Mean absolute error across capture-held-out folds.
Driftistan R&D / Public evidenceMeasured.
Measured.
Not mocked.
Review versioned aggregate outputs, real field telemetry and capture-held-out validation while Driftistan's proprietary model code, weights and training data stay protected.
Current evidence release
What has actually been trained.
Every value below is generated from the versioned training metadata. A new training run refreshes the protected lab, this page, the public JSON and the documentation together.
Raw accuracy-field baseline on the same held-out samples.
Lower MAE than the comparison baseline.
8 complete capture groups.
Interactive model evidence
Compare real model scenario runs.
Each result was generated by the private shadow-only tree runtime during evidence publication. The public lab exposes reproducible inputs and outputs without exposing the proprietary model, and it never controls navigation.
Loading versioned model evidence…
Latest verified telemetry snapshot
Real field evidence behind the result.
Generated from the exact consented, complete capture groups accepted into the current training release. It refreshes with the training pipeline rather than using marketing counters.
Validated GPS observations used by this version.
Aggregate movement reconstructed from successive eligible fixes.
Combined eligible intervals represented in the dataset.
Typical interval between accepted field observations.
Median Android provider accuracy field.
Provider accuracy at the measured 95th percentile.
Measured snapshot · refreshed 8 Aug 2026
Coordinates, route geometry, capture identifiers, device fingerprints, feature vectors and model internals are excluded.
Validation design
Whole drives were held out.
Leave-one-complete-capture-out validation reduces leakage from adjacent samples in the same drive. The chart compares aggregate model MAE in cyan with baseline MAE in grey.
Version record
Training updates leave a trail.
The public evidence manifest keeps the latest ten distinct training releases. Documentation is generated in the repository during every successful training run.
OEM and Tier-1 navigation teams
Review automotive fitEvaluate a confidence layer for real vehicle navigation.
Driftistan is building this signal for sensor-quality awareness, confidence-aware map matching, degraded-GPS detection and navigation R&D. The current v1 is an R&D shadow model, not a certified production driving component.