Public-repository study / Brex Substation
A Narrow Yes Is More Useful Than an Agent Score.
We independently reconstructed three historical workflow families in Brex’s public Substation repository. Two exact coding-agent deployments completed and reviewed each workflow three times. Both qualified across all 18 decisions.
What was approved
| Reconstructed workflow | Held-out result | Operating decision |
|---|---|---|
| Preserve every digit in large financial identifiers | 6/6 passed | bounded delegation |
| Route records with missing fields correctly | 6/6 passed | bounded delegation |
| Process large records without silent truncation | 6/6 passed | bounded delegation |
This was not a declaration that the agents were safe for arbitrary work. It was a limited approval for three exact workflow families, under one frozen deployment definition and one fixed evidence set.
Why an all-pass result is useful
A control system that can only reject work becomes a brake. A useful system must also show where autonomy has been earned. Here, repository tests, disclosed checks, held-out checks, independent review, and frozen handoff all agreed. The appropriate decision was a narrow yes.
Both model-role pairings finished 9/9. At this sample size, that is a consistency result—not evidence that one model is globally superior. The output that matters is the workflow boundary, not a leaderboard.
How the study was run
One Claude model repaired the pinned public repository and the other reviewed the exact frozen result. We reversed the roles, repeated each pairing three times per workflow, and evaluated all candidates against rules established before execution. Failures, had they occurred, would have been preserved rather than retried away.
Inspect the basis
Public source repository · plans and checks · Veyl methodology
This is an independent reconstruction from public brexhq/substation history at pinned commits. It is not a Brex engagement, customer result, production assessment, or statement about Brex’s private systems. A real deployment decision requires company-selected workflows, private evidence, and company approval.