Algorithmic Evidence Processes

This project develops a framework for statistical inference based on repeated randomized executions of an algorithm applied to a fixed dataset. Instead of modeling the data-generating distribution directly, the framework treats the stochastic algorithm as inducing a distribution over outputs.

Together with Claudia Tarantola, I use sequential likelihood-ratio-type evidence processes to compare output distributions under competing hypotheses. The resulting process does not accumulate new population information. It quantifies how strongly the perturbation-induced output law associated with the observed dataset aligns with alternative inferential claims.

This perspective gives a formal language for stability, reproducibility, and algorithmic uncertainty in settings such as variable selection, feature importance, model comparison, and meta-analysis. Depending on calibration, the process can target directional drift, finite-threshold stability, or exact finite-sample validity through e-value-style supermartingale control.

Current manuscript: non-peer-reviewed working paper.