Period & support
State the data period, available and evaluated instruments, events and independent time blocks. After a history reset, evaluation begins again.
X → Y & comparison
Compare conditional frequency with a benchmark. For ATR and flow, evaluate incremental information beyond range, volume and past volatility. Keep descriptive frequencies separate from calibrated forecasts.
Temporal validation
Fix definitions and training before evaluated outcomes. Purge boundary-crossing windows and account for dependence and multiple searches. Inspected history is never labelled untouched holdout.
Weekly reports
A report is computed once after the UTC week closes and never changes. Only contracts with a full week of daily bars are included; excluded contracts are counted and shown.
Model cards
Model state comes from the periodic review registry. A new model version resets the state until it is reviewed again; reviews run every 14 days.
Public boundary
Publish aggregate findings and clear limitations. Internal company research, model formulas and parameters, source code, keys, user data and infrastructure details stay out of publications.