Under the hood, Sharp is Edge Desk — a small agency of domain-specialist AI agents that watches Polymarket order books and fresh real-world evidence, flags markets whose odds lag reality, and ships a cited alert. It never places the trade for you. Here is the exact loop, end to end.
Every alert is fully traceable — each step below persists its inputs, status, latency and cost to PostgreSQL, so any run can be opened end to end.
A live event feed (a goal, an injury, breaking news), a manual market check, or a versioned replay fixture hits the ingest API. Every event carries a stable sourceEventId for deduplication and keeps occurredAt separate from receivedAt.
The API validates and writes the event plus an initial run durably and idempotently, then returns 202 Accepted with a runId — it doesn't block on the agents. Re-posting the same key returns the existing run with duplicate: true.
The Hermes orchestrator loads market config and baseline, then routes the event to one domain specialist — sports, geopolitics or crypto. Reading a red card is nothing like reading an election injunction, so each is its own agent, yet all emit the same normalized signal contract.
Shared adapters fan out at once: Polymarket snapshots (price, best bid/ask, spread, depth) over the CLOB WebSocket, and fresh cited evidence from Linkup or a live feed. Every fact keeps its source and retrieval timestamp — a current price alone can never establish lag without a pre-event baseline.
A hybrid lag detector joins the signal to the market outcome. Deterministic code owns the price math, thresholds and final safety gates; a model assists only with feature extraction and the human-readable reason. It emits notify, ignore or needs_review — never a trade.
A notify decision writes an alert and an outbox row in one transaction. A sender worker claims the job with row-locking, composes the card and delivers it through the Hermes Telegram gateway — an idempotency key stops two workers double-sending the same alert.
Once sent, the tracker schedules price checks at +10, +20 and +40 minutes from the send time. Each labels the call correct, wrong, flat or invalid_data under a versioned policy — and records lateness rather than silently shifting the horizon.
The public scoreboard reports directional hit rate at each horizon, median move and sample size — with a versioned metric definition, and live results kept separate from replay. No cherry-picking a favorable horizon.
Everything is arranged around one hub: PostgreSQL. Triggers enter through the API; market data and evidence enter through Hermes; specialists write their signals to the database, and every downstream worker reads from it. Follow the arrows.
The matcher's job is one question: has reality moved faster than the market? It answers deterministically, in basis points, and only then does a model write the explanation.
lagBps meets the minimum in the predicted directionPulled straight from the money-scouts monorepo — a TypeScript workspace of focused apps and packages.
apps/* + packages/*events, marketChecks, replays202-fast dispatchjsonb + timestamptzsports-v1Trigger to cited alert to scored outcome — every step inspectable.
Start free trial