Augustus

No, and it says so directly: Augustus is independent of TypeSafe, the official skill and current provider docs own API contracts, and Augustus supplies composition, implementation, evaluation and improvement methods. It also reaches past Jev to the wider decision-model class, naming GLiNER, Laya, kev, OpenJev, SemIf, NanoJev, Jeff-1 and localjev.

The stated working model is evidence, bounded judgment, explicit policy, checked action, observed outcome, and sometimes the best result is a formula, a checklist, or no new model. Its examples route expensive generated-JSON email routing to a bounded intent classifier placed before existing handlers, keeping eligibility and payments in code, and its recipe cards warn against treating a schema-valid response as correct.

Prompt text. It ships as an agent skill, plus a companion augustus-train, installed through the Claude Code plugin marketplace or npx skills add, and you invoke it as a slash command. It needs no API key to give design guidance; calling Jev or another hosted provider is a separate optional integration with its own credentials and costs.

Not on its own. It guides an agent through task-specific data assembly, candidate selection, fitting, export and reload, and bounded improvement under application costs and constraints. The README rules out calling it a pretrained model, a hosted training service, or a recipe for reproducing a general instruction-conditioned Jev engine, and describes its acceptance audit as public-proxy and fixture executions rather than production evidence.

The published release is 0.8.1: add the marketplace at 24601/[email protected] and install augustus@augustus, or use npx skills add for Codex, Cursor and other Skills CLI agents. Current development is 0.8.2-dev, so a default-branch install can receive later unreleased work. Check with claude plugin details augustus@augustus or the metadata.version line in the installed SKILL.md, and invoke it as /augustus in Claude Code or $augustus in Codex.

Different artifacts for different asks: a concise design card and falsifier for advice, working adapters plus policy and evals for build requests, and a bounded incumbent-versus-challenger loop for improvement requests. The README is firm that the agent reads only the references relevant to the ask, and that source popularity and proxy-score gains do not establish improvement. Its own tests and structural lint are described the same way: they do not establish model quality.

Augustus De Morgan, mentor of William Stanley Jevons, which is also where the project's short film picks up Jevons' logic piano and De Morgan's boundary. The naming tracks the line the skill keeps drawing: a model that reasons about a decision is not the same thing as a reliable decision, so evidence, policy and an explicit abstention path sit between the judgment and the action.