Overview

Bushbush’s vertical is the internal running of an AI lab and its fleet: money, resources, and the daily state of the other agents. That vertical has executable ground truth because most of it is checkable after the fact. An invoice reconciles or it does not, a balance is a number, an application gets a real answer, and a claim that something is working is either supported by the record or it is not. She runs today, in beta, and she holds the lab to that record rather than to a confident guess. She reads, reconciles, and reports; she has no way to act on the outside world.

The thesis, applied to operations

Executable ground truth here is the part of operations that settles into fact: costs that reconcile against source documents, balances that can be read, applications that get a real answer, and statements about the fleet that can be checked against the measurement layer. The overseer runs on a different failure family than the agents she watches, so her read on the fleet does not quietly inherit their blind spots. Cross-family checking is a fixed rule: anything bound for a human gets a second read from a different family, and that second read must cite a concrete counter or let the work pass.

How it works

Fig. 1 · One daily read, and where it has to touch fact

Select a stage of the operations loop to read what it does

1. 1Read the record
2. 2Reconcile money
3. 3Second read
4. 4Deliver the brief

the record

She reads the fleet's own state and the lab's economics from the record: what each agent did and what it cost. The read is only as good as what it is checked against next.

ground truth

A figure about money is checked against its source before it is reported: an invoice, a balance, a statement. The source decides, not a fluent summary of it.

second read

Anything bound for a human gets a second read from a different failure family, which must cite a concrete counter or let it pass. A vague objection does not hold.

deliver

What survives becomes one plain daily read for the founders. She also prepares non-dilutive applications the lab qualifies for, verified against their own source, then hands them back for a founder to submit.

• ground truth, a source document
• the operations loop

The discipline is that nothing reaches a human without a cross-family check, and no figure is reported before it reconciles against a source. A single bottom-line briefing is synthesised over the fleet’s own state, but a briefing bound for the founders has to survive a second read from a different failure family, one that must cite a concrete counter or let it pass. Money is reconciled against source facts before it is reported. And where the lab qualifies for non-dilutive resources, she verifies each against its primary source and prepares the application, then hands it back: she never submits.

Where it stands today

Fig. 2 · What runs today, what is being earned, and the line she does not cross

• ◐beta

  Read, reconcile, second-read, deliver

  The daily read runs today, in read-only: she pulls the fleet's state, reconciles money against its source, passes a cross-family second read, and delivers one brief.

• ◷earning

  Trusted, not just running

  A week of hard testing turns "runs" into "trusted", and sets the accuracy bar a figure has to clear before anyone leans on it.

• ◷by design

  Acting on the world

  She has no send path, on purpose. She prepares and hands back; a founder decides and submits. That boundary is a design choice, not a missing feature.

She runs today, in beta and read-only: the read, reconcile, and second-read loop is live, and she has no send path by design. What is not yet proven is the week of hard testing that turns “runs” into “trusted”, and the accuracy bar money is held to before any figure is leaned on. The point is the discipline: that nothing reaches a human without a cross-family check, held to over time.

The bet is the fleet’s, applied to operations: a reconciled figure and a briefing that survived a second read are worth more than a fluent guess, because the first two can be checked and the third cannot.