Prediction Methods

Prediction methods

STRATHEX sends calculations to the STRATHMARK engine chosen for the competition. It does not choose between hidden local predictors.

Active engines

Engine	How it predicts
V2	A hierarchical model combines prior-only personal history with population estimates, calibrated uncertainty and joint mark optimization.
Linux V3 competition	Independent Formula and trained CatBoost ML forecasts, capability updates and earned component weights, followed by complete-field mark calculation.

Linux V3's LLM council is unavailable. The wider three-member council design and local diagnostic experiments are not an active numeric council in this profile. Older preview and Windows rehearsal installations have their own limitations; see Engine selection.

History and wood

V2 sends stable competitor IDs and valid dated history under one exclusive cutoff. Undated, same-day, future and invalid observations are excluded. Quality and same-tournament weighting are accepted but do not change V2 numbers.

V3 freezes evidence for all fields in a round. Valid settled results become eligible at the next round boundary. An advancing field is reconstructed and recalculated, rather than inheriting individual marks.

What to review

A prediction includes more than a time: uncertainty, performance spread, cutoff, engine/model versions, warnings, degraded state and provenance. V2 returns optimizer evidence. V3 retains component forecasts and signed field, review, issue and result receipts.

An explicit V2 manual override remains a judge decision. It is not evidence that an experimental model has been promoted.

What is historical or experimental?

The STRATHEX-side baseline/XGBoost/Ollama selector and QAA interpolation are retired from live v7 prediction. Old reports remain as dated records.

Accuracy Preview is implemented separately. Its candidate is not enabled in competition predictions. For numerical details, read V2 or V3.