + gaps + middle block' template with no layer-level structural analysis. Reviewer initially mis-rated this as sound; the manual note is correct." |
| }, |
| { |
| "sample_id": "156", |
| "label": "process_false_positive", |
| "severity": "moderate", |
| "confidence": "high", |
| "error_layer": "reasoning", |
| "human_note_zh": "假阳性，但没148严重。“塔身没有出现摇晃或不稳定的迹象，并且在整个视频序列中塔身都保持着其结构，因此我预测塔身将保持稳定”扯淡了。但是“剩余块体的排列和平衡”这个rubric是对的。", |
| "false_claim": "no wobbling observed during the video, therefore the tower will remain stable after removal", |
| "why_wrong": "The video ends before the block is fully removed, so absence of wobble carries no information about post-removal stability. The valid rubric is the arrangement/balance of the remaining blocks.", |
| "reviewer_crosscheck": "Confirmed. The solver also cited 'alternating colour pattern implies good weight distribution', which is not a load-bearing property." |
| }, |
| { |
| "sample_id": "110", |
| "label": "process_false_positive", |
| "severity": "moderate", |
| "confidence": "high", |
| "error_layer": "reasoning", |
| "human_note_zh": "ai 自己总结的rubric 不对，导致是假阳性。同样一套rubric 用在 #109 就错了。", |
| "false_claim": "centroid drift across frames therefore the vehicle shows subtle movement", |
| "why_wrong": "The rubric has no ego-motion compensation, so apparent centroid drift conflates camera motion with target motion. It is not generalisable.", |
| "reviewer_crosscheck": "Empirically confirmed on #109 (same subtask, same rubric, answer wrong). Raw tracks: #110 obj_0 dx=+0.287 (large, genuinely moving) vs #109 obj_0 dx=-0.051 with a global mean of -0.007 across 7 tracked vehicles, i.e. a small residual the rubric cannot resolve." |
| }, |
| { |
| "sample_id": "23", |
| "label": "process_false_positive", |
| "severity": "severe", |
| "confidence": "high", |
| "error_layer": "data_fabrication", |
| "human_note_zh": "这几个 case 感觉完全就是幻觉了，“目标球确实进入了袋口”，显然没有。其次问题问的是将来时的“会不会”，他给答成完成时了。边界框小+落到球袋 这个完全是幻觉rubric了，可执行性也贼差。", |
| "false_claim": "the ball's bbox becomes very small and it is positioned at the corner, i.e. it disappeared into the pocket", |
| "why_wrong": "Fabricated: obj_4 bbox area is 0.00076 at the last frame vs 0.0007 at the first (no contraction); final centroid (0.733, 0.862) is not a corner; visibility stays 1.0 through the final frame. A 4-frame tracking gap (frames 20-23) is followed by 49 more tracked frames (24-72), so nothing was pocketed. Also a tense error: the task asks for a prediction, the solver reported a completed event.", |
| "reviewer_crosscheck": "Verified against workspace tracks/tracks.json. Additionally the tracking JSON the solver received was TRUNCATED (51KB mid-section, no opening brace), so part of its cited data was never visible to it." |
| }, |
| { |
| "sample_id": "26", |
| "label": "process_false_positive", |
| "severity": "moderate", |
| "confidence": "derived", |
| "error_layer": "data_fabrication", |
| "human_note_zh": "（覆盖于“# 23-29 这几个 case 感觉完全就是幻觉了”这条批注）", |
| "false_claim": "balls clustered at the pocket location in the final frames imply the target ball goes in", |
| "why_wrong": "Same hallucinated rubric family as #23 (proximity/size treated as pocket entry) plus the future-vs-completed tense conflation.", |
| "reviewer_crosscheck": "Frame inspection is real, but the pocket-entry inference rests on the same unvalidated rubric." |
| }, |
| { |
| "sample_id": "27", |
| "label": "gray_zone", |
| "severity": "unclear", |
| "confidence": "derived", |
| "error_layer": "reasoning", |
| "human_note_zh": "（“# 23-29 这几个 case”按 id 区间覆盖了 27，但该条批注的具体指控针对答 Yes 的案例；27 答 No。需人工确认。）", |
| "false_claim": "pending explicit confirmation", |
| "why_wrong": "Manual note covers the billiards cluster as hallucination-prone; the specific complaint (claiming the ball went in) does not apply to this No answer. Needs a second pass.", |
| "reviewer_crosscheck": "Our tool-grounded critic flagged direction errors in obj_0/obj_4 motion claims and a target-ball misidentification, which supports a FP label, but this was not independently confirmed by the human annotator." |
| }, |
| { |
| "sample_id": "63", |
| "label": "process_ok", |
| "severity": null, |
| "confidence": "high", |
| "error_layer": null, |
| "human_note_zh": "模型知道用 tracking 去间接估计位姿，这很好，这是我人都想不出来的，说明有自进化的潜力。但当相机消失时，model 是不是可以用新的tracking 标记物去指示自己的方位呢？（当然 这不构成指导意见）", |
| "false_claim": null, |
| "why_wrong": null, |
| "reviewer_crosscheck": "Uses the exit-edge rule (target leaves via the left edge then vanishes therefore back-left). Consistent with the tracking data." |
| }, |
| { |
| "sample_id": "57", |
| "label": "process_ok", |
| "severity": null, |
| "confidence": "high", |
| "error_layer": null, |
| "human_note_zh": "建议同 63", |
| "false_claim": null, |
| "why_wrong": null, |
| "reviewer_crosscheck": "Multi-evidence convergence (bbox shrink + centroid to the left edge + depth). One overstatement: 'washer remains on the left throughout' while the track starts at centroid x=0.837 (right side). Numbers cited are imprecise (final width 0.016 vs actual 0.005) but the trend and conclusion hold." |
| }, |
| { |
| "sample_id": "59", |
| "label": "process_ok", |
| "severity": null, |
| "confidence": "high", |
| "error_layer": null, |
| "human_note_zh": "建议同 63", |
| "false_claim": null, |
| "why_wrong": null, |
| "reviewer_crosscheck": "Sound, though it announced a frame-sampling verification step and then concluded without doing it." |
| }, |
| { |
| "sample_id": "52", |
| "label": "process_ok", |
| "severity": null, |
| "confidence": "high", |
| "error_layer": null, |
| "human_note_zh": "建议同 63", |
| "false_claim": null, |
| "why_wrong": null, |
| "reviewer_crosscheck": "Last-seen-side plus camera-egress reasoning, consistent with the data." |
| }, |
| { |
| "sample_id": "120", |
| "label": "gray_zone", |
| "severity": null, |
| "confidence": "high", |
| "error_layer": null, |
| "human_note_zh": "推理合理，但最终还是纯借助了 VLM 本身的判断能力，而没有形成一套可验证的rubric。", |
| "false_claim": null, |
| "why_wrong": "Conclusion follows from frame comparison, but no quantified, reusable criterion was produced (how much displacement counts as faster, how perspective is handled).", |
| "reviewer_crosscheck": "Agreed: correct answer, low rubric executability." |
| }, |
| { |
| "sample_id": "81", |
| "label": "process_ok", |
| "severity": null, |
| "confidence": "reviewer_only", |
| "error_layer": null, |
| "human_note_zh": "（人工核验笔记未覆盖）", |
| "false_claim": null, |
| "why_wrong": null, |
| "reviewer_crosscheck": "Real tracking data plus frame cross-check. Pixel conversions are wrong (claimed 8px/37px vs actual ~15px/~47px) but direction and magnitude class are unaffected." |
| }, |
| { |
| "sample_id": "96", |
| "label": "process_ok", |
| "severity": null, |
| "confidence": "reviewer_only", |
| "error_layer": null, |
| "human_note_zh": "（人工核验笔记未覆盖）", |
| "false_claim": null, |
| "why_wrong": null, |
| "reviewer_crosscheck": "Under-reports the displacement magnitude by ~10x (claims final x=0.419, actual 0.384), but the error direction strengthens rather than undermines the 'movement exists' conclusion." |
| }, |
| { |
| "sample_id": "71", |
| "label": "gray_zone", |
| "severity": "reviewer_flagged", |
| "confidence": "reviewer_only", |
| "error_layer": "reasoning", |
| "human_note_zh": "（人工核验笔记未覆盖）", |
| "false_claim": "the object drifted right-to-left and exited, therefore it is behind and to the RIGHT of the camera", |
| "why_wrong": "Under pure translation, exiting via the left edge implies back-LEFT. The mapping contradicts #63, which derived back-left from the same evidence pattern. Answer happened to match ground truth (camera likely yawed).", |
| "reviewer_crosscheck": "Tracking quality is also poor: 62/408 frames have the target, final-frame confidence is the -10000 sentinel. Needs human adjudication." |
| } |
| ] |
| } |
