Turn a Paper into a Running Experiment

Load the original Research tutorial with no AI connection, review its evidence, change a setting, generate code, run a smoke check and a full run, and read the Dashboard figures.

What you will do

Walk the whole Research workflow on an original, public-domain tutorial specification, so you need no AI key and no real paper. The same stages apply when you start from your own PDFs.

Before you start

A project open in the Studio
A modern Chrome or Edge

Step 1: Load the tutorial

Open Research. Choose Load original tutorial (also under Import). A complete specification is imported, and no AI connection is used. The header shows the session and its revision.

Step 2: Read the stage rail

The stage rail shows Sources, Review, Configuration, Generate and run, and Results, each with a status line. Only a completed check or run earns a success mark. Experiment and target selectors in the rail apply to every stage.

Step 3: Check the evidence

Open Review. For each experiment you see its operations and inferred shapes. Click any Evidence link: a reader opens beside the stage showing the cited quote and the line of extracted text it came from. Note anything marked "Not stated in the sources" and read the assumptions the specification makes.

Step 4: Change a setting

Open Configuration. Filter by name, then open one setting's audit to see its rule, where its value came from and the code that consumes it. Change the number of epochs, then press Apply configuration and regenerate. That creates a new specification revision. Discard changes would drop the draft instead.

Step 5: Generate

Open Generate and run and generate the project. Open the code to see the files: train.mjs or train.py, the configuration, and a REPRODUCTION.md listing deviations. The Model Builder shows the pinned architecture, which Research code controls.

Step 6: Smoke check

Run the smoke check. It uses synthetic data to prove the code trains, evaluates, saves a checkpoint and reloads it. The result is labelled as a smoke result everywhere. The reproduction ladder keeps "generated", "smoke-checked" and "fully run" separate.

Step 7: Full run

For a full run, bind complete data: the sample set includes a 120-row CSV and JSON dataset. Bind the feature order, labels, class order and split protocol, and run. Watch the measured history draw as epochs arrive.

Step 8: Results and figures

Open Results. A run shows its metrics, training history, confusion counts, the figure intents it could and could not fill, and its frozen configuration and hashes. Open the Dashboard: automatic figures such as a learning curve and a confusion matrix are ordinary editable widgets with the run's provenance. Edit a title or palette; your edits survive regeneration.

Step 9: Export

Export a Research ZIP to move the code, configuration, revisions, runs and figures to another machine. Importing it needs no AI connection.

Using your own paper

Create a New research session and drop in a PDF or an arXiv identifier.
Set the AI provider in Settings and read the request and token budget shown on the Sources stage.
Mark one primary document, then Analyze sources.
Settle any conflicts under Decisions in Review and continue as above.

Limits to remember

A passing smoke or full run on a fixture proves the template works, not that a paper's reported accuracy was reproduced. Only supported families are generated: tabular MLPs, small CNNs, residual classifiers and a logistic-regression baseline.

Research Guide and Research.