e-values

A local CLI for logging predictions before the outcome and checking whether your "80% sure" comes true 80% of the time.

Runs ~20 options strategies against live market data in shadow mode, records every hypothetical fill under worst/base/optimistic assumptions, and grades each with anytime-valid e-processes. Places no orders.

Repository for the paper "Auditing Pay-Per-Token in Large Language Models", AISTATS'26

This repository contains the code for the paper "Optimizing Social Utility in Sequential Experiments".

Measure your agent harness, find where it wastes the model, and prove the fix worked. Harness-agnostic, agent-agnostic, zero dependencies. Reference implementation of HTP-1.

A kernel-userland protocol enforcing information-theoretic bounds on AI adaptivity leakage, benchmark gaming, and capability spillover.

Real-time cross-chain bridge exploit detection that accuses only with proof. Caught 422/422 attack payouts across Nomad, Ronin and practice attacks with 0 honest payouts accused. Conservation ledger, e-value false-alarm control, Hawkes contagion radar, signers recovered with ecrecover. CSB-01

Certified pairwise leaderboards for retrieval evaluation. WITNESS samples a few fresh human labels at known probabilities and uses official qrels and LLM judges only as variance-reducing predictors: unbiased for any judgment source, false-discovery-rate controlled, with label-budget forecasts. One-command reproduction.

Benchmark for statistically valid AI scientist systems, using audit-closed protocols, transparency logs, and sequential inference to prevent false discoveries in autonomous research agents.

Publicly verifiable proof that an LLM endpoint actually spent the compute you paid for — zero provider cooperation. Closes the effort gap (Hollow-LLM) and the transferability gap (IRIS).

Anonymous code and result release for block-calibrated false-discovery control in open-vocabulary detection

Research implementation of anytime-valid group-invariance tests with kernel, adaptive Gaussian, and neural statistics.

A Bayesian statistics monograph built on de Finetti, Sanov and Bayes: exchangeable prediction under moment constraints gives maximum entropy laws, and testing, shrinkage, asymptotics and decision are read on one moderate deviation scale.

Volume I of a two volume series. Bayesian clinical trials built up from the traditional trial and tied to regulation: calibration, monitoring, safety, borrowing, adaptive designs, and reanalyses of landmark trials, with an R package and an online lab.

MSc thesis: designing e-value evidence around a decision-maker's loss, with an application to Polymarket prices for the 2026 World Cup final.

Windows instrument for fine-tuning runs: every sample drawn, a report that leads with its answer, and a local-model workbench that builds formula tools and tests what each knob does, with e-value evidence across folders.

Anytime valid inference in Python: confidence sequences, sequential tests and e-values

Codebase of Concise and Logically Consistent Conformal Sets for Neuro-Symbolic Concept-Based Models, accepted at NeurIPS2025

Label-free drift attribution (model / noise / world / annotator) with anytime-valid guarantees and a frozen pre-registration — closes the zero-label misattribution gap 0.50→0.00 on the recoverable cause; negatives reported as measured boundaries.

Statistically valid gates for non-deterministic eval suites. Stop early without losing error control.