The model

A semantic encoder with a judge. It turns text into vectors, compares meanings and scores candidates with calibrated confidence, and abstains when the evidence is not enough.

tardi nano · mining 0.1 understands the operator’s question however it is phrased, recognises whether it belongs to the situation and tags each figure with what it measures. With those figures, the procedure’s rules do the calculation: AI understands, it never makes up the result. If the question is not about the situation, it abstains. It weighs 6.9 MB and runs in the browser without sending anything out.

Under 15 MB. It runs on the processor of ordinary servers and laptops, Intel/AMD (x86_64) or ARM (arm64), Apple Silicon included. No GPU, no network, no external services. tardi nano, the ultralight variant, also runs inside the browser of a computer or a phone.

Yes, and by construction: the runtime computes with integer products and functions defined operation by operation, with no library free to reorder them. Each version is checked against a reference implementation and between machines of different architecture, over a frozen set of texts, before it ships. The sandbox on this site runs in the browser on a different engine: it is for trying tardi out and does not carry that bit-for-bit guarantee.

It starts from an open base under a permissive licence and is transformed for each domain: a multilingual vocabulary pruned to the domain’s tokens, a much smaller body and its own layers to understand and evaluate, trained with that domain’s data, labelling and calibration. The line continues towards a tokeniser and an architecture of our own, with no inherited weights.

With a reference set built before touching the model and then frozen: queries written by people from the domain and pairs labelled by two annotators. The test partition is read once, every target is fixed before models are compared, and every slice is reported — negation, numbers, similar names, cross-language and abstention — with none allowed to go backwards. The sandbox model, tardi nano · mining 0.1, was measured on synthetic data: it shows how tardi works, not its quality in a real operation.

It doesn’t chat, doesn’t write and doesn’t reason in several steps. In each new domain it is measured with that domain’s data before it goes to work, because vocabulary changes from one sector to another. And if the calibration does not reach its thresholds, the judge abstains instead of guessing: that is the designed behaviour, not a failure.

Commercial, per delivered version, to use inside your own infrastructure. No telemetry, no remote kill switch.

Not today: it is built to understand and evaluate, and everything in it — the size, the judge, the determinism — comes from that decision. Short generation is a separate research line within the family.