University of Chicago

AI Prophet turns LLM forecasts into live prediction-market orders. Ledger state is not held in memory between ticks: _live_ledger_state rebuilds it from the database each cycle, so anything the database fails to record, or records optimistically, becomes the bot's reality on the very next tick.

• Traced a state-divergence bug in the NET-flip path, which sells the opposite side before buying the new one: the BUY leg executed even when the SELL leg raised or returned an ERROR status. The old position was still open on the exchange and its cash had never come back, so the BUY over-committed available funds — and the next tick read a database that had logged the SELL as attempted, diverging from the exchange for good.
• Made the flip transactional from the caller's point of view: after catching the SELL exception, the routine now checks sell_status == ERROR and returns immediately with order_placed=False and an explanatory error string. The BUY is only ever attempted after a confirmed SELL.
• Found a second, quieter failure in _save_order(), which short-circuited whenever signal_id was None — always the case for make_trade() calls. Those orders never reached the database at all, so _live_ledger_state had no record of them and would cheerfully place the same position again on the next tick.
• Made BettingOrder.signal_id nullable so make_trade() rows persist without a parent signal, while the foreign key still constrains the normal process_forecasts path, then dropped the signal_id guard from _save_order() and left only the database-engine check.

DefaultBettingStrategy.evaluate decides which side of a market to take and how many shares to buy, by comparing the model's probability against the market's ask. Both halves of that decision depended on the assumption that the YES and NO asks sum to exactly 1.

• Showed that using (1 - yes_ask) as the NO price is only correct when yes_ask + no_ask == 1. On an overround book (ask_sum > 1) or a dutch-book (ask_sum < 1) both the NO edge and the NO share count came out wrong, which corrupted side selection and sizing together.
• Rewrote the decision to compute yes_edge = p_yes - yes_ask and no_edge = (1 - p_yes) - no_ask independently, pick the larger edge, and size the position by that edge at the actual ask price of the side chosen.
• Renamed the local spread in both DefaultBettingStrategy and RebalancingStrategy to ask_sum — it had always held yes_ask + no_ask, not a bid-ask spread. The public max_spread keyword stays for backwards compatibility, with the misnomer now documented.
• Added regression tests for the three cases the old code got wrong: the NO side on an overround book, the NO side on an underround book, and picking the larger-edge side.