AI assist

What exists today

Worth being honest about the starting point: the only AI in the system now is the parameter optimiser (random, Bayesian TPE, and an LLM-driven mode) plus one entity-resolution bot that matches equivalent questions across venues. Neither is a user-facing assistant. All three surfaces below are new work.

Surface A — trade determination

Answers: "I don't know what to trade."

Takes the user's stated interest and the current market state, and returns a ranked candidate list with reasoning. It does not decide; it shortlists, and every item is traceable to the data that produced it.

interface TradeSuggestion {
  marketId: string;              // must resolve; hallucinated ids hard-reject
  side: "buy" | "sell";
  conviction: number;            // 0..1, calibrated, not vibes
  rationale: string;             // plain English, shown verbatim to the user
  evidence: Array<{             // REQUIRED. No evidence ⇒ no suggestion
    kind: "signal" | "book" | "news" | "resolution_rule";
    ref: string;                 // signal name, snapshot id, article id
    observedAt: number;
  }>;
  suggestedSizePctOfBankroll?: number;   // a hint; the chain decides
  modelVersion: string;          // pinned, logged, comparable over time
}

Rules

• Every market id is validated against the registry before the suggestion is shown. A model inventing a plausible-looking id is the expected failure, not an edge case.
• Evidence is mandatory. A suggestion with an empty evidence array is dropped, not shown with a caveat.
• Conviction is calibrated and measured. Log suggested conviction against realised outcome. A model whose 0.8 is not roughly 80 % is worse than no number, because users will size on it.
• The user or their bot acts, never the surface. Accepting a suggestion creates a proposal, which enters the chain like any other.
• Prices and sizes are never asserted as facts from model output. They are read from the book at proposal time.

Surface B — placement assistance

Answers: "How do I actually get this on without paying up?"

Given an approved intent, suggests how to work it: passive or aggressive, one clip or several, limit placement relative to the touch, timing around known events. Bounded by construction — it operates strictly inside the envelope the chain already approved.

Overlaps deliberately with exec.smartrouter and exec.antitoxicfill. Those are deterministic and stay authoritative. Surface B explains and proposes alternatives; it does not replace them. When they disagree, the deterministic utility wins and the disagreement is logged — that log is how we find out whether the model is adding anything.

Surface C — position, correlation, and risk monitoring

Answers: "What am I actually holding, and what could hurt me?"

Read-only, Governance authority, and the surface users will value most. Turns a position set into something a human can act on.

• Concentration. "Four of your six positions resolve from the same oracle."
• Hidden correlation. Reads intel.crossmarketgraph and says the thing the user cannot see: that six positions are one bet. This is the highest-value output on the whole surface.
• Drawdown trajectory. Distance to the user's own limit, and what happens at it.
• Resolution risk. Ambiguous rules, disputed oracles, markets resolving imminently.
• Behaviour drift. "This bot's win rate has been falling for two weeks while its size has been rising."
• Plain-English decision explanations. Why a bot did or did not act, from the reason codes rather than a free-text reconstruction.

Failure modes, and what we do about them

Where AI must not be used

• Inside a guardrail decision. Guardrails are deterministic and replayable; a model in the chain destroys both.
• As the signer, or anywhere in the signing path.
• To compute exposure, drawdown, correlation clusters, or reconciliation. Those are arithmetic. Use arithmetic, then let the model explain the result.
• To generate reason codes. Codes come from the registry.
• To decide whether a user bot may go live.