Model Accuracy

What accuracy means for probabilistic predictions

DimensionsPot returns a statistical estimate for each dimension, not a measurement. Accuracy has two components:

1. Point accuracy — how close the predicted value is to the true measurement, on average. Expressed as Mean Absolute Error (MAE).
2. Interval calibration — whether the 95% prediction interval (range_95) actually contains the true measurement 95% of the time. A well-calibrated API is one where a stated 95% PI contains the true value in at least 95% of cases.

Both are validated separately. Point accuracy tells you how far off a typical prediction is. Interval calibration tells you whether the API is honest about its uncertainty.

Validation framework

API v1.4.0 (model adult_ridge_v4.0) was validated before deployment across 7 test tracks totalling 5,200+ API calls against ground-truth measurements from three independent public datasets. No cherry-picking: precision thresholds were set before the runs, and every track is included in the report.

Server errors across all 5,200+ calls: 0.

Selected dimension performance — height + mass input

Under the most common production condition (height and mass only, no additional anchors), MAE and 95% PI coverage for selected BONE dimensions (Track 1, n=300):

FLESH dimensions (circumferences) carry higher MAE by design: soft-tissue volume is not fully determined by height and mass alone. Supplying circumference anchors directly reduces FLESH MAE — see Anchor Strategy for the precision impact of additional inputs.

NHANES civilian holdout (Tracks 3–4)

ANSUR II is a military dataset of athletic adults. To assess generalization to a civilian population, a separate holdout of 500 NHANES subjects (Track 3) and 300 with random anchors (Track 4) was run.

Note on NHANES landmark differences: NHANES measures upper_arm_length from a different anatomical landmark than ANSUR II, producing a systematic ~50 mm offset. The engine produces values consistent with ANSUR II / ISO 7250-1 landmark definitions. If your application requires NHANES-convention measurements, apply a fixed 50 mm correction to upper_arm_length outputs.

Effect of circumference anchors — Track 6

When clients supply circumference measurements alongside height and mass (PRIMARY_RICH tier), the engine constrains soft-tissue predictions directly rather than estimating them from body mass.

With 5 circumference anchors, 34 of 56 tested dimensions improve and zero degrade for athletic and civilian body types. Additional input signal propagates correctly through the prediction pipeline — it does not introduce regression artifacts.

Reproducibility

All results above were generated by tests/precision_validation_v4.py (RANDOM_SEED=42). Three independent runs — local development, local post-cleanup, and production Google Cloud Run EU — produced numerically identical output across all MAE, bias, and coverage metrics (0.0 mm delta). The validation is not environment-specific; it is a property of the API itself.

Limitations