Are the benchmark numbers real?

The completed study ran hundreds of thousands of evaluation questions across over a thousand independent online environments, with committed datasets, aggregate reports, a data-completeness audit, and hundreds of passing tests. Plots rebuild without further API calls.

So should I always feed posterior uncertainty?

Not always. Adding posterior means cut regret versus counts alone, but adding fuller Beta parameters and intervals then raised regret again. The test covers a representation package, so wording and length are not isolated from uncertainty content.

Can Jev just do Thompson sampling if I name it?

No. Merely naming the strategy did not reproduce that algorithm, and exploration prose helped some choices while still missing states where a lower-mean arm was optimal. Sampling the answer distribution helped more at ten arms than at two.

How do I reproduce without burning budget?

Install the pinned environment and run the pytest and lint checks, then rebuild figures from committed records. Live collection commands take a spending cap, isolate one API writer, and generate reports from local artifacts.

What license is it under, and what stack does it use?

Apache License 2.0, copyright 2026 Thomas J Richner, with a NOTICE file. It is a Python 3.12 project: install the pinned environment with uv sync --frozen --python 3.12, run the suite with pytest, and lint with ruff.

What is the planned next phase of the research?

The continuation plan replicates the strongest findings on fresh tasks and service dates, isolates which uncertainty fields and advice formats change behavior, and tests whether more faithful probability forecasts improve sequential reward. A longer-term direction is a reasoning-model harness with access to Jev, Bayesian algorithms, both tools, or neither.

What did it find about supplying external Bayesian computations?

With exact action values supplied, it observed zero local decision loss across 4,000 sequential decisions and followed supplied Bayes-UCB indices about 99.93% of the time. A recommendation alone was far more sensitive to its format and surrounding evidence, so the benefit depended on how the advice was presented.

CommunityJev Tools & CLIs

Can Jev Bayes

Bayesian decision study probing Jev on stationary Bernoulli multi-armed bandits, testing whether posterior statistics, exploration instructions, and supplied action values improve cumulative reward.

See highlights

Open github.com

Stars: 1
Language: Python
License: APACHE-2.0
Published by: Community
Author: TomRichner
Last verified: Oct 2026

TomRichner/can-jev-bayes

User repository on GitHub

View on GitHub

Jev Bayes, No? Testing TypeSafe AI's Jev against Bayesian-optimal strategies, and testing if Jev can effectivly use Bayesian priors.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

How well can Jev make sequential decisions under uncertainty, and how can Bayesian methods help it learn and act more effectively?

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Compares 2 to 15 arms against Thompson sampling, Bayes-UCB, knowledge gradient, IDS, and exact methods.
Posterior means cut regret versus counts alone while fuller Beta summaries did not uniformly help.
Supplied exact values gave zero local loss and near-total use of supplied Bayes-UCB indices.
Joint Choice versus per-arm Noul trades coherence and cost, with sampling gains by arm count.

Watch out

Findings are specific to short stationary Bernoulli bandits on one model version and do not claim general sequential optimality; live reruns spend API credits under a cap. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-deep-dive

Long-form Jev writeup that maps each conclusion to official docs, vendor evals, integrations, community reproductions, and press, separating contract facts from unverified claims.