# FuturesDesk

FuturesDesk is a sports-market intelligence platform. It separates observed market evidence from inference and does not treat movement, public concentration, promotion, or handle as automatic proof of edge, sharp action, or sportsbook liability.

## Agent path

1. Read this file and the public capability manifest: https://api.futuresdesk.io/.well-known/futuresdesk-agent.json
2. Read the OpenAPI contract for callable public tools: https://api.futuresdesk.io/agent-openapi.json
3. Check every response for freshness, provenance, evidence state, and authority limits before using it.
4. Preserve sportsbook, jurisdiction/context, exact market, line/price, and observation timestamp when present.
5. Fail closed when required evidence is missing or stale. Do not upgrade uncertainty into certainty.

## Agent harness boundary

FuturesDesk exposes typed, allowlisted capabilities rather than raw database access. Public agents do not receive SQL, secret, deployment, publication, wagering, spending, or account-mutation tools. The public surface is API-triggered only; internal schedules/events continue to use the existing governed automation estate rather than a duplicate public scheduler. Model/provider selection is not delegated by this public surface.

Internally, provider-neutral inference, model/cost routing, receipts, and runtime health already have canonical FuturesDesk controls. Do not create a second generic gateway, model registry, telemetry plane, scheduler, or harness merely to imitate another architecture. New infrastructure is justified only by a verified semantic gap.

## Public agent-safe surfaces

For generic agent acceptance, current-state checks, or provenance/freshness verification, prefer the read-only Heisman GET path below. Do not enter human-facing UI or the Market Tape form unless the task specifically requires user-supplied tape classification. A bounded failure/self-repair check can use an unknown candidate token on the same Heisman GET path and then follow the returned fix.current_sample guidance.

- Heisman Market Topology: https://api.futuresdesk.io/api/heisman/market-topology
  Preferred machine capability for generic acceptance/current-state checks. Start with https://api.futuresdesk.io/api/heisman/market-topology?limit=1. Canonical read failures return a stale-labelled 503 rather than pretending the fallback is current.
- Market Tape Decoder: https://api.futuresdesk.io/api/agent/market-tape-decode
  Read-only deterministic POST tool. It classifies user-supplied tape evidence; it does not independently verify sportsbook data or establish betting edge.
- Human-facing Market Tape Decoder: https://api.futuresdesk.io/market-tape-decoder
- Human-facing Heisman Market Topology: https://api.futuresdesk.io/heisman/market-topology

## Evidence doctrine

- Exposure != Edge.
- Promotion != Liability.
- Movement != Flow.
- Ticket share is wager-count concentration, not handle.
- Handle concentration is not automatically liability or sharp action.
- A value or execution claim requires an independent fair-price comparison and current executable evidence.
- Different books, jurisdictions, markets, terms, or timestamps are not silently merged into one movement claim.

## Authority and safety

Public agent surfaces are research-only. They grant no wagering, publication, spending, account, secret, deployment, or production-mutation authority. Agents should not interpret a successful API response as authorization for a consequential action.

No public MCP server is currently advertised. Do not infer MCP availability from this file. A discovery card will be published only after a public MCP runtime is independently verified.

## Errors and retries

Public APIs return structured JSON errors. Repair invalid input using the response detail and the OpenAPI contract. For 429 responses, honor Retry-After when supplied. For 5xx or stale-data responses, preserve the failure state rather than substituting unverified data.

## Citation

When summarizing a FuturesDesk response, include the exact endpoint, returned freshness/provenance fields when present, and the observation or generation timestamp. Distinguish observed facts, supported inference, hypothesis, and execution decisions.