Projection and Model Evidence
Understand projection eligibility, chronological testing, calibration gates, and why unsafe probability outputs are withheld.
Current athlete-facing rule
StrideSignal does not show a long-term linear forecast, next-PR mark, PR chance, or A/B-standard probability. Those outputs are withdrawn until adequate matching data demonstrate calibration.
Early short-term estimate
The Short-term Projection page uses every positive non-excluded stored performance for the selected event. Exact-date marks enter rolling-origin candidate selection. Coarse-date marks enter the robust all-history anchor without fabricating an exact date. One-result and sparse estimates are shown as preliminary with wider uncertainty. Forecasts stop at 90 days and label the range model spread, not probability.
Model evidence page
Candidate models are tested chronologically for technical review. Earlier results train a candidate and later unseen results measure error and interval coverage. One or more positive non-excluded stored performances can produce a preliminary next-performance estimate and descriptive model spread. The 80% and 95% probability-interval labels remain withheld unless the selected model has at least eight held-out tests, observed 80% coverage between 72% and 88%, and observed 95% coverage of at least 90%.
Exact dates and canonical inputs
Season-only dates stay out of elapsed-time fitting but contribute to the robust all-history anchor; no exact day is fabricated. Exact duplicate source assertions count once. Classification changes, uncertain conditions, and other data-quality concerns remain visible and down-weighted. Only another event, an invalid mark, an exact duplicate assertion, or the user's explicit exclusion removes a row.
Why probability is withheld
A numerical probability must be calibrated: events predicted at 50% should occur about half the time in adequate comparable samples. The current evidence does not support that claim for PRs or standards. Stored legacy predictions remain visible for audit, but their former probability and confidence labels are marked withdrawn.
Important limitation
A projection is an association under conditions similar to the analyzed data. It is not a target, guarantee, training-plan effect, readiness measure, health statement, or cross-class equivalence. Long-term development scenarios remain unavailable until class-appropriate governed population data and acceptable rolling-origin or prospective validation exist.
Model review version: 2026.08.01-5.