Model policy

Every model role runs as a fresh, short session with a fixed model, a turn cap and a dollar cap. The operator policy is one line: Sonnet 5 is the floor, the deep sessions (the PM's plans and the brief) get the top model, Fable 5.1, and adversarial review runs on Opus 5.5. "Fast" means Sonnet 5 with a tight budget, never a smaller model.

Each role is also tested on its own work, and the result decides how much choice it has. Today the deep roles keep their standing assignment: the PM's plans and the Researcher's brief run on Fable 5.1, the PM's intraday checks on Sonnet 5 and the Critic on Opus 5.5, and they stay locked to those models until a full round of testing says otherwise. Consult and the Quant are a choice among the models that scored about the same on their checks. Within a role, a model that did not pass is marked not available, with the reason in words. On Consult the model menu shows that list with each reason. On an agent's Settings tab the picker lists every allowed model, and a model the role may not run is refused when you save, with the reason.

Who runs on what today

The Quant's nightly review runs as code and starts no model session today (the Quant). The Trader is built and in testing and stays switched off for now; when it runs it is assigned Sonnet 5 with a small budget ($0.40 and 12 turns). If a model is unavailable the session falls back to Opus 5.5 (the fallback model setting); the Critic, whose own model is Opus 5.5, falls back to Opus 5. A PM or Researcher run you ask for from Consult uses the same model and cap as its scheduled kind.

Every session on your book (PM, Researcher, Critic, Quant and Consult alike) draws on one daily AI budget, which the platform owner sets for each book. A session may spend the lesser of its own cap and what is left of the day, and some is always kept back for the end-of-day check. When the day's budget is spent the desk pauses until midnight CT and says so: Consult answers with a polite refusal and the Critic is skipped, while your positions, stops and open orders are untouched (when the budget runs out). The Overview's model-spend figure and the day's one-line summary on Desk show what was actually spent.

Where you change it

Cadence and models live with the agents. Desk › Agents has one card per agent with the model it runs on, when it works and today's sessions. The Settings link on a card opens that agent's Settings tab (Portfolio Manager › Settings, the same for the Researcher and the Critic). It is the one place that edits that agent:

• Model for the role: one picker that covers the PM's morning plan, its daily review and any run you ask for (the check-in model has its own picker under Schedule, and the wake-up model is not set on this page). Models are listed most capable first because this is where you set a role's standing model: Fable 5.1 (the most capable), Opus 5.5, Opus 5, then Sonnet 5 (the everyday model); no smaller model is offered. (Consult's per-message menu runs the other way, quickest and cheapest first, because there you are choosing what to spend on one question.) A Fallback model and the Effort level sit beside it.
• Max turns per session and Budget per session: lowering either is one click; raising it again above what runs now asks for your authenticator code, like loosening a risk limit.
• Intraday sessions on or off, the Intraday interval (15, 30 or 60 minutes, 09:00–⁠14:30 CT; the default is 60, a posture preset carries its own advisory interval and this setting overrides it; a slot after the day's first is skipped on a quiet tape, so the count is a ceiling) and the Intraday model. These are on the PM's page only, because only the PM has a schedule you set. Premarket and postmarket always run; while intraday is on, the fixed 14:48 CT close-confirmation pass runs too (it is not on the interval grid and is never skipped as quiet). The Researcher's page has one switch, Daily briefs on or off; its two brief times are fixed.
• Wake-ups: how many a day, the quiet time between them, the budget and model per wake-up, and what counts as close enough to a stop, a large move or a drawdown step. It is shown read only, and these limits can only ever be made tighter.
• Settings that belong to another screen (the approval line, posture, risk limits, Guidance) appear as read-only chips that link to the screen that owns them; nothing is editable in two places.

Each value carries a quiet mark saying where it comes from: the role's own definition, a change you made in the console, or a value fixed for the whole platform. A platform-wide value wins over a console change and the page says so. Consult's default model and caps are set for the whole platform and its model is picked per message, so its page explains rather than edits.

Saving opens a diff of exactly what changes, and the change is audited. Consult's model is picked per message from the menu under the composer on Consult: Sonnet 5 answers in seconds for cents, Opus 5.5 goes deeper for tens of cents, Fable 5.1 takes the hardest questions and is the slowest and dearest; Consult can also draft a cadence change as a card you confirm, applied with the same bounds. A change applies from the role's next session and is audited.

Model budgets are set with each agent's cadence and models on its Settings tab on Desk (Portfolio Manager › Settings). Your account's Settings screen holds no budget control. The per-session cap is never allowed above the daily cap.

When a change takes effect

Every session is short-lived and nothing of a model keeps running between them, so a saved cadence or model applies at the next session with no restart. A session already running finishes on the model it started with.

What the policy will not do

• No role ever runs on a model outside the allowed list, and the console refuses a model it does not know.
• Sessions stop at their dollar cap mid-turn rather than overrunning; the day stops at the daily cap.
• Changing a model changes how a role thinks, not what it may do: the tool allowlist, the memory it may write and the Executor's checks are the same for every model. Each role has one definition, and that definition alone decides which tools a session of that role is given, which it may call and which memory it may write; a kind of session the definition does not name gets read-only tools, never the PM's by default. The tool counts on the agent's Settings tab are read from the same definition, so what it shows is what the session gets.