This is what it does
Try it
The Assistant Roster is nine small research assistants, each given one job, reporting to a chief of staff that will not let a number stand unless somebody measured it.
Nothing to run here — internal research, not a service
There is no demo on this page because there is nothing to demonstrate. The nine are command-line research scripts run on one machine against recorded data, and the ninth is a brief rather than an interface. None is hosted, none is downloadable, and no order has ever been placed by any of them.
What the page can do is show you the record: which jobs have a success condition that has been measured, and which are still unmeasured.
- What it is: nine narrow tools plus a written discipline — each assistant states its success condition before it is allowed to claim anything, and one thesis has already been falsified by measurement.
- What it runs on: recorded public prices and order books, append-only outcome logs, and a paper ledger of simulated fills that is labelled fictional in its own output. No model sits in the reporting path.
- What it does not do: it holds no positions, places no orders and moves no money. That is the design, not a stage it is passing through.
- Where it stops: nothing on this page is runnable. The record is the interesting part, and the feedback link at the bottom is the way to ask about it.
The record
Nine jobs, and whether their success condition has been measured
Every row carries its own success condition, and the state beside it is whether the project has measured that condition. An unmeasured row is not a failure; it is the honest answer until the measurement exists.
Where an assistant has been pointed at recorded data, the answer is often that it found nothing. One thesis was falsified on five independent routes; one buyer scores zero candidates by design because it invents no base rate; below five acted decisions the report prints counts and withholds the percentage. The measured figures are in the numbers below.
The numbers
Constants from the build
The build keeps more counts than these, per assistant, in its own self-tests — the chief of staff runs 82 controls, the payment operator 99, the operator-input form 28 refusals, the records lookup 25. Each is printed by the command that runs it rather than quoted from memory.
The rule that holds
It will not report a number nobody measured
- One thesis was falsified by measurement. A strategy that looked persistent on paper was run against recorded data across five independent routes; all five failed, for five different reasons, and the record says so.
- One assistant scores zero candidates by design. It refuses to invent the probability it would need, so with no supplied base rate it buys nothing and says why.
- Small samples are printed as counts, never as rates. Below a stated threshold the report gives the raw numbers and withholds the percentage.
- The paper ledger is never reported as profit. It is labelled fictional in its own output, because a simulated fill is not a result.
Under the hood
What it is built on
- Stack
- One Python script per assistant plus a registry file that names and locates each of them, a recorder that writes public market data to disk, and a local dashboard that reads the records rather than re-describing them. No model sits in the reporting path.
- Surfaces
- A command line per assistant for its plan, a tuning brief the chief of staff writes, and a local dashboard served on the loopback interface only.
- Runs where
- On one machine. The project’s single network client is a separate read-only recorder that lives outside the assistant folders, so the folders themselves stay offline by construction.
- Data
- Recorded public prices and order books, each stamped with the moment it was taken; append-only outcome logs; a decision register; and a paper ledger of simulated fills labelled fictional in its own output.
- Payments
- None. No money is connected to any of it, and nothing here is sold, subscribed to or invested in.
- Tested by
- Per-assistant self-tests — 82 controls for the chief of staff, 99 for the payment operator, 28 refusal checks for the operator-input form — plus a safety gate that scans every script inside every assistant folder for a network client, an order send or a key path and fails the run if it finds one.
Status
Where this actually stands today
- It is internal research, not a product, not a service and not a fund. There is no sign-up, no subscription and nothing to buy on this page.
- It trades nothing. No order has ever been placed by any assistant here, and no money is connected to it.
- Several of the nine are blocked on a human input that no amount of code closes — a base rate, a margin floor, a channel decision. Those are open decisions and the brief names them as such rather than inventing a proxy.
- Where an answer is that a strategy does not work, that is recorded as the outcome rather than tidied away.
Plainly
What this is not
- It is not investment advice, and nothing in it is a recommendation to buy, sell or hold anything.
- It makes no claim of profitability. Its own records state that it cannot, and it is built to refuse that claim.
- It is published here as a description of a method — how a small system behaves when it is made to answer for its numbers — and not as an offer of any kind.
Questions about the roster
If you want to know how a particular finding was measured, or why an assistant refuses to answer, ask — the record is the interesting part.