Polymarket research desk

Short-duration prediction markets settle every few minutes, which invites a quick conclusion from one good streak. I wanted a desk where a strategy reaches real money only after it passes rules written before the test, and where the part that trades is safe from the agents that help build and run it, including when I push them hardest.

I split the system by authority. The money loop is a deterministic daemon with no language model in it. Agents work on research and on a control plane that can observe and advise and holds no trading power by default. The statistics were written down before any learning loop existed, and candidate rules are frozen before they are scored. Claude and Codex agents built it in parallel worktrees, each run ending with a written report, test counts and an independent read-only review. On one handoff, Codex treated five Claude-built PRs as work from another author instead of trusting their completion reports: it merged them in a disposable worktree and ran the combined suite, 1,117 passed and 90 skipped, before anything reached main.

1. 01

   A money loop with no model in it

   A Python daemon under systemd ticks once a second and records spot prices, order books and the trade tape to SQLite. In shadow mode it quotes virtual orders, counts a fill only when the tape prints through its price, and settles at resolution. Going live needs three things at once: live mode in config, an ARMED flag file and no HALT file. Touching HALT cancels quotes, a daily-loss stop halts on its own, and cancels go by order id because the account can hold manual orders. The operator extension arms a pod only after the exact typed phrase ARM followed by the pod name, and a single-instance guard went in after two daemons once drove the same pod.

2. 02

   A control plane that fails closed

   My agent, Jeisan (it has its own page), runs on NousResearch's Hermes Agent on a separate host and reaches the pods only through a chain of least-privilege hops that ends at a strict server-side parser. The only production mutation is performed by a root systemd timer with no model in it. A live start would need an expiring, non-replayable mandate, root-attested config hashes and fresh signed exchange evidence, and the policy and mandate ship disabled, so installing the control plane cannot arm or trade. Separate watch jobs hold a sticky HALT that overrides the manager.

3. 03

   Statistics before the learning loop

   Before building a loop that retunes itself, I wrote the statistical guardrails down as an issue and a document. They set how much evidence a change must carry before it counts, and the checks are pure functions in code. The decision layer is limited to operational facts and pre-registered experiments.

4. 04

   A research fleet that checks itself

   Four parallel Opus lanes researched at once and checked each other's work. A backtest from one lane was tested against another lane's live, millisecond-timestamped book recording, and the check exposed two data traps that then went to every lane: stale book snapshots, and trade timestamps that are block time, about two seconds late. Candidate rules were frozen and hash-stamped before scoring, then scored once against a pre-registered bar.

5. 05

   A simulator that has to prove its fills

   A Codex research sprint reworked the simulator so that each simulated fill has to prove its price was tradable, using exchange timestamps and trading status. The fill, quote-age, fee and settlement accounting are pinned by 70 execution assertions, 50 trading-invariant scenarios and 26 tests for time-weighted average prices. A secret scanner in CI blocked two research bundles from publishing.

Results stay private: balances, positions, fills and which strategies survived are not on this page, and no figure here is a return. The strategies that passed run in pods behind the kill switches above: shadow until armed, armed only by the typed phrase, halted by a file or by the daily-loss stop. The engineering is checkable: a control plane that ships disarmed, a simulator whose fills must prove their prices, and a written method for testing the next idea before it touches money.