The platform
One system connects research,
conviction, and execution.
An event-driven trading engine supports historical tests and Interactive Brokers order execution. The research layer maps market forces and ranks portfolio ideas.
This page explains how the research, decision, portfolio, and execution layers work together.
The runtimes
Four runtimes maintain four distinct books
Each runtime coordinates a defined set of agents and owns one named book. This boundary keeps every portfolio action attributable.
| Runtime | Agents | Book | Orders | What it does |
|---|---|---|---|---|
| hedge-fund | 8 agent slots | MainBook | Research only | The broad idea desk. Macro, scans, catalysts, charts, trades, and a portfolio manager, in one chain. |
| breakout | 6 agent slots | ConcentratedBreakout | Execution enabled | A concentrated book of stocks making 52-week highs. Few names, high conviction, hard stops. |
| breakout-options | 6 agent slots | ConcentratedBreakoutOptions | Execution enabled | The same book expressed in long calls. A second portfolio manager sizes premium and picks the contract. |
| conviction | 11 agent slots | ConvictionBook | Paper, on purpose | The flagship. Narratives select the names, and the board holds them long before the chart is ready. |
The two breakout runtimes support execution. The conviction runtime remains in model status while it develops a scored decision history. All four books are model portfolios.
The continuous runtime
A continuous supervisor.
Work starts when evidence changes.
The conviction desk uses a continuous process. Every 60 seconds, a supervisor checks whether any of its 14 loops has eligible work.
Each loop has its own interval and its own wake condition. Budgets are hard capped: at most 3 agents run at the same time, and at most 24 agent runs happen in an hour. The cheap deterministic work runs constantly. The expensive judgment work runs only when something changed.
Deterministic. Samples every tracked name, measures volume and gaps, and raises candidates.
Curators lease work from the research queue. An empty queue means no agent runs at all.
Wakes only when another loop has filed an intent to change the board.
Re-orders the leaderboard 1..N. An unchanged board means it stays asleep.
What the desk remembers
Two stores, and one rule that binds them
Everything the desk knows lives in two places that persist between runs. Agents do not carry knowledge in their heads from one session to the next.
One record per company. What the business does, where the revenue comes from, which exposures it carries, and who its customers, suppliers, and competitors are.
Dossiers are written progressively across the 2,648 name universe, a batch at a time, so the desk gets deeper every day it runs.
One row per company and narrative pair. The same company can sit on the board twice under two different stories, and each row is judged on its own.
A row cannot exist without a dossier. That is the binding rule. Nothing gets ranked before the business behind it is written down.
Every name above $1B market cap.
Figures measured August 2026.
Where names come from
Three doors in, and one rule at every door
A name reaches the board from a narrative, from a scan, or from a sibling runtime. Whichever door it uses, it has to come from screening the universe — never from an agent's memory.
An agent takes a live narrative and expands it into a transmission map: who gets paid directly, who supplies them, and who benefits second-hand. Every name on the map comes from screening the universe.
The full universe is scanned with no cut applied. Deterministic detectors then flag the odd behaviour: a volume spike of 2.5x or more on a move smaller than 1.5%, a relative-strength crossing, a day-over-day jump in score.
The other three runtimes hand over what they find. A name the breakout scan surfaces can become a conviction row, and a conviction row can be handed the other way.
Grounding is enforced, not requested. Every factual claim on a row carries a source that was validated before the claim was written. A claim without one does not get stamped.
How a row is judged
Five axes, scored and written — and never added up
Each row carries five judgments. Each one is a score from 1 to 10 plus a written note that cites the stamped data behind it. A score without its note is invalid and gets rejected.
Narrative strength
Is the story real, and is it getting stronger or weaker?
Transmission
Does the money actually reach this company, and through which line item?
Expectation gap
Is this already in the price, or is the market still asleep?
Timing
Where is the chart, and how near is the catalyst?
Crowding
Who else is already here, and how much room is left?
The five scores are never summed into a total. There is no composite number and no weighting table. A weighted sum would let a strong story paper over a broken one.
Conviction is a judgment written in words: LOW, MEDIUM, or HIGH. The 1..N rank is a separate agent's judged order, and every change to it carries a written reason.
The debate gate
Every promotion faces an opposing case
A formal debate tests each row before it can enter a stage near capital.
An advocate argues the case. An adversary must present the strongest opposing case. A judge then rules and records explicit conditions.
Those conditions get stamped onto the row. Every later pass re-checks them. When a condition breaks, the row comes back down, and the reason is written next to it.
One agent is the only writer of the board. Every change is serialized through that single hand, so nothing lands anonymously and nothing lands twice.
Opportunity triggers
A measurement raises it. A judgment rules on it.
Alerting on price alone produces noise. So the platform splits the job in two, and only the second half reaches a member.
The deterministic sensor raises a candidate from evidence it can measure: a surge in volume, a one-day move of 4% or more, or a catalyst that lands within two sessions. No language model is involved.
A judge agent reads the evidence, the chart, and the tape, then rules in one paragraph. It either fires the alert or mutes the candidate with a written reason. Muted candidates stay on the record.
Execution and backtesting
One engine.
History and the live tape.
The code that replays years of market data is the code that places the order. A rule tested on history runs unchanged in the market.
Trades go out as 3-leg OCA bracket orders through Interactive Brokers. The entry, the stop, and the target are submitted together, so an exit exists from the first second the position does.
Every plan is checked against hard limits before it is submitted, and positions are reconciled against the broker afterwards. What the platform thinks it holds has to match what the broker says it holds.
Position size and stop distance are computed from the playbook formulas. An agent never invents a quantity. It judges whether to take the trade, not how many shares the maths allows.
The learning loop
Decisions get graded on a schedule set before the outcome is known.
Every decision is scored twice after the fact, at 10 and 20 market days. The dates are fixed in advance, so nobody chooses a flattering window.
A retro agent then grades the judgment against the mechanical counterfactual: what a plain rule would have done with the same row. That separates a good decision from a lucky one.
What survives becomes a proposed change to the playbook, opened as a pull request. A human reviews it and merges it, or does not. No agent edits the playbook on its own.
This loop has a short decision history. Its fixed review dates let that evidence grow without changing the method after an outcome.
See the system through its current decisions.
Open the public dashboard to follow the Book, Pipeline, Horizon, Leaderboard, Coverage, and Reports.