Frequently asked
questions.

A Behavior Package is a versioned, signed bundle that captures everything about how an LLM should behave — prompts, fine-tune recipes, LoRA adapters, eval suites, safety policies, and license metadata, all packaged together with a manifest.

Instead of copying prompts between repos and rebuilding evals from scratch every time you switch models, you install a Behavior Package the same way you install an npm or PyPI dependency.

Prompt repos store text. Joseki packages the full behavior contract: the prompt, the recipe that produced the model behavior, the evals that prove it works, the safety policy that guards it, and the license that says who can use it — versioned together and signed.

You can audit, diff, sign, install, and roll back a Behavior Package. You generally cannot do that with a shared prompt or a chatbot link.

Teams that ship LLM-powered features and need their prompts, fine-tunes, and safety rules to be reusable, testable, and governed. The first wave of users we are building with: AI platform teams at regulated companies, public-sector contractors, and AI consultancies that need to ship the same behavior to many clients without copy-paste drift.

If you have ever copied a prompt out of Slack into production, this is for you.

We are running closed early access through the summer. Join the waitlist and we will reach out as cohorts open up.

A Behavior Package earns a green badge only when, at the published version:

• Every required eval passes against the listed compatible models
• A prompt-injection and secrets scan returns clean
• The manifest is signed and matches the artifact hashes
• A machine-readable license is declared

If any of those checks fails, the badge degrades to review or failing and consumers see the difference at install time.

Every package carries an evidence trail: a signed manifest, content-addressed artifact hashes, a license spec, and the output of automated safety scans. Today the safety scan covers leaked credentials and prompt-injection phrases; the rule set will grow as we learn from real packages in the wild.

Licensing is captured as metadata on every version, so consumers can verify they have the right to use a package programmatically — not by digging through a README.

Yes. Visibility is per-package — public, organization-only, or private. Private packages live in your organization's namespace and are only visible to members you grant access to. We treat your packages as your IP.

Joseki WrapperHub is built by SenteGuard, a security-focused product studio working on trust infrastructure for AI systems. You can read more about how we think on the SenteGuard blog.

The fastest paths in:

• Email [email protected] — best for partnership, press, and security disclosure.
• Signal: djwide.13 — for off-the-record conversations.
• The waitlist form — if you want early access and a follow-up call.

Behavior Packages declare their compatible models in the manifest. We are launching with first-class support for the major commercial models (OpenAI, Anthropic, Google) and for self-hosted Llama-class models. Adapter packages can target any base model the manifest names.

Yes — joseki install, joseki inspect, and joseki eval are the headline commands. The CLI ships alongside early access.

The Behavior Package manifest format and the eval-runner spec are designed to be open. The registry, hosted evals, and the trust layer are commercial. We will publish more on this as we get closer to public launch.