What is Retoque

Retoque turns one specific, recurring failure of an open-weight model into a versioned fix you can install, test and remove. A fix is a patch: a small LoRA adapter trained for one task on one exact base checkpoint. You keep running the model yourself. Retoque handles the evidence, the license, the signed download and the checks that run before anything is installed.

What you get with a patch

• One adapter file next to your base model. The base weights are only ever read. Turning a patch off is a state change, and the base model behaves exactly as it did before.
• A held-out before and after. Each version cites one evaluation run on cases it never trained on, with the baseline, the patched result, a prompt-only control and the regressions it introduced. See Evaluation methodology.
• A signed package. The publisher signs the sha256 of the manifest with ed25519, and the manifest lists the sha256 and size of every other file. The CLI refuses a package unless all of it checks out. See Artifacts and verification.
• A license you can read. Each version has written terms identified by their sha256. You accept that exact digest, for free on the patch page or through an onchain purchase. See Licenses and entitlements.
• Rollback. retoque disable, retoque enable and retoque rollback --to switch what runs without reinstalling. See Disable and roll back.

What Retoque supports today

Retoque is at its first release. The registry supports one base model, Qwen/Qwen2.5-1.5B-Instruct at commit 989aa7980e4cf806f80c7fef2b1adb7bc71aa306, and the first-party patch the pipeline was built around is ledger-tool-call, which teaches that model to emit a strict create_payment tool call. What is actually published, and the evidence for each version, is listed on Explore and Supported models. These docs don't quote results, because the numbers belong to a specific run and live on the patch page.

What Retoque is not

• It isn't a hosted model API. The registry key you create in your workspace authenticates you to the Retoque registry only. It can't call any model, and it is never a substitute for an OpenAI, Anthropic or Hugging Face key.
• It doesn't redistribute base weights. You download the base checkpoint from Hugging Face yourself, under its own license.
• It doesn't make a model better in general. A patch is evaluated on one declared task, on one checkpoint, under one decoding setup.
• A signature or an onchain record proves who published which bytes and when. Neither one proves that a patch works. The evaluation report is the evidence, and its limits are written next to it.

Where to go next

1. Getting started walks through installing the CLI, logging in, installing a patch and rolling it back.
2. Concepts defines patches, versions, compatibility and evidence states.
3. CLI reference lists every command, flag and exit code.
4. Creator guide covers building and submitting your own patch.

These docs are also available as plain Markdown for assistants and scripts: every page has a raw version under /docs/md/<page>, and the whole set is indexed at /llms.txt.