Thomas.

It checks itself.

thomas ask "When did Apollo 11 land, and who walked on the Moon?"
thomas check           # a fresh session that did not write the answer goes through it

After an answer, thomas check hands the question and the answer (none of your persona, notes or earlier turns) to a fresh session of your brain that did not write it, and asks it to find what is wrong. Each checkable claim it finds comes back marked OK, DOUBT or UNSURE, with the reason; a reply it cannot read that way is shown as written, up to its first 600 characters. On the free brain the check tries your free models starting from the next one in your list, so it may still land on the model that answered. Nothing is added to your conversation, and it needs one: with (memory :keep nil) there is no last answer to check.

A real run on the Claude brain, unedited, on an answer we planted with four mistakes in it ("Apollo 11 landed on July 21, 1968. Neil Armstrong and Michael Collins walked on the surface while Buzz Aldrin stayed in orbit."):

Checked by a fresh session of the brain that did not write the answer -- a second opinion, not proof: it
can share the first answer's blind spots, and it can be wrong itself.

  DOUBT  Apollo 11 landed on July 21, 1968
         Both parts are wrong: the landing was July 20, 1969 (20:17 UTC), not 1968. Apollo 11 launched July 16, 1969 — in 1968 the program was still flying Apollo 7/8.
  DOUBT  Neil Armstrong walked on the surface
         Armstrong did walk on the Moon (first, at 02:56 UTC July 21, 1969), but the claim is only half-right as stated since it's paired with a false partner; Armstrong himself is correct.
  DOUBT  Michael Collins walked on the surface
         False. Collins was the Command Module pilot and never landed; he remained in lunar orbit aboard Columbia and is the one who stayed behind.
  DOUBT  Buzz Aldrin stayed in orbit
         False — exactly inverted. Aldrin was the Lunar Module pilot and was the second person to walk on the Moon, joining Armstrong on the surface.
  DOUBT  Implied crew roles (two on surface, one in orbit)
         The structure is right — two moonwalkers, one in orbit — but the names are swapped: it should be Armstrong and Aldrin on the surface, Collins in orbit.

5 claim(s): 0 OK, 5 doubted, 0 unsure.

It caught all four mistakes. It also doubted two things that are true: Armstrong did walk on the Moon (its own reason says so), and two on the surface with one in orbit is right. That is what a second opinion looks like. It is not proof. The checker is the same kind of model as the writer, so it can share the same blind spots, and it can be wrong itself. It costs one more call to your brain.

What it can see.

Thomas runs on your machine; your prompts go to the brain's provider (Anthropic, OpenAI or Google) under your own login, or, on the free brain, to OpenRouter and the company running the free model. The free brain only answers in text. With the default settings the only other brain Thomas runs is Claude, in an empty scratch folder, and Thomas refuses every request it is asked to approve: reading, writing or running anything. Claude may still run a few read-only commands such as whoami and pwd without asking, so what reaches the provider includes your prompt, your persona settings, your account name and a temporary folder path, along with what Claude itself sends (such as your platform and the date). So that it remembers you, each message also carries your recent turns and your notes, kept in two files under ~/.thomas that Thomas creates readable only by your user account; (memory :keep nil) stops keeping and sending them; files already written stay until you delete them.

Codex and Gemini read files outside that folder even in their asking modes, so Thomas runs them only after you turn autonomy on. Reading another app's stored key is off unless you turn it on.

On the free brain, your prompt, persona, notes and recent turns go to OpenRouter and to the outside company running the free model, which may keep, train on, or publish them.