The two-book experiment — history¶

This page is history, not the current agent. It records the agent that ran from 2026-08-13 to 2026-08-25, as one half of what this project called the two-book experiment: a real-portfolio book that mirrored a live Webull account, run beside this agent, which traded a fixed watchlist it did not choose for itself.

Both books ended on 2026-09-01. The commit before that removal is tagged v1-two-book-experiment. The agent below merged with the analyst experiment into the single agent of experiment 1 — see the experiment 1 journey for its record.

Newest first, the same as the current journal.

2026-08-25 — the agent may move its own exits. A side: "adjust" action, with a new stop, target or both, on something already held. Before this, exits were fixed when a position opened and untouched until it closed, so re-reading a holding every morning taught the agent nothing it could act on short of selling: GOOG spent a week with a $377.09 take-profit while each day's analysis put the end of the move at $345.00. Python still refuses a level that would execute on placement, and a level already where it was asked for is skipped rather than re-sent.

2026-08-25 — holdings now show what is resting under them. The prompt lists each position's live stop and target, and says NOTHING is resting to close it when there is none. Added with the adjust action, and required by it: the agent cannot sensibly move an exit it cannot see.

2026-08-25 — signals are filtered to the configured model. Running a second model for comparison puts two signals per ticker in the table, sometimes disagreeing. Without the filter the agent traded on the mixture, folding an experiment into the live book.

2026-08-25 — a conviction floor, switched off. Minimum chance of working and minimum risk/reward, both defaulting to zero. Off deliberately: the chance of working is the model's own claim, and until the Scorecard's calibration says it is honest and that it sorts outcomes, a threshold on it is arbitrary discipline. A signal stating no number fails the floor rather than passing it, or the floor could be dodged by not answering.

2026-08-13 — buys go out as brackets. Previously the exits were armed after the buy returned, which meant they were validated while the account still held nothing and read as a new short. Two positions were bought that day and neither got its exits.

2026-08-13 — an ATR stop is derived when the stated one is unusable. Both of that day's unprotected positions were bought days after their signal, by which time the price had fallen through the stated stop and the level was correctly discarded — leaving nothing. record_signal already substituted an ATR stop, but only for Buy and Overweight, and the agent buys on Holds too.

2026-08-13 — an unguarded position is announced. It used to be silent: no alert, no ledger row, and the only way to find out was to look at the broker.