Five investors rebuilt from their own writing run a real $10,000 portfolio against the Dow 30, in public.
A real $10,000 account runs the committee's decisions. Every week the portfolio value is recorded against the Dow Jones Industrial Average from a common baseline; the analytics below reconstruct the account day by day from the project's own position records and free daily market data. This is one line of evidence, not the verdict on the project.
Loading analytics…
These are the account's actual filled positions, reconciled weekly against the brokerage's own export. The committee's conviction-weighted target weights are published separately; fills drift from targets with market movement between rebalances.
The weekly record captures the account intraday; the engine values are end-of-day closes, so small deltas between the two are expected and shown, not hidden.
These are the committee's target weights: conviction-weighted from the latest screen, not the live brokerage balances. Buys are sized in proportion to conviction; the committee is long-only by construction. The actual filled positions (shares and market value) are tracked in the analytics above and reconciled against these targets.
Each investor is reconstructed only from their own writing: letters, books, memos, and reports. The question is whether a language model, grounded that way, reproduces a coherent and recognizable investment mind, and whether five of them can genuinely deliberate rather than simply average.
If the simulation holds, does the committee's decision-making translate into investment decisions worth acting on? A real $10,000 portfolio, benchmarked against the Dow 30 it selects from, supplies the evidence, measured over time rather than declared in advance.
Chosen to span complementary and often conflicting philosophies, so that genuine disagreement carries information. Each is simulated exclusively from primary sources.
The engine separates independent judgment from social influence, so the committee deliberates rather than merely polls.
Live market data on the security is gathered and shaped into a structured investment brief.
Each investor issues an independent verdict and conviction, without seeing any colleague's view.
Members see each other's positions and respond to the sharpest disagreement, revising if warranted.
A secretary writes the official committee statement: the verdict, the carrying argument, and the dissent.
The method, the source, and the findings are published so the work can be read, checked, and challenged.
The full apparatus: libraries, the committee engine, retrieval grounding, the cost economics, and the portfolio process. Published before results, against a fixed method.
Read the paper →The engine, retrieval pipeline, and tooling, published as a public GitHub repository for transparency and scrutiny. No API keys and no copyrighted corpora are included; the method is documented so it can be read and checked.
View the code on GitHub →Future write-ups examining fidelity of the simulation, the committee's structural tilts, and whether performance holds up to scrutiny over time.
Forthcoming