What you get —
and what you don’t

You frame the problem — the target, the kind of problem, what one row stands for — and the study tries every engine on several parts of your development data, each with its own cross-validation. The standard package races gradient-boosted trees (LightGBM and scikit-learn) against a reference baseline and, since v2.8, our own neural network (in both packages, with its own time: 60 s per attempt, added to the budget, so no other engine loses a second). It only competes if it was measured on every part of the cross-validation, and the study picks per dataset: it wins only where it is the best. The GPU package keeps our previous network too, as an experimental engine; in the CPU package the new network takes its place. The champion is the one with the best mean over its parts, and the screen names it, says what that engine is and gives the reason in plain words. When the paired comparison against the runner-up does not show an improvement, it says that too: a choice under uncertainty, not a clear winner. A few engines, not hundreds, and no hyperparameter search.

The figure that counts comes from data no candidate saw while competing: the test set is kept intact until the end, and the comparison shows each engine’s mean over its parts — the one that decides — with the range, so a lucky split cannot pick the champion. Below it, the evidence says what applies and what does not, and why, and the receipt records what competed, what won and under which rule. The receipt comes out Unsigned until you configure your own signing key; its signature is an HMAC, so a third party cannot verify it without that key. It documents the choice — it is not a certificate.

Build the package and you get the winner with its preparation recipe and a predict.py that predicts like the screen does — with Python, outside the Studio, but it needs the Studio’s engines: it does not run on a machine without them. The receipt is not in that package: it travels in estudio-completo.zip, the full study, which you reopen in another Studio. A page that predicts in the browser (ONNX) is offered only when LightGBM wins a binary classification, with its decision chosen and no date columns; otherwise the option is shown switched off with its reason. Your data never leaves your machine — except the 5 rows and the counts you send to your LLM if you tick the optional proposal described below.

If you have an LLM configured in Settings, you can tick “Let the LLM propose the fields”. It then proposes the empty fields of the framing — target column, problem type, positive class, observation unit, prediction time and horizon — and each proposed field carries a “Proposed by the LLM” mark until you edit it. What goes out: the column names, a few counts for each column (type, % empty, distinct values) and 5 real rows of your CSV, to the LLM configured at that moment, which the notice names before you tick. What you send is your responsibility: with personal data, do not tick the box or remove it from the CSV first. What it does not do: it does not fill in what you already wrote, it confirms nothing — you review and confirm — it needs a configured LLM, and it is not in the hosted demo. Without the tick, this sends nothing.

Three ways in: a sentence, your CSV, or a template from the catalogue. Whichever door you take, the core reads the data, proposes a target with its reason — and the alternative it considered — and prints the schema it inferred: every field, its type, its range, all editable. It proposes; you decide. And the screen warns you about the part that matters: a wrong range makes the model learn over a domain that does not exist.

The audit is not a training summary. Six checks run against the model itself — parser, verifier_agent, safety_agent, typecheck, python_compiler, backend_contract — and they re-run on every change, so the verdict cannot go stale behind your back. The pipeline stages that carry no pass/fail are listed apart with whatever they did report, instead of disappearing under a green tick. The whole verification downloads as a file, and the trace links the model to the data it was prepared with and the training run it belongs to.

Before training, the model is already an object you can verify: the type check types the network and states its interpretability level, and the parameter count is printed so you can redo the arithmetic yourself — a 64 → 32 → 1 network over three inputs is 2,369 parameters, and that is what the screen says. The resource estimate arrives with figures, not adjectives, and it works in both directions: a small model announces kilobytes, and when we asked for a 2,953,150,466-parameter network, the answer was that training it needs ~26.8 GiB — the hosted demo refuses it outright, and the downloadable Studio shows the warning with its figures and leaves the decision to you. A tool that tells you the bill before you commit is the difference between an estimate and a promise.

This is the expert-mode export, not the study’s package. The self-contained bundle carries the model in ONNX, a predict.py that loads it and validates the schema, the manifest with its hashes, and the model card with its metrics and its limits. You feed it raw values — no MatrixAI installed, no service to call. And the parity is measured, not assumed: in an August audit of that export, the exported model, loaded with a plain ONNX runtime, returned the same answer as the Studio on all six test inputs — six of six.

The template catalogue includes a case built on five years of real daily history for Santander, Spain (Open-Meteo): temperatures, pressure, humidity, wind, cloud cover, dew point and today’s rain, with the target built by shifting rainfall one day back — each row predicts tomorrow using only what was known that day, never data from the future. The card also carries, in red and before you create anything, the limit of its data: it does not allow commercial use. Next to it, a wave-height case on five years of real hourly marine data. That is the pattern everywhere: what something is good for and what it is not, printed where you decide.

The full expert-mode cycle, screen by screen, is in its manual — captured from a driven session and honest about its numbers: its walkthrough trains on generated rows and reports a chance-level result as exactly that, because showing that the product declares a bad run is worth more than a prettier screenshot. The main path, the study, is in the tour.