Knowledge store and executed answers. After that pass, a question can be answered another way when all of these hold:
- it is the request's only question;
- its options are all plain numbers, or all yes / no (true / false) words;
- the first pass gives no option at least 0.925 (numeric options) or 0.75 (yes / no options).
A numeric question first goes to a bundled knowledge store (country-level statistics; see Third-party components). If the store
returns a confident value that matches exactly one option, that option is the answer. Otherwise the model writes one short Python
function (greedy decoding, at most 320 new tokens), and a bundled exact-arithmetic executor runs it, in-process, as a GPU kernel.
Where the kernel cannot load, the executor's interpreter runs instead, with identical results. The answer changes only when the
returned value equals exactly one option and the function computed it (a bare constant return does not count).
In both cases the chosen option gets probability 0.99, and the others share 0.01 in proportion to the first pass. In every other
case the output is the first pass's, bit for bit.