Enciclopedia Boulder v1.0

Inicio · Capítulos · 14. Señales para visión artificial

15.237 14.8.27 Pruebas de aceptación

ID Prueba Esperado
CV148-A01 All 310 movement and 34 failure records remain structurally unchanged. PASS
CV148-A02 All 14.1–14.7 submodels remain structurally unchanged. PASS
CV148-A03 All 344 classifier labels resolve to one class evaluation profile. PASS
CV148-A04 Component and end-to-end evaluation are represented separately. PASS
CV148-A05 Test data are grouped and quarantined before clip/frame generation. PASS
CV148-A06 A benchmark profile freezes metrics, tolerances and gates before test access. PASS
CV148-A07 No single opaque aggregate score is required for approval. PASS
CV148-A08 Every primary metric reports support and uncertainty. PASS
CV148-A09 Dependent frames are not treated as independent statistical replicates. PASS
CV148-A10 Model comparisons use paired test units. PASS
CV148-A11 Pose metrics keep aligned and unaligned errors separate. PASS
CV148-A12 Scene segmentation includes panoptic, part and temporal identity metrics. PASS
CV148-A13 Contact evaluation distinguishes touch, support and force. PASS
CV148-A14 Temporal segmentation reports edit and overlap metrics, not frame accuracy alone. PASS
CV148-A15 Kinematic errors use native units and preserve filter/provenance. PASS
CV148-A16 All 28 high-priority contrast sets have dedicated evaluation. PASS
CV148-A17 Open-set, abstention, risk-coverage and calibration are evaluated. PASS
CV148-A18 Conformal coverage is reported with set efficiency and group breakdowns. PASS
CV148-A19 Distribution shift includes athlete, problem, gym, device and time views. PASS
CV148-A20 Robustness includes occlusion, timing, depth, calibration and sensor dropout. PASS
CV148-A21 Missing modality degrades safely or abstains rather than silently hallucinating. PASS
CV148-A22 Latency is compared only at a prespecified quality floor. PASS
CV148-A23 Run schema captures code, model, container, data, split and preprocessing hashes. PASS
CV148-A24 Crashes, timeouts and unscorable attempts remain in the denominator policy. PASS
CV148-A25 End-to-end errors are attributable to upstream stages or irreducible uncertainty. PASS
CV148-A26 Ablations are prespecified and evaluated on paired units. PASS
CV148-A27 Reproducibility is tested through immutable rerun artifacts. PASS
CV148-A28 Field deployment claims require a prospective frozen pilot. PASS
CV148-A29 Coach utility claims require a separate human study. PASS
CV148-A30 Privacy, consent and withdrawal remain release blockers. PASS
CV148-A31 Unsupported classes and insufficient-support results are explicitly published. PASS
CV148-A32 This section defines a benchmark protocol and does not claim it has been executed. PASS

15.237.1 Resultado estructural de esta entrega

Comprobación Valor Estado
Movimientos preservados 310 PASS
Patrones de fallo preservados 34 PASS
Perfiles de clase 344 PASS
Tareas 18 PASS
Tracks 12 PASS
Vistas de dataset 14 PASS
Condiciones de robustez 24 PASS
Métricas end-to-end 24 PASS
Decision gates 20 PASS
Pruebas 14.8 32 PASS
Fuentes resueltas True PASS
14.1–14.7 preservados True PASS