Pip — Prediction-Market Forecasting Agent

The brief

Forecast public prediction markets, in writing, in advance, with the same instrument and the same gates every day — no improvising on a hot streak, no flinching on a cold one. The goal was not to win a single market. It was to publish a track record honest enough that a buyer could grade it.

Pip’s edge thesis is documented in the playbook: read literal resolution criteria, find markets where crowd sentiment misprices the strict letter of resolution, write the trade only when the documented risk gates pass.

What we built

Four LLM roles — Analyst, Researcher, Trader, Risk — each behind a structured tool-call schema. A fourteen-gate risk middleware sitting between Trader and exchange, deterministic, LLM-immune. A Supabase ledger of every decision, every gate fire, every reconciliation. Daily mark-to-market through a recovery loop that closed the audit gap in May, when missing exits — not bad signal — were the actual blowup.

On June 1, 2026 the Researcher became an ensemble. Five concurrent trials at temperature one, Platt √3 sharpening on the mean log-odds, a supervisor agent that only fires when the trials disagree, and a final 0.33/0.67 blend against the market prior. Baseline single-trial Brier was 0.111. The published architecture’s reported Brier under similar conditions sits at 0.106. Expected post-ship range: 0.085 to 0.095.

The receipt

Pip’s playbook reads like a postmortem because that is the operating mode — every losing trade lands in writing, every fix lands in code, the agent that survives is boring on a per-trade basis. The forecasting stack documents both the wins and the audits where capital recovered when the exit loop closed.

What this signals for a buyer: an agent that holds its discipline across thousands of decisions, that grades itself against a public market, and that publishes the gate fires in the same channel as the wins.