Before I work it out
A working prototype · first pages
Math has more than
one kind of wrong.
Bring a new math problem, check a sequence of steps, scan written work, or work through a lesson with fresh values. The local engine works on new supported inputs across algebra, trigonometry, calculus, statistics, probability, and common geometry formulas, with a distinct checked range for each operation.
Eleven reusable lessons now span arithmetic, algebra, linear algebra, geometry, trigonometry, calculus, and statistics; sixteen fixed examples add more starting points. The local algebra path factors exact-rational polynomials through degree two and solves every real quadratic case, including exact radical roots and a checked no-real-roots result. One shared, revision-bound lesson loop connects fresh examples, practice, checked repair, and transfer. Repair is available for confirmed equation or inequality tasks and supported exact-fraction sequences; calculus candidates are compared with the requested result, not treated as line-to-line steps. When configured, an optional AI tutor can explain unfamiliar topics and problems; its answers are not independently checked. The local solver still abstains on unsupported work and cannot read diagrams.
Open the notebookA fresh problem is welcome
Give the math an instruction.
Type a new expression and say what to do with it. The notebook reads back the operation, input, variable, and key options before calculating. It uses a symbolic algebra engine in this local browser tab, not a list of matching examples.
Try “solve x^2 − 5x + 6 = 0 for x”, “solve -2x + 3 < 7 for x”, “differentiate sin(x^2) with respect to x”, “How fast is x^3 changing at 2?”, “probability of rolling an even number on a fair six-sided die”, “area under x from -1 to 1”, “mean of 2, 4, 6”, “area of a circle with radius 5”, or “multiply matrices [[1,2],[3,4]] and [[5,6],[7,8]]”.
Local reading stays in this tab. Optional model interpretation is separately disclosed before any text is sent.
A worked result
This explanation will use only the checked steps and task shown above; the tutor's response is still AI-generated and unverified.
What the symbolic engine can do
It can evaluate or simplify expressions, factor or expand polynomials, solve supported equations, inequalities, and systems, differentiate, integrate, find supported limits, calculate common geometry formulas, and summarize numeric data. The local parser understands a few common task phrasings, including linear inequalities, instantaneous rate of change, and area under a curve. If an OpenAI model is configured, you may separately consent to send the exact text in the question field for a proposed reading; the model can only route toward a registered local calculation and never verifies it. For fresh single-variable rational polynomials through degree eight, it builds a term-by-term power-rule lesson and offers two optional checked practice problems. It also teaches exact factorization for rational polynomials through degree two and solves real quadratics by exact factorization, the quadratic formula, or negative-discriminant classification. That lesson is limited to at most 12 nonzero terms after collecting like terms, rejects any explicitly written power above eight, and uses bounded expression-tree size. It reads back the operation and math input, shows the exact request sent to the local engine, target variable, domain caveat, and assumptions; it asks when an equation, variable, region meaning, integral bounds, or sample-versus-population choice is unclear. Calculation still requires confirmation. Editing a request clears its old interpretation and answer.
Single-variable rational polynomials through degree two get a checked factorization lesson; the quadratic work checker accepts reordered factors and reports exact coefficient mismatches. Real quadratic solving checks completeness through exact discriminant classification: rational roots use factor reconstruction and zero-product reasoning, irrational roots use exact radicals with symbolic substitution, and a negative discriminant gives no real solutions. Complex roots are not returned. Numeric-coefficient linear systems written with polynomial equations get generated Gaussian-elimination steps (up to six rows and variables). Indefinite integrals of single-variable rational polynomials through degree eight now show checked term-by-term power-rule steps, recombination, and the constant of integration; polynomial definite integrals show an antiderivative and bound substitution. One-sided limits can state a left or right approach. Geometric area is limited to quadratic-or-lower curves with rational coefficients, bounds, and intercepts. Systems with variable denominators, functions, or non-integer powers stop for review. Typed decimal literals become exact fractions up to 12 decimal places. Many equation and derivative patterns still get only a calculation rather than a full lesson. Rule traces are consistency-checked with exact local arithmetic and the bundled computer-algebra engine, not independently proven, and real-domain restrictions are not fully checked. A consented model may help map a word problem to a registered operation, but proofs, diagrams, open-ended concept teaching, and unsupported math still need review.
The first mark is yours
What does the work say?
Paste fresh work, choose a starter example, or bring a photograph. The checker analyzes new inputs with symbolic and exact-arithmetic tools: polynomial and selected equation steps, one-variable linear inequality solution sets, supported calculus traces, safe sin/cos/exp identities, rational arithmetic, exact probability values, rational quadratic factorizations with reordered factors, matrix final answers, and one final numeric answer for confirmed percentage or same-rate cost tasks. Matrix answers are compared exactly, but their intermediate reasoning is not checked yet. It checks probability answers but not every reasoning step, and it still abstains on intermediate ratio work and domain-sensitive expressions. Exact authored matches can open an additional example path. A written change is not a claim about the person.
Four strands, four authored patterns in each. Review status: teacher review pending.
Try a fresh identity such as sin(x)^2 + cos(x)^2 followed by 1. For probability, start with probability of A: 3 favorable outcomes out of 8 equally likely outcomes, then enter 3/8. For calculus, start with the task, then add each candidate line: differentiate sin(x^2) with respect to x, cos(x^2), 2*x*cos(x^2). Safe polynomial, sin, cos, and exp expressions can be compared beyond the fixed example atlas. If symbolic evidence cannot settle a change, Y? marks it unresolved; domain-sensitive notation goes to review.
Your typed work stays in this tab. It is not saved or sent.
Bring in a written solution one line at a time. Each source crop stays beside its editable transcription so you can correct and reorder the work.
Keep the task separate from the photographed steps. For equation or inequality work, this becomes the starting statement, so crop only the later written steps. For calculus, enter the full operation and variable here, then crop the candidate answers. Use a supported equation, inequality, or command; open-ended word problems are not interpreted yet.
On-device formula OCR loads about 53 MB of open model weights on first use. The photo itself stays in this browser; the model files load from their pinned CDN and model-host URLs. Handwriting accuracy has not been validated, and every symbol needs review.
Written lines, in page order
Compare each crop with its text. Move lines into the learner's order, edit unclear symbols, or remove a bad crop. No correction is made silently.
Add one or more reviewed crops, then confirm the task and their order.
Start with the question
Choose an idea to work through.
Fresh values are generated for each lesson and checked within its stated limits. These author-written lessons still need teacher review.
Eleven reusable lessons across five strands
Each lesson gives fresh values, a worked example, a quick check, simpler feedback after a miss, and a different transfer problem. Topics include equation balance, inequalities, fractions, percentages, linear systems, matrix multiplication, geometry, trigonometry, calculus, and statistics. A guided repair path also covers changed work in linear equations, inequalities, exact fraction arithmetic, and polynomial derivatives. One correct response is not a mastery claim.
Prerequisites, boundaries, and review status
A fresh worked example
Check the idea
Choose an answer to continue.
Return to your own work
Repair the first changed line
A different problem
Ask about a math idea
The authored pages are a small starting atlas. When connected, ask the AI tutor about other topics or ask it to work through a problem, even when the local solver cannot check that kind of math. AI-generated calculations and proofs stay marked unverified.
Follow-up questions can use the last two tutor exchanges from this tab. “Clear question / conversation” also clears that temporary conversation.
Notebook search stays in this tab. No question or image is sent.
Checking whether the optional AI tutor is configured…
Let’s look closely.
Choose the next move
What would help right now?
The consent preview shows the typed or OCR-corrected text and local checker evidence that would be sent. The server does not independently authenticate that evidence. No image is sent; the tutor’s explanation stays unverified.
A possible next page
Why did the notebook offer this?
One more question
Try a nearby example.
Choose an answer to see what this response may suggest.
A small repair, then a new try
Sixteen places to begin.
Four authored patterns in each strand make a testable first instrument. Add a pattern only when its math and question have been reviewed.
What this page can claim
Matrix requests now calculate fresh explicit matrices rather than selecting a stored worked example. Addition, subtraction, multiplication, transpose, determinants through 3×3, and inverses through 2×2 use exact rational entries. Generated explanation lines are recomputed by the same engine. Typed or OCR-imported work can compare one final matrix answer only; it does not locate a matrix mistake in intermediate work yet.
The symbolic workbench reads a supported typed request back to the learner, asks for missing meaning, and calculates only after confirmation. It uses a locally bundled computer-algebra engine and has no custom-trained tutor model. A separate optional AI tutor can explain concepts and work through problems or proofs across math topics when the server is configured; its calculations and proofs are AI-generated and not independently checked. For rational arithmetic, it evaluates bounded integer fractions, decimals, parentheses, and all four operations with exact BigInt rational arithmetic; a confirmed task reader also handles a percentage of a stated quantity, an increase/decrease from a nonzero base, and cost scaling when the prompt explicitly says the rate stays the same. Probability support calculates exact fractions from explicit favorable/total counts, one fair coin or six-sided die, conditional counts, and P(A|B) when P(A ∩ B) and P(B) are supplied. It does not infer independence, hidden sample spaces, or Bayes' theorem. It checks written probability values, not every conceptual step. It verifies the displayed operations and checks handwritten or typed arithmetic transitions. Division by zero, ambiguous notation, and work outside these grammars stop for review. The arithmetic checker verifies exact values; when a task explicitly asks to simplify, it also checks reduced integer or fraction form and leaves mixed numbers or unfamiliar formats unclassified. For equations with exact rational coefficients, it shows linear solving steps with an explicit check beside each transformation. For one-variable linear inequalities, it compares exact real solution sets, preserves strict versus inclusive endpoints, recognizes one-sided intervals, and checks that negative multiplication or division reverses the sign; variable denominators, nonlinear forms, and chained inequalities stop for review. A task-headed linear solve also separates valid equivalent progress from a completed x = value answer, and identities or contradictions need an explicit all-values or no-solution conclusion. The first changed equation remains marked if a later line repairs the final answer. For rational polynomials of degree at most two, it checks factor reconstruction and accepts reordered factors; real quadratic solutions are classified exactly; irrational roots use the quadratic formula and complex roots are outside the returned real-domain result. For single-variable rational polynomials through degree eight, it shows constant, sum, and power-rule applications with exact coefficient/exponent checks and optional practice. For indefinite integrals over the same bounded polynomial family, it shows term-by-term rational integration, recombination, the arbitrary constant, and optional checked practice. In that derivative lesson, the 12-term cap applies after collecting like terms; written powers above eight and oversized expression trees fall back without a checked lesson. The original expression is parsed into canonical terms and compared with the reconstructed term list before rule checks. For the supported quadratic-area route, it shows a complete exact root classification through degree two and checks the antiderivative, each interval contribution, and the sum. These checks reuse the same symbolic engine and are internal consistency checks, not independent proofs. It can also show generated row-reduction steps for numeric linear systems and check some task-headed calculus traces against a derivative or antiderivative target. Eleven parameterized, author-written lessons generate fresh equation-balance, linear-inequality sign-reversal, linear-system, fraction-equivalence, exact rational-arithmetic, polynomial power-rule, percentage, rectangle-area, right-triangle-sine, arithmetic-mean, and matrix-multiplication examples. Transfers use exact arithmetic, exact-rational equation checks, or the symbolic workbench within each stated contract. Each lesson has a quick check, simpler feedback after a miss, and a different transfer question; one correct response is evidence about one item, not mastery or a learning gain. After a verified linear-equation change, Y? can offer related equation-balance practice; after a verified inequality change, it can offer sign-reversal practice; after exact rational arithmetic, it can offer optional fraction practice; after a differing percentage/proportion final answer, it can offer related percentage practice; after matching supported polynomial-derivative work, it can offer optional power-rule practice. The routing record shows the task, checker contract, result, rule, input revisions, and lesson version. A wrong answer does not establish its cause, and unsupported or ambiguous work does not receive an inferred lesson. Lesson content is author-written and teacher review is pending. The authored error atlas still contains 16 fixed examples with probes. OCR can assemble multiple individually cropped lines into a reordered, editable sequence; the task must be entered separately and the learner confirms the sequence before checking. The image remains in the browser, while model files load from pinned public hosts. Handwriting has not been validated.
For a confirmed linear equation or inequality task, and for a starting expression checked by the exact fraction engine, guided repair can begin at the first changed move. Later fraction errors can start from a preceding verified step. The learner tries a fresh example, edits the original sequence, runs the full checker again, then attempts a distinct transfer. Calculus candidates are compared with the requested result; they are not treated as line-to-line derivations. The checker does not infer why a line changed; lessons and probes are author-written and teacher review is pending.
The local solver does not yet handle every natural-language problem, geometry proof, or diagram. When configured, the optional AI tutor can explain and work through math beyond the local solver, but its calculations and proofs are not independently checked. It receives only the text shown in the consent preview; it cannot read the page photo. Unsupported local-solver work stays unclassified for teacher review.
The earlier StepVerify experiment remains a separate math-word-problem baseline; it does not power this notebook.
Open the earlier typed-equation experiment Formula OCR model and runtime notices