How we use AI

Each of these is a separate surface with its own rules. Nothing here places, cancels or sizes an order.

• The AI tutor. Answers a question about a concept while you are learning, and can keep a concept map of what you have covered.
  What it sees: Your question and the lesson you are on. Your account id is replaced by a one-way pseudonym before the call, and names, addresses and phone-shaped text are removed.
  What it does not see: Your name, your email, your book, your positions, or any other member's work.
  Runs on the restricted tutor lane: organization keys only, and an allowlist of vendors.
  You choose: Keeping tutor conversations and the concept map. You can take it back at any time.
• The research assistant and workflows. Drafts, sources and reviews research you asked it to work on, in stages you confirm.
  What it sees: The prompt you wrote, the documents you attached, and public market data.
  What it does not see: Another member's private research, and nothing at all unless you started the job.
  You choose: Logging research prompts for fairness analytics. You can take it back at any time.
• News screening. Reads public market news and decides which articles are worth surfacing against a ticker.
  What it sees: The article text and the ticker symbols involved.
  What it does not see: Anything about you. No member data reaches this call.
• The thesis judge and the compliance sidecar. Scores a submitted thesis against a published rubric, and flags a Call that reads as out of mandate.
  What it sees: The submitted work and the rubric. Names and addresses are removed before the call.
  What it does not see: Who wrote it. A judge call carries the submission, never a member id.
  Runs on the restricted judge lane: organization keys only, and an allowlist of vendors.
• The order copilot. Comments on an order you are about to place, before you place it.
  What it sees: The ticket you are filling in and the shape of your own book.
  What it does not see: Any other member's book. It never places, cancels or sizes an order; it writes a comment and you decide.
• Guide answers. Answers a question about how the platform works, from the platform's own guide.
  What it sees: The question and the page you asked it from.
  What it does not see: Your account. A guide question carries no member id.
• Claim adjudication. Checks a mechanism claim an organization has approved for review.
  What it sees: The claim text and the approved mechanism it is checked against.
  What it does not see: Anything the organization did not put in the claim.
  Runs on the restricted adjudication lane: organization keys only, and an allowlist of vendors.
• The resident agent fleet, where an organization turns it on. Runs bot members that trade a simulated book and write their reasoning, so a season has depth before people join.
  What it sees: Public market data and the bot's own book.
  What it does not see: Any human member's book, and it never ranks against one on a board that says otherwise.

• A season title is nominated by the platform and confirmed by a named person. `SeasonAward.confirmedByUserId` is required with no default, so a title with no confirmer cannot exist.
• An unlock test is human-only. A request carrying an access token instead of a signed-in browser session is refused before any input is read.
• A compliance override is a decision a person takes and the platform records, with who took it and why.
• The thesis judge's score can be weighted against a human grade, and the scorecard shows the model, the rubric, the temperature, the agreement across runs, and the human weight.
• A disqualification takes two distinct people, neither of whom is seated in the season.
• No model places, cancels or sizes an order. Every trading decision on this platform is made by a person or by a declared, published bot strategy.

Generated from the model configuration. A vendor appears here when the code can route a call to it, whether or not a key for it is currently set. “Organization key” means the platform’s own credential; “your key” means a key you added yourself under Settings, which only your own calls use.

The restricted lanes never reach a self-hosted endpoint, whatever a configuration override says. Nobody here can state the terms of an endpoint somebody else runs, so a student’s learning conversation and a graded submission never go to one.

Prompt logging is never switched on for our account. Switching it on would grant the vendor a broad, permanent licence over everything logged, so it stays off and the restricted lanes ask, on every single call, to be routed only to a provider that does not keep the prompt at all.

We run open-weight models: models whose weights are published under a licence anyone can read. That is a cost decision before it is anything else, and it has a consequence worth stating, which is that each model carries its own terms and we are the ones responsible for meeting them. This is that list, generated from the same catalog the platform routes on.

Attribution the licences ask for: Built with Llama.

This notice is versioned by its content. Version 1c67a91543476b70, last material change 2026-09-09. A material change is a new vendor, a new lane, a change to whose key is used, a change to what a surface is shown, or a change to what a vendor’s terms say. Adding a new model from a vendor already listed is not material: the same vendor sees the same things.

The version is computed from those fields by the build, and the build refuses to publish a new one without a dated entry explaining it, so a change cannot land here quietly. If you have turned on anything AI-related in your settings, you are shown a one-line notice on your next visit after a material change.

See also the Privacy Policy for what is kept and for how long, and the Terms.