PRBot, run through the published rubric

The full teardown is being written against the live PRBot codebase and publishes here, warts included. A diligence sample only counts if it is real, so it ships when the claims file does. Below is the rubric it follows.

• Claims inventory

  Every AI capability the company asserts, restated as a falsifiable claim with a named source.

• P&L wiring

  Which claims connect to revenue or cost in the data, and which are demo theater.

• Code-level verification

  The actual repositories, read directly: architecture, vendor exposure, what breaks at scale.

• Eval discipline

  Whether outputs are measured against a locked evaluation set, or nobody can say if the system works.

• Data and workflow readiness

  Whether the data layer and operating workflows can carry the AI roadmap being sold.

• Drift and dependency exposure

  What happens when models, vendors, or prices change, and whether anyone would notice.