DISCOVER → CHALLENGE → MEASURE
Find the relationship.
Then try to break it.
Search beyond a single guess. Compare compact mathematical explanations, test unseen observations, and identify a measurement that could tell the explanations apart.
OBSERVATIONS
Give it something to explain
1–4 inputs, with the target in the last column. At least 30 distinct input groups; up to 2,000 rows / 1 MiB. Headers: letters, digits, underscores.
Search and validation settings
Ready. The demo's answer is not supplied to the search.
Data stay in page memory. Exports include your numeric data; share them deliberately. The hosting provider receives normal page requests, not your CSV.
BETTER QUESTIONS, NOT JUST BETTER FITS
What would make
this explanation wrong?
The search fits only training observations. A separate validation partition selects the model. Final test answers stay out of both selection and experiment design.
Percentages apply to distinct input groups. This is bounded symbolic regression, not a general AI scientist or proof of a physical law.
EVIDENCE, NOT CERTAINTY
Selected explanation · coefficients fitted on training only
Dots are final-test observations. The diagonal is a perfect prediction; systematic departures expose model error. RMSE is in target units.
ASK A MORE INFORMATIVE QUESTION
The next measurement
Competing explanations
| Explanation | Selection RMSE |
|---|
Challenge it with a new observation
Enter inputs in header order and a measured target. This compares a new measurement against the frozen model. It does not operate any equipment.
What this does not establish
JSON includes the dataset, SHA-256, split membership, full-precision coefficients, candidate scores, proposed measurement, and limitations. It is a record—not an independently signed attestation.
A discovery instrument, not a claim of superintelligence.
It searches a fixed grammar of candidate representations, combines useful terms, retains competing explanations, and asks where they disagree. No frontier model runs behind this page. Synthetic examples demonstrate software behavior; they are not new scientific discoveries. Repeated tuning after seeing test results requires genuinely new validation data.