Language models read text as tokens, not letters, and generate answers one piece at a time. That makes some simple-looking tasks unreliable: counting how many times a letter appears, spelling backwards, long arithmetic. Keplar handles these with code.
What Keplar treats as exact
The classifier flags two kinds of question as error-prone:
- Exact: counting and arithmetic, such as letter counts, totals, vowel counts, expressions using
+ - * / ^ ( ) %, and "N% of M". - Fact: a precise number, year, date, size or distance ("what year", "how tall", "population of").
What happens
- The question is never left to a single model. It gets a small panel plus a check (on Free too, within the three-model cap).
- For counting and arithmetic, Keplar's own code computes the result.
- That result is placed in the models' prompt as a verified fact, so they explain it rather than guess it.
- Under the answer, Sources lists an Exact check line saying the value was computed by Keplar's own code.
Question: How many times does the letter "r" appear in "strawberry"?
Computed: 3 (checked by code, not by a model)That example shows the shape of the feature; the live answer text is written by the models around the computed value.
What it does not cover
Only clearly countable or computable questions are handled this way. Anything else is untouched. The "fact" category (a year, a distance) gets the extra panel but not a code computation, because code has no way to know the true value. For those, a primary source is the check.
Why not give every question a calculator?
Because nearly every question is not a calculation, and a false positive would change an answer that was right. The rule set is deliberately narrow.
A caution from testing
On 3 October 2026 a letter-counting question on the Free plan was classed as simple and sent to a single model, which gave a wrong count, before this check was extended to cover it. The check exists because that failure is real. If you find a counting or arithmetic question that still gets a wrong answer, please send it to team@keplar.one.
Related
- How Keplar understands a question: The classifier that decides task type, complexity and needed capabilities, what signals it uses, and why it is a transparent rule set rather than another model.
- Which questions use more models: A practical guide to what makes Keplar consult one model, three, or five: question type, length, stakes, forecasts, attachments, Deep Research and thoroughness, with examples.
- Sources, citations and web lookups: What the Sources section really contains, why Keplar does not open or check the links, the status of web lookups, and a checklist for verifying what matters.