Release gate passed

Better models.
Evidence to ship.

A geometry classification release, evaluated across clean, noisy, rotated, and scaled point clouds. Every decision traces back to the data and model that produced it.

DRAG TO ROTATEDrag or use the arrow keys to rotate the shape. Change its condition with the buttons below. Faint points show the clean reference.
cube / clean / 256 points
Try the release gate
Would you ship this model?
PASS

All release checks passed.

    Explore two saved evaluations of real model artifacts. Switching updates the evidence below; it does not call a live backend.

    01 Schema→02 Content hashes→03 Grouped holdout→04 Slice policy→05 Atomic promotion
    Candidate accuracy
    Across all four holdout slices
    Baseline improvement
    Percentage points, paired groups
    Independent shapes
    Weakest class / slice
    01 / Robustness

    One score never tells the story.

    Same latent shapes. Four conditions. No shape shared across training and holdout.

    Axis-dependent baselineInvariant candidate
    02 / Error analysis

    See where it misses.

    Rows: actual class · columns: predicted class
    Counts are evaluation samples, not independent shapes.

    03 / Failure drills

    A release gate must know when to stop.

    These drills change real artifacts and re-run the gate. No fabricated model scores.

    REGRESSION BLOCKED

    Aggregate accuracy hides a weak slice

    INTEGRITY BLOCKED

    Edited evidence cannot authorize a release

    RECOVERY VERIFIED

    A previous release is still recoverable

    04 / Distribution health

    Drift is visible, not hidden.

    Candidate features compared with clean training data. Train-only quantile bins; 0.5 pseudo-count smoothing.

    Holdout sliceMean PSIMaximum mean shift

    PSI and these demo thresholds are diagnostic choices for this synthetic fixture. They are not calibrated production or CAD acceptance criteria.

    05 / Release policy

    Explicit checks. A clear decision.

    Artifact lineage

    Canonical JSON SHA-256 binds models, dataset, report, and immutable release metadata. Promotion re-evaluates the exact inputs and compares against the active model.

    What this result establishes

    Real softmax models are trained locally on generated cube, sphere, and cylinder surface clouds. Invariant features improve this seeded synthetic task. The dataset does not contain real CAD assemblies, sensor captures, or industrial defects.

    Confidence intervals resample latent shape groups. This small, deliberately simple benchmark demonstrates release engineering; it does not establish real-world geometric recognition quality.