A calibrated decision API
Calibrated means that across many decisions where Myles gives an option 0.9, that option is right about 90% of the time. That is what lets you set a threshold, measure the error rate it keeps, and trust it.
Use cases
What builders use the Decision API for.
Myles answers the small decisions inside software: yes or no, one of several labels, or a level on a scale. You send the situation as text, JSON or images and get a calibrated probability for every option, read in one forward pass, so your code automates what clears a threshold and hands the rest to a person. Myles decides; it does not write text.
Decisions with a probability
One endpoint, POST /v1/decide, with three kinds of question. Several questions about one situation go in one request.
Yes or no classification with probabilities
A noul question returns the probability that the answer is yes. Is this a scam? Is the refund within policy? Does the answer cite its source?
Routing and labelling
A choice question returns a probability for every label, up to 26. Which team should take this? Which category? Which tool should the agent call?
Scores on a scale
A score question returns the most likely level, a probability-weighted score and a probability for every level. How urgent? How risky?
Guardrails for LLM apps and agents
Checks that run before and after your main model, in milliseconds, with a probability your code can act on instead of text to parse.
LLM input and output guardrails
Check for prompt injection, toxicity and personal data before and after your main model runs, without sending each check to a large model.
Grounding checks
Check that an answer is supported by the retrieved source before it reaches the user.
Tap-safety checks for computer-use agents
Rate how risky an agent action is before it runs, and ask a person when it matters. For computer-use agents, Myles-G 12B reads the screenshot and gives the probability that the tap is safe.
Decisions from images
Myles-G 12B and Myles-G 31B read PNG and JPEG images in the same request as the text, and the answers come back exactly as they do for text.
Receipt approval
Send the receipt photo with the claim and ask, for example, whether the total on the receipt matches the amount claimed. You get the probability of yes, and send the unsure ones to a person.
Yes or no questions about an image
Photos, scans and screenshots: approve this receipt, is it safe to tap this, is it fit to ship, which item is the target. The median answer with one image at the default size takes about a fifth of a second on our current GPU.
Billed as input tokens
An image counts as ordinary input tokens. Choose the detail per request with image_budget: 70, 140 or 280 tokens per image.
Where it runs
The same API on our hosted endpoint and on your own hardware, so code written against one works on the other.
A self-hosted decision model
Run Myles on servers and GPUs you operate, including disconnected networks, under a licence per deployment, monthly or annual, with no per-decision fees and nothing sent to us.
A UK sovereign decision model
Built in the UK, with a published lineage record for every new model and measured error rates. No Chinese-origin models or data anywhere in the lineage.
On a CPU or in the browser
Myles-400M, 150M and 22M need no GPU and run offline, called from your own program. Myles-22M runs in a browser in about 27 ms.