Changelog

2026-11-04

Switched the ECI to a Bayesian model. Model capabilities, benchmark difficulties and slopes are now fit jointly with a Bayesian model, replacing non-linear least squares with a ridge penalty. Uncertainty is now reported as 90% credible intervals, instead of 90% bootstrap confidence intervals from 500 resamples. We now include a model only if at least four of its benchmark scores fall between 0.1 and 0.9 after normalization, replacing the previous rule of dropping models with fewer than four benchmark scores. Domain-specific ECIs have not yet been refit with the Bayesian method: they still use least-squares domain estimates combined with samples from the general ECI fit. See the methodology and our report on the update for details.

2026-05-04

ECI documentation published to epoch.ai.