Above 90 percent accuracy is the expected figure for the free OSS scanner. That number does not prove the scanner is safe to run on production repositories, because the document says reports ship without human review and may contain errors. It also omits which model, price, context, weights, or benchmark produced the result.

Die Brief reads this as a policy move, not a silicon spec. The constrained party is the maintainer, who receives automated vulnerability reports from a lab program that names Claude but leaves the model's cost and benchmark unmeasured. For a buyer, the change is access to a free scanner; for the lab, the change is a trust surface it must defend.

The filing keeps the model price, context window, weights, and benchmark hidden, so an engineer cannot compare the scanner against a human reviewer or another tool. The missing failure rate, patch quality, and liability for a mislabeled report remain unmeasured.

Die Brief reads this as a policy move, not a silicon spec. The constrained party is the maintainer, who receives automated reports from a lab program that names Claude but leaves model cost and benchmark unmeasured.

The document leaves out model price, context window, weights, and benchmark.

The scanner is a trust surface for maintainers, not a measured model spec. Without price, context, weights, or benchmark, an engineer cannot compare its reports against a human reviewer or another tool.