FORECAST EVALUATION PROTOCOL
No empirical AI improvement is claimed.

1. Define currency, revenue measure, cohort, forecast horizon and evaluation periods.
2. Freeze baseline and assisted forecasts BEFORE outcomes are known. Record snapshot timestamps.
3. Record model/version, prompt, input availability and human overrides privately.
4. Exclude future information. Use identical periods and cohort for both methods.
5. Populate period,actual,baseline,assisted only from authorised snapshots and realised results.
6. Run evaluate-forecast.py; inspect per-period errors, WAPE, signed bias and MAE.
7. Record missing periods, exclusions and zero-actual cases; retain negative results.
8. Obtain permission before publishing any real company data; publish aggregates with confidentiality review.
9. Report sample size and limitations. Observational differences are not proof of causation.