Your AI model can install a fix.
Verified, versioned improvements that clip onto compatible open-weight models. Compare before and after on sealed tests, install the exact adapter, and unclip it whenever you like.
First patch
in evaluationretoque · unreleased · lora
A first-party repair for strict payment tool calls on Qwen2.5-1.5B-Instruct. It appears here after it passes held-out testing and review.
Bound to
Qwen/Qwen2.5-1.5B-Instruct@989aa79
One exact revision, tokenizer and chat template. Anything else is refused at install.
Before / after
ledger-tool-call
sealed held-out suite · strict grader · no run published
Base model
—
held-out cases passed
With fix
—
held-out cases passed
Values appear after the run is reviewed and published. Dataset hash, seed and decoding are pinned on both sides.
How a fix ships
Reproduce it, measure it, clip it on, unclip it.
- 01
Identify the failure
A patch starts from one narrow, testable mistake, measured on the unmodified base. If the base already handles it, there's nothing to patch.
- 02
Verify on sealed cases
Base and patched runs share checkpoint, prompts and decoding. Held-out cases use sentence shapes and names the adapter never saw.
- 03
Install, then roll back
The CLI checks revision, signature and every file hash before writing anything. Unclip or roll back with one command.
$ retoque install <patch>@<version> \ --model 'Qwen/Qwen2.5-1.5B-Instruct@989aa79'+ clipped on <patch> $ retoque status$ retoque disable <patch>$ retoque run --patch <patch> --input case.json --no-patch$ retoque rollback <patch> --to <version>Compatibility
One exact checkpoint, nothing implied.
A patch is bound to the repository, revision, tokenizer and chat template it was trained and tested on. The installer refuses anything else unless you pass an explicit experimental override, and it writes that override into your lockfile.
Supported modelsRobinhood Chain
A record of what exists and who licensed it.
The chain stores each version's manifest and adapter digests and each paid license. Weights and training data stay off chain. A registration proves a record was made, not that a patch works.
contracts not deployed in this environment
Before you install
Fair questions, straight answers.
The short version: you run the model, we ship the evidence and the adapter, and you can always take it off.
Does it work with Claude, GPT or Gemini?
No. A patch changes model weights, and closed API models don't give you their weights. Retoque supports exact open-weight checkpoints that you run yourself, and each patch names the one revision it was tested on.
What happens to my base model?
Nothing. The patch is a separate LoRA adapter loaded next to the base weights. Unclip it and you have the original model back; the CLI never writes to the base checkpoint.
Can a patch make things worse?
Yes, and we publish it when it does. Every version runs a regression suite of unrelated tasks. Any case that passed before and fails after is listed on the patch page, and so is any behaviour change outside the task.
What does a license give me?
Permission to use one specific patch version under its written terms. It doesn't license the base model beyond that model's own terms, and it doesn't cover future versions unless the terms say so.
Do I need a wallet?
Not for free patches: accept the terms and install. A wallet is only needed for paid licenses settled on Robinhood Chain.
Can I publish my own patch?
Create a project, upload rights-cleared data, train and evaluate in the studio, and submit for review. Paid publication is limited to verified publishers while review is new.