Roostcapital

• Roost Preflight Finance Action — Preflight a SINGLE outbound finance action — an invoice, purchase order, or transfer identified by `vendor` + `amount_cents` + `external_ref` — before you send it. Call this before any outbound payment, transfer, payout, or spend of this shape; it scopes to exactly these fields, not arbitrary action payloads or multi-step plans. Flags likely duplicates, split bills and amount anomalies against earlier checks under the same anonymous token, and returns `status`: PROCEED
• Check Payment — Check a payment before sending it. Compatible ALIAS of `roost_preflight_finance_action`: same parameters, same backend call, same envelope superset — existing callers keep reading `action` / `verdict` / `reason_codes` / `flags` while new callers read `status` / `assurance`. Advisory only; you decide. Prefer `roost_preflight_finance_action` for new integrations.
• Roost Inspect Result — Retrieve a result Roost returned earlier, by the `request_id` it reported. Reads Roost's durable, ownership-scoped result store — only results created under your own identity are retrievable, for as long as Roost retains them — and returns the stored envelope unchanged. An unknown, expired, or not-your-own `request_id` returns an `unavailable` result (no `status` field) rather than an error or a fabricated result. The `filter` parameter is accepted for forward compatibili
• Roost Compare Results — Diff two results Roost returned earlier, by their `request_id`s. Use it to verify a fix landed or to track drift across runs — not for a first-time assessment (use `submit_eval`) and not to clear a live action (use `roost_preflight_finance_action`). Both results must exist in Roost's ownership-scoped store under your identity and share a result type; any miss or mismatch returns an `unavailable` result instead of a comparison.
• List Eval Scenarios — List finance-action test scenarios from a pack. Each carries a `context` situation; have your agent decide proceed/hold/flag per scenario, then call `submit_eval`.
• Submit Eval — Submit your agent's decisions (`responses`: [{scenario_id, action, reason?}]) and get a scorecard envelope: a top-level `status` (BALANCED/UNSAFE/OVER_CONSERVATIVE/MIXED/NO_DATA), component `metrics` (coverage, correctness, consistency, severity-adjusted risk, unsafe-proceed and conservative-friction rates), every failing case under `evidence.failed_scenarios` with a fix, and an `assurance` block stating what Roost did and did not independently verify. Branch your con

Schemas and invoke metadata: Directory API