Paul Starrett

Credits are decided on the evidence: the record of what was uncertain, what was tried, and how it was evaluated. That record is easiest to capture while the work is happening; months later, people have moved on and systems have changed. A record built during the work is the strongest one there is.

• While the research is happening: setting up how evidence of experimentation is captured from the systems your engineers already use, and a records retention policy that keeps it for the life of the credit, not a default three years, so the claim can be proven later
• Before you claim or amend: characterizing which work qualifies, and testing whether the record proves it, before an examiner does. Reviews of prior claims are conducted under attorney-client privilege
• When the IRS or FTB examines: defense from the first document request, built on what the engineers actually did
• When you raise money or sell: whether the credits in the deal hold up, and what the exposure is, including how AI assets are characterized in the acquisition

A record built while the work happens serves more than one purpose.

• AI governance: whether a system does what its policy says it does, documented before a regulator or plaintiff asks
• Disputes and investigations: reconstructing what a system actually did when its outputs are challenged, as counsel, consulting expert, or testifying expert witness