VALIDATION DATA
Proof that the system is tested, not just described. Reference cases, version comparisons, accuracy summaries, and reproducibility logs.
Canonical Examples
Seven reference cases with known ground truth. These are the benchmark runs against which all new analyses are compared.
Canonical reference model. 27" waist, 40" seat, dual-sphere FSCWG silhouette. Input: 4 high-resolution images (front, side, rear, 3/4 view). Ground truth annotated from physical measurements and 3D scan.
Below-average projection depth (10.2"). Tests system behavior at the lower bound of the FSCWG range. Expected: lower PD percentile, higher TDI (more uniform tension), ZGC pass on 511 but gap on 501.
Above-average projection depth (18.7"). Tests upper bound behavior. Expected: high PD percentile, lower TDI (stress concentration), ZGC fail on 501, pass on 550 with custom adjustment.
Same subject as Case 001 but with only a front-view image. Expected: GFS drops to 0.72, PD confidence interval widens to Β±1.5", CUS flagged below 0.60.
Version Comparison: v6.0 β v7.0
How key metrics and system behavior changed between versions. Deltas are computed on the shared holdout set (N=247).
| Metric | v6.0 Behavior | v7.0 Behavior | Delta | Notes |
|---|---|---|---|---|
| Overall Score | Opaque "Perfecto Score" (0β100) | Transparent "Composite Score" (0β100) with weighted components | + Transparency | Formula fully documented in Methodology |
| Projection Depth | Reported as absolute value only | Absolute value + confidence interval + percentile rank | + Uncertainty | CI computed from image quality and pose coverage |
| Curvature | Single scalar "Curvature Score" | Full 12-angle Curvature Radius Profile with error bars | + Resolution | Enables detection of localized asymmetry |
| Tension | Qualitative "Tension: Optimal" label | Quantitative TDI (0β1) + Tension Distribution Index | + Quantification | Enables comparison across denim SKUs |
| Confidence | Not displayed | CUS displayed on every metric + overall | + New | Below 0.60 triggers visible warning |
| Provenance | Run ID only | Full provenance: version, timestamp, input quality, model version, consent state | + Auditability | Required for reproducibility |
| Reproducibility | Not measured | RS computed on every run (Nβ₯5 repeated runs) | + New | RS displayed in validation report |
| Calibration | Not tracked | CE computed on holdout set, updated each release | + New | Baseline CE: 0.03 (Composite Score) |
Accuracy Summary
Calibration Error and Reproducibility Score computed on the v7.0 holdout set (N=247).
Known Edge Cases & Failure Modes
Documented failure modes with example inputs. These are actively monitored for model improvement.
Single-View Input
GFS drops below 0.75, PD CI widens beyond Β±1.0", CUS flagged. System recommends multi-view re-upload. Example: Case 004.
Extreme Asymmetry
Sphere fit residual exceeds 95th percentile of training distribution. GFS degrades gracefully but CRP shows high variance at lateral angles. Example: 3Ο asymmetry case (N=12 in holdout).
Uncalibrated Denim SKU
TDI and ZGC use generic fabric estimates with expanded confidence intervals. System flags "uncalibrated SKU" on report. Example: Vintage Levi's 501 (pre-1990) with unknown stretch %.
Rapid Motion Sequences
MSS trained on five standard motions. Athletic or dance movements produce uncalibrated scores. System displays "motion out of distribution" warning.
Reproducibility Log
10 repeated runs on Case 001 (Vicente Noriega) with identical inputs and model version. All runs used the same random seed for stochastic components.
Explore the Full System
Read the methodology, run the Live Study Room, or check the changelog for v7.0 details.