About
About Reflexmark
Why a decision layer deserves to be its own piece of software, who is building it, and how far along it is.
The idea
Agents spend a surprising share of their budget asking a large model for one word. Route this. Is it stale. Does a human need to see it. Which of these nine actions comes next. Each of those is a full generation call: the state is read again, tokens are sampled, and a word comes back that the calling code has to parse — with no honest signal of how sure the model was.
Reflexmark treats that class of question as its own problem. You send a state and a set of typed questions; it prefills the state once, runs each question as an isolated branch on a copy of that cache, reads only the option-label rows of the model’s output head, and returns calibrated probabilities. It never generates text. Psychologists would call this the System One layer: fast, automatic judgement underneath the deliberate work.
Why local
The states worth deciding on are the ones you would rather not upload: support inboxes, internal documents, half-finished plans, customer records. A decision layer only becomes cheap enough to use everywhere when it runs next to the thing that needs the decision. So Reflexmark targets a small open-weights model on hardware you already own, starting with Apple Silicon. Local is not a privacy feature bolted on top; it is why the design works.
Who builds it
Reflexmark is built by Thorsten Meyer, an independent developer in Iffeldorf, Bavaria. It is a one-person project, which is the reason this site makes no promises about roadmaps, support windows or availability dates. What it can promise is that the numbers printed here are measured rather than estimated, and that nothing claimed is bigger than what has been built.
Where it stands
In active development. Not publicly released. The service contract and the agent surface — the HTTP endpoint, llms.txt, the manifest with JSON Schemas, the MCP server and the CLI — are built and tested, and the unmodified official TypeSafe SDK runs against the compatible endpoint. The on-device engine is being built now.
There is no download, no pricing, no waiting list and no accuracy table yet. When there is one, it will carry majority-class and bag-of-words baselines alongside it.
Principles
- Local first. Your state never leaves your machines.
- Calibrated or it doesn’t ship. Every eval table carries majority-class and bag-of-words baselines.
- One open contract. The CLI, the MCP server and the HTTP endpoint are views on a single published request and response shape.
- Honest numbers. Prototype measurements are labelled as prototype measurements.
Independence
Reflexmark is an independent project and is not affiliated with, endorsed by, or sponsored by TypeSafe. Jev is referenced only to describe API compatibility.
Talk to me
Email contact@reflexmark.com. A description of a real decision you keep spending a full model call on is the most useful thing you can send.