A forecaster who says 70% should be right about 70% of the time across all their 70% forecasts. That property is calibration, and it is entirely separate from accuracy. A forecaster can be well calibrated and rarely confident, or highly accurate on easy questions and badly calibrated on hard ones.
Brier = Σ (forecast probability − actual outcome)²
Lower is better. Forecasting 0.9 for something that happens scores 0.01; forecasting 0.9 for something that does not scores 0.81. It rewards confidence when justified and punishes it heavily when not.
A Brier score alone means little without a baseline. The relevant comparison is against always forecasting the base rate — a model that cannot beat "always say 50%" is adding nothing, however sophisticated. VILIQ reports the model score, the uniform baseline, and the difference as a skill figure, always with the sample size.
Common belief
"The model was right 8 times out of 10, so it is good."
What is actually true
Ten forecasts is far too small a sample to distinguish skill from luck, and hit rate ignores confidence entirely — being right at 51% and right at 99% count identically. Calibration over a meaningful sample is the measure that survives scrutiny.
This is why every VILIQ forecast is written to an append-only ledger with its outcome definition fixed before the outcome is known, and why wrong predictions are never deleted. A model that can quietly drop its misses cannot be calibrated, and an uncalibratable model is an opinion with extra decimal places.