Known biases

• Selection bias: Century Worth forecasts subjects with enough evidence to support a falsifiable public prediction, not a random sample of all companies or AI events.
• Coverage bias: tracked entities and source classes are overrepresented; missing coverage is not evidence that an outcome did or did not occur.
• Grading lag: unresolved, indeterminate, invalidated, and superseded-before-resolution forecasts are excluded from hit rate and calibration and are reported separately.
• Small and correlated samples: buckets may contain few forecasts, and multiple forecasts can respond to the same underlying event, so percentages are descriptive rather than independent performance guarantees.
• Methodology and correction effects: calibration retains the probability and methodology captured at publication and uses the latest append-only grade correction without rewriting earlier history.