Tuiscint — answers with receipts

Reads what it retrieves, and shows the source for every claim. Says “not stated” rather than guessing.

Colour theme

Reads the open web and repeats what it finds — sometimes wrongly. Not advice, and not for a crisis: Samaritans 116 123.

Settings
Retrieval

A canon is a fixed, curated corpus the operator chose — here, this project’s design docs. Deterministic and fast. The web is live and messy.

More sources, more to read — and more places to go wrong.

Verification

Rejects an answer whose quoted span doesn’t answer this question. On noisy sources it is the main defence.

Lower vetoes less. 0 disables it.

On failure, re-reads each source alone and keeps the answer most of them agree on.

Decoding
Mode Think writes a scratchpad before answering, and measured worse: it refuses less reliably and its reasoning is often unfaithful to the sources. Turn it on to watch the model talk itself into an answer.

0 is greedy, the setting every published number was measured under. Raise it to watch a grounded answer drift into invention. Nothing is cached above 0.

Sample only from the k likeliest next tokens; 0 means no limit. Does nothing at temperature 0. Not the same as “sources to read”.

Try: