Your working
Original equation + stepsNumbers, decimals, fractions, x and brackets. No powers, inequalities or division by x. Maximum 12 lines.
The reasoning repair lab
Don’t just mark it wrong.
Find the step that changed it.
Check your algebra one line at a time. See an exact counterexample, repair the reasoning, then test the skill on a fresh problem.
Linear equations in x
The first divergence
Exact math firstThe final answer isn’t the whole story.
Trace the steps to see what changed.
Trace the steps to see what changed.
01 / Verify, don’t guess
Rational arithmetic compares the solution sets. The neural model never marks your math correct.
02 / Ask a small model
A trained neural ranker suggests a practice family. It can be wrong and will sometimes abstain.
03 / Transfer, not copy
Try a new-number problem. The exact checker scores it; one click is not a mastery certificate.
Make the idea travel
New numbers, same skillChoose the next equation with the same solution set. The practice card and its answer are checked symbolically.
Choose a step to test your reasoning.
Adaptive choice uses smoothed practice counts, not a learned diagnosis. Progress stays in this tab unless you export it.
Your local learning record
Nothing uploadsSave the entered equations, exact checks and attempted practice. Imported correctness claims are ignored and recomputed. No names or email addresses are requested.
Trace a current chain to enable export. Exported files contain the math you entered; choose who sees them.
No current verification.
What the AI does, and where it stops
The 62 → 24 → 5 neural network is trained on 3,200 generated erroneous step pairs. It ranks distribution, balance, division, bracket signs and arithmetic. It does not generate proofs or infer why a real learner made an error.
The development numeric holdout shares the five templates: 93.7% top-class accuracy on 600 cases. A side-order normalization fixes a severe earlier failure. These are synthetic development results, not student validation. A later fresh coefficient split reached 487/600 top-class matches; the model issued 475 suggestions, all matching those synthetic injected-edit labels. This narrow result is not a guarantee for other equation forms. Full failures are in the source package.
All real-number linear solution-set comparisons use a separate BigInt rational parser. A malformed or unsupported line stops the check. Equivalent equations can still be pointless steps or use bad explanations; this tool cannot validate unwritten reasoning.
No account, API key, large-model download, external request or browser storage is required. The downloadable Python training script and JavaScript inference checks document the learned weights.