Reading

How a Lodera question is made

A Lodera question begins as a controlled math structure. Numbers can change, but the task, rules, checks, and skill target stay known.

1. Start with a question template

No Lodera question is written by a chatbot. An author starts with a question template. The template names one solution process, such as using the product rule or solving an exponential equation by substitution.

The template also records the main skill, other skills used, a difficulty estimate, a hint, and worked steps. A branch is used when the structure changes enough to need different steps or different wrong choices.

This fixed structure matters. Review can focus on one repeatable task instead of treating every generated item as unrelated text.

2. Choose numbers under rules

Each template chooses values from allowed ranges. The rules avoid cases that break the intended task. A denominator cannot become zero unless that is part of the lesson. A quadratic can be required to have the kind of roots the question expects.

The generator retries a draw when a value creates a bad edge case. That means variety comes from controlled inputs, not a new prompt sent to a text model.

The same template can make many versions. The numbers and surface details may change, while the skill and solution path remain stable.

3. Compute the answer and wrong choices

The generator uses Python and SymPy, a computer algebra system, to compute math values. Templates build the correct answer from the actual expression. They also build wrong choices from planned mistakes, such as losing a sign or using the wrong exponent rule.

Wrong choices are not added as random numbers after the fact. A shared helper removes any choice that equals the key. It also removes duplicates and fills a standard item with three distinct wrong answers.

Some conceptual questions use a definition rather than a numeric calculation. Those still pass through the same item shape and checks. The claim is precise. Math software computes and checks computational answers. The bank also includes reviewed conceptual choices.

4. Run automatic checks

Every generated item passes a shared validation step. It requires a nonempty key, three distinct wrong choices, and no choice matching the key. It also requires a valid difficulty value, a real hint, and worked solution steps.

Other checks look for display problems and answer choices that reveal their own mistake. Family audits can sample many versions, look for repeated stems, frozen choice sets, missing branches, and suspicious answer patterns.

These checks are useful but limited. The repository guide says that automatic cross-checks can catch wrong answers. A reviewer still needs to inspect the worked steps. Software cannot fully judge whether a sentence is clear or a difficulty label feels fair.

5. Review the question template

The workflow produces browser previews and audit reports for review. A reviewer looks at sample versions, wording, steps, wrong choices, and whether the task fits the course. Branches with design problems can be blocked from future top-ups. Existing rows can be placed outside active serving while the source is fixed.

The active bank is built from reviewed question templates. That does not mean every item was typed one at a time by a teacher. It means the repeatable structure is checked, then the software makes controlled versions from it.

Template and branch names stay with each stored question. That makes it possible to trace a bad item back to the code that made it.

6. Listen when students report a problem

Students can rate or flag a question after answering it. The report is tied to the student, the exact stored question, and its template. That gives the reviewer the real item instead of a vague note about a topic.

When enough reports flag the same question, its status can change so active serving no longer selects it. The team can then check the item and its source template. The goal is to fix the cause, not patch one visible copy.

This feedback loop is part of quality control because no bank should claim to be error-free. Controlled generation makes problems easier to trace. Student reports help find what automatic checks and review missed.

Why this process is slower on purpose

A chatbot can produce a page of new questions in seconds. Lodera chooses a narrower path: reviewed question templates, controlled values, computed answers, automatic checks, and a route for reports.

That does not remove every risk. It creates stable places to test and improve. For math practice, repeatable structure is more useful than endless one-off text.

References

This article describes Lodera's current question-bank and practice-engine code.

Start with no credit card

Find the math skills you still need.

Start 7 days free