Generated live, never retrieved

1. 1

   Your question is checked

   Not for content policing — for two specific things. A link to a paywalled study site is refused, because unlocking those is not what this is. And a pasted-in full exam paper is refused, because that is not a student stuck on a problem.

2. 2

   We look for the same question already answered

   Questions are normalised — punctuation, filler words and capitalisation stripped — so 'differentiate 3x^2 ln(x)' and 'Differentiate f(x) = 3x² · ln(x)?' are recognised as the same thing. A match returns instantly and costs nothing.

3. 3

   A subject-specific instruction is chosen

   Physics is marked differently from writing. Each subject carries its own teaching instructions — what counts as a step, what to check, how to phrase a correction.

4. 4

   The model writes the method

   Not prose. A structured list of steps, each with its own confidence rating and any technical terms flagged for definition. The structure is enforced, not requested — a response that ignores the format is rejected rather than shown to you.

5. 5

   It streams to you as it is written

   Steps appear one at a time because they are produced one at a time. You are watching the work happen rather than waiting for a finished page.

Language models are fluent whether or not they are right. That fluency is the actual danger: a confident wrong explanation reads exactly like a confident correct one.

So the model is asked to rate each step, and those ratings are shown. A low-confidence step is a signal to check that line against your notes first. It is not a disclaimer bolted onto the bottom — it is per-step, so it points somewhere specific.

When Homeworkify marks an attempt, it works the problem out independently first, then compares your working against it line by line, stopping at the first genuine divergence.

The rule underneath is deliberately lopsided: telling you that correct work is wrong is far worse than missing a mistake. A false accusation makes you distrust every later verdict and may push you to "fix" something that was already right.

So an uncertain verdict is not a failure — it is a real outcome, presented as "not certain about step 3" rather than a judgement. And that caution is enforced in the code, not just requested in the instructions: a low-confidence accusation is downgraded on the way out, where no change to the wording can bypass it.

It can be wrong. Confidently. That is why the ratings exist, why you can flag any answer as wrong, and why every explanation carries the reminder to check against your own course materials.

It does not know your syllabus. It does not know which board you sit, which notation your teacher prefers, or what was covered in class. A valid method that is not your class's method is still a valid method — but your marker may want theirs.

It is not your instructor. For anything graded or that counts, check the working yourself.

Your questions, the answers you received, practice sets, and marked attempts — all tied to your account so the dashboard can tell you something true about your own progress. That history is the whole reason it can spot a pattern in your mistakes; a tool with no memory of you can only ever answer the question in front of it.