AI adjusting the books

v3 · 4–17 Aug · AI adjusting the books · see v6, the current engine →

v3 was my first attempt at putting a language model in the trading loop, and the most useful thing it produced was a reason to take it out. The model sat beside five books, could only veto entries or move parameters inside bounds written down in advance, and each of those books ran next to a twin that never changed.

Figure 01 · v3 · 4–17 August 2026

A model beside five books, inside written bounds

The model could only veto, shrink or retune inside a charter. Retired on 18 August.

Data
Proposal
Next-open fill
Score

fig. 1 — v3. ① RSS headline titles feed a morning news brief. ② Two gaters read it and may veto or shrink the day's entries; three tuners may move parameters on Sundays. ③ A validator reads its bounds from each book's frozen charter and rejects anything outside them; with no gate file the book trades its plain algorithm. ④ Each AI book is compared with its untouched twin, and only the spread at 26 weeks was going to count.

What the model was allowed to do

A gater could veto or shrink an entry but never add one, and it could touch at most half of a day's candidates. A tuner could change at most two parameters in a session, by no more than a quarter each, and never beyond half either side of the registered value. The validator read those limits from each book's charter and threw out a tuner proposal whole if any part of it failed. The model also sat outside the nightly path: the strategy used a gate file if one existed and traded its own rules if it did not.

What actually happened

On 4 August the analyst wrote one brief from 133 headlines, and one gater produced one file: eight candidates, all marked unclear, no vetoes. Three manual tuner sessions that day declined to change anything. From 5 August every scheduled session failed because the login they depended on had expired. The failures went into TODO: lines in a log nobody read, and the layer stayed dead for thirteen days before anyone noticed.

Why I retired it

I agree with the log's own verdict: “The defect was not the failure, it was the silence.” The model acted in one session out of about fifteen, so any spread against the twins was noise, and calling it an AI result would have been dishonest. The more lasting lesson was about where a model belongs. Frequent, low-stakes calls that a rule could make are the worst use of one, and anything a grid can decide should be decided by a grid. The five books and the two twins created for them were retired, not deleted, and the league went from 25 books to 18.

v4 replaced the model with a parameter grid and a stricter walk-forward. The twin idea stayed. Every agent result the engine reports today sits next to a paired control.

dates: 4–17 August 2026
AI books: 5, each paired with a frozen twin
model call: one turn, no tools, JSON out · about 18 a week
schedule (UTC): news 11:00 · gaters 21:40 weekdays · tuners Sunday 10:30
gater bounds: veto or scale 0.25–1.00 · never add · at most half the candidates
tuner bounds: ≤ 2 parameters a session · ≤ 25 % a move · within ±50 % of registered
judged at: 26 weeks, AI minus twin, net of costs (2027-02-01)
output: 1 gate file · 1 news brief · 0 scheduled tuner changes
retired: 18 August · league 25 → 18 books