Proposes single-instrument US equity theses based on transparent quantitative signals (RSI, 20d return, realised vol) and sanitised news flow. Never trades directly.
Every agent in the fleet, cataloged and versioned.
Discoverable by role or by desk scope. Each entry pins model, prompt hash, and policy version. A demoted agent can be rolled back deterministically. This is the primitive that makes the fleet deployable across research, risk, and execution desks under one governance regime.
Deterministic policy evaluator. Runs kill-switch, universe, conviction, and exposure checks against the InvestmentMemo. LLM narration is applied post-verdict and cannot override it.
The only agent identity with an Alpaca API key. Receives approved theses from the orchestrator and places paper-trading orders. Cannot propose. Cannot review.
Read-only surface into the receipt ledger. Verifies the hash chain, exports audit bundles, and powers the auditor dashboard.
Google TimesFM 2.0 (500M) zero-shot probabilistic forecaster. Consumes recent price series and returns p10 / p50 / p90 quantile forecasts over a configurable horizon. Deterministic for a fixed model version and context window — the forecast is a reproducible audit primitive, unlike LLM narration. Snapshotted into every receipt so an auditor can replay signal → forecast → thesis → verdict → fill end-to-end.
