Method
An autonomous pipeline screens a universe, argues with itself about each name, sizes what survives, and closes positions by rule. The language models in it argue and one holds a veto; none touches a risk limit or a position size. What the fund measures is published in full; how it is built is not.
Loading the published snapshot the figures below read from…
One name in, five readings, one position
A prism splits one input into its components. That is literally what the quorum does to a ticker: the claims in its thesis are divided between five synthesis lenses, and the five shares recombine into a single conviction. Pick a name and the split is its own.
5 synthesis lenses · no published theses in this snapshot
The five lenses are not independent experts. They are five prompts over largely the same evidence, and the published cells show they divide the evidence rather than cross-check it, so almost every claim carries exactly one lens’s endorsement. Five bands leaving a prism are five slices of one reading, never five confirmations of it — and W146 found the quorum unproven against its cost. The split is real; the independence is not.
How a decision gets made
Five phases from the screen to the exit, and which of them a model may touch. Only one is highlighted: the risk veto is the single language-model layer that has measured a positive signal — retrospectively, on n=6. The forward measurement of that same claim is the card directly beneath this one, and it does not support it yet.
Three of the five phases are deterministic code, not model output. That is the central invariant and the reason the risk figures are worth anything: a limit cannot be moved by anything a model says, and position sizes are arithmetic.
Read the pipeline as text
- 1Screendeterministic codeDeterministic and periodic, outside the per-name loop. Cuts roughly twelve thousand listed names to a tradeable universe of a few dozen.
- 2ArgueSeveral independent readings of each name, which are made to disagree before they are combined — and which may refuse to form a thesis at all. Readings that carry no weight in the live blend are still scored, so their information coefficients stay measurable; the research page publishes them. The most expensive part of the chain and the least proven: W146 measured it at about zero and kept it anyway, because nothing justified cutting it either.
- 3Risk vetoThe guardian pass, and the one language-model surface that has measured a positive signal: the names it vetoed went on to return -4.51% forward alpha, approve-minus-veto +0.034 at t=2.26 on n=6.
- 4Size and limitdeterministic codeDeterministic arithmetic, conviction-tiered, with a 5% cash floor. No model output sizes a position or moves a risk limit. Long-only: no shorts, no margin, no derivatives.
- 5Exitdeterministic codeBy rule rather than by judgement. The portfolio manager holds no exit lever.
Invariants
Risk limits and exits are code, never model output
Position sizing and every risk limit are deterministic arithmetic. A language model can express a view and can veto; it cannot move a limit, and it cannot size a position. Positions close by rule rather than by judgement, and the portfolio manager has no force-exit lever.
Long-only, paper, virtual capital
No short positions, no margin, no derivatives. The fund trades on live market data with virtual capital, and every figure on this site is from that paper run.
No look-ahead, enforced in three places
Three independent checks, at different layers, rather than one convention — so a single mistake cannot reintroduce it.
Every decision is logged before it is acted on
Each trade decision is written with its full reasoning to Postgres before it is acted on. The scorecard and research pages are reads of that ledger.
How this site gets its numbers
A field reaches this site only if the contract names it
Redaction is an allowlist, so a new database column is invisible here by default. Machine-written prose about named public companies is not published at all; the structure it produced is.
Stale is shown, never hidden
The fund runs on a laptop that sleeps. When the last publish is old, every page says so above its figures rather than serving yesterday’s numbers as today’s.