Equity committee
A $100,000 paper book of undervalued US companies. Six stages produce the portfolio, and they do not all run at the same time. Two run every night: one reads the market regime, the other screens about 500 companies. The remaining four run only on the eve of a rebalance, which is the one morning a month anything is bought or sold.
Every stage is ordinary code following fixed rules, and the code picks the companies. Language models write up the reasoning and cast a second opinion. Where a model disagrees, the disagreement is recorded and the rule is followed.
Book value
repriced nightly; dashed line is buy-and-hold SPY over the same window
- book
- buy-and-hold SPY
Holdings
weights drift with prices between rebalances
9 companies. The largest is NEM at 16.1% of the book, then CTSH at 15.1%. Weights drift with prices until the next monthly rebalance.
| ticker | weight | entry | last | change | value |
|---|---|---|---|---|---|
| NEM | 16.1% | 89.76 | 135.14 | up 50.55% | $16,776 |
| CTSH | 15.1% | 44.80 | 63.06 | up 40.75% | $15,684 |
| FOX | 13.0% | 51.67 | 62.69 | up 21.34% | $13,520 |
| EXE | 11.5% | 88.19 | 94.66 | up 7.33% | $11,960 |
| BBY | 10.7% | 85.47 | 85.29 | down 0.21% | $11,119 |
| TGT | 9.6% | 139.72 | 163.47 | up 17.00% | $10,029 |
| GD | 8.4% | 368.84 | 376.58 | up 2.10% | $8,751 |
| TXT | 7.8% | 91.54 | 87.13 | down 4.82% | $8,158 |
| ATO | 7.8% | 177.80 | 167.56 | down 5.76% | $8,078 |
How a decision is made
one night, start to finish
every night, 21:30 New York
- 01macro monitorreads the market regime from SPY and its moving averages
- 02analystscreens about 500 US companies, using only data that was public that day
These two run so the daily report has a regime reading and a watchlist. They do not produce a trade proposal.
eve of the rebalance only
- 03quanttests the locked strategy over a rolling window of history
- 04reviewerrejects the cycle if return, drawdown or universe size miss fixed thresholds
- 05portfolio managersizes each position under the sector and single-company caps
- 06risk monitorvotes to approve or reject the finished proposal
if the risk monitor rejects
- First and second rejection: back to the portfolio manager to re-propose.
- Third: back to the quant, with tighter parameters.
- Fourth: the cycle halts and nothing trades that month.
if approved
The proposal waits for a person to approve it. Nothing reaches the book overnight.
first trading day of the month, 08:30 New York
The approved proposal executes against simulated prices, paying spread, slippage and commission. This is the only step that moves the book.
after the fact, and outside all of the above
- llm reviewerwrites a second opinion on the night's proposal
- llm analystwrites a short case for each company picked
- llm riskcasts its own approve or reject vote
These three are language models and they run once the decision is already made. None of them can change a holding, block a trade or halt a cycle. If the llm risk vote disagrees with the rule-based risk monitor, the rule wins and the disagreement is recorded.
A worked example
the night of 17 July 2026, one company
21:30, 17 July 2026
the market regime is classified
SPY closed at 743.29 and sat above its long moving averages, so the regime was scored bull with 0.719 confidence. This decides how much the book is allowed to hold, before any company is looked at.then
about 500 companies are screened
Each one is scored on how cheap it is against what it earns, and on the quality of the business. Expand Energy came out near the top: it earned 14.8% relative to its price, generated free cash worth 9.8% of it, traded at 8.2 times earnings, and scored 9 out of 9 on a standard nine-point financial health test. It had also fallen 16% over the previous year, which is what made it cheap.then
the strategy is tested against history
Not this company on its own. Instead the whole rulebook, replayed over past data, to check it still behaves: 22.4% a year, a 1.21 return-per-unit-of-risk score, and a worst fall of 19% along the way.then
the reviewer checks the numbers clear the thresholds
Return, worst loss and the number of eligible companies all had to clear fixed limits. They did, so the cycle continued. Had any of them missed, the night would have ended here with no proposal.then
the portfolio manager sizes the position
Expand Energy was one of four names marked high conviction, so it was given 11.3% of the book against the 8.7% that medium-conviction names received. No single company may exceed the position cap, and no sector may exceed 30%.then
the risk monitor votes, and raises a flag
Approved, but not silently: it recorded that Basic Materials made up 20% of the book across two companies, and that those two tend to move together. Inside the limit, worth noting. A rejection here would have sent the proposal back to be rebuilt.then
the language models had nothing to add
All three advisers returned UNAVAILABLE that night. The API account was out of credit. The cycle completed, the proposal stood, and the trade went ahead unchanged, because the advisers do not sit in the decision path. The design held that night under a real failure.then
a person approves it
Nothing reaches the book overnight. The proposal waited for approval before anything was allowed to execute.11:38, 18 July 2026
it executes
126.35 shares of Expand Energy at $88.19, a position of $11,142.90, part of $90,000.10 put to work across nine companies. Simulated prices, simulated fills, and the book stood at $99,937.04 by the close.
Readiness
every check must pass before real money is considered
These are the conditions this system has to meet before it would be trusted with real money. It has not met them. Until every one passes, the verdict stays NO-GO and the book stays simulated.
not passedenough historyNeeds more history: 0 of 0 hours so far.
not passedresearch nightsMissed 1 research night.
missing: 2026-08-02
passedlosses inside the limitDrawdown is inside the −25% buying halt.
book 107961.37 · drawdown 0%
not passedadvisory reviewsAdvisory reviews are missing or could not be read.
llm_review_memo: UNAVAILABLE, llm_risk_vote: UNAVAILABLE