About this site

A small site for a specific question.

NanoJev proves that one 0.6B-parameter model can return full probability distributions over actions, with zero output-token decoding. That single capability changes what a "decision" can be — and what a "multi-agent system" can be when many of those decisions compose. This site collects the evidence and the vocabulary around that question.

01What this site is

A static mirror of three recorded replays (Maze, Snake, and a 40-map navigation benchmark) plus four short-form content pages — Compare, Guides, Projects, and this About — that frame those replays inside the broader research arc from single-model decisions to multi-agent systems.

static site recorded evidence research vocabulary open upstream

02What this site is not

Not a chatbot, not a tool-use demo

No live API endpoint, no agent loop you can talk to. The replays are recorded, not generated. You cannot type a question and get an answer; you can scrub through what already happened.

Not a benchmark paper

The numbers in the upstream README (success rates, comparison splits) are reported as-is and linked back. This site does not add new evaluations or claim new state of the art.

Not a product

No signup, no paywall, no telemetry that captures user input. Pageviews only when an operator opts in to analytics.

Not a tutorial

The Guides page is a reading list, not a step-by-step course. It points to papers, READMEs, and runbooks maintained upstream.

03Who it is for

Three audiences find this site useful, in roughly this order:

ML researchers: Comparing small decision models. The recorded replays show three different systems on the same game; the Compare page is for people deciding which baseline matters.
Multi-agent builders: Designing the decision interface between agents. The homepage vocabulary section defines what a "decision agent" actually is; the Projects page lists candidate compositions.
Curious readers: Following the Jev → NanoJev story. The Guides page is the entry point for non-specialists.

04Principles

Structured outputs over free-form text

The site exists because typed probability distributions make multi-agent composition tractable. The content follows the same rule: definitions are sharp, claims are linked, examples are concrete.

Recorded evidence over claimed behaviour

Numbers come from the upstream README and point back to it. The replays are the ground truth; the prose is commentary.

Open upstream, neutral mirror

This site does not fork the model or the data. It mirrors what is already published and adds framing that the upstream README does not (and probably should not) carry.

Static by default

No server, no runtime, no user data. The entire site is a directory of HTML, CSS, and a few JSON files — small enough to read in full.

05Where the content comes from

Every replay on this site comes from the upstream NanoJev project. The vocabulary, glossary, and FAQ on the homepage are derived from the same source, plus the Jev paradigm introduction at typesafe.ai. The Compare page summarises the comparison table from the upstream README and replays. The Guides page is a curated reading list of upstream documentation.

Continue to the comparison page — the same three systems, side by side.

Open Compare