A correct final answer can hide incorrect working. Find the break, repair it, write a fresh step, then try a changed structure. Keep the questions and the help you used.
The example ends at the right answer, but the two transformations are wrong. The exact checker finds the first break.
The trained ranker suggests a family, not a diagnosis. Override it here. An abstention asks you to choose; it does not invent a recommendation.
One repaired chain and two written steps. The record below keeps each question, every submitted response, and the help used on that card.
Correct after a worked example is not relabelled as an unassisted first response. This short exercise is not proof of lasting learning.
Counterstep compares real solution sets using an exact rational parser. A second check requires the requested step's resulting left/right coefficients, allowing equation-side swaps. Expansion must remove brackets; arithmetic must produce a numeric literal. This rejects a copied question or a final answer when expansion was requested, even if the solution set is unchanged.
The unchanged neural model chooses among five authored practice families. The new changed-structure questions do not expand the classifier's validation: its existing fresh numeric test was 487/600 top-class matches within its original five generators. No teacher review, learner trial, calibrated diagnosis or measured learning gain is claimed.
Supported: 2–12 linear equations in x, constants, rational numbers and parentheses. No nonlinear products, variable denominators, inequalities or other variables. The repair path requires a unique solution. A valid equation chain cannot establish unwritten reasoning or understanding.
Saved files replay inputs and help requests; saved outcome claims are ignored. Fingerprints identify bytes, not authorship. A file holder can construct another valid history, and the generators are in the source. This is a personal practice tool, not an exam or authenticated grade.