Open source as proof

I publish infrastructure because I do not want my own claims to be private.

A README that says "production-ready" without a coverage gate is advertising. A protocol that hides limitations is worse than a protocol that is honest about being experimental.

AgentEval is MIT, installable from PyPI, and runnable with no network for the flagship Failure Memory demo. KarmaSakshi Protocol is explicit: evaluation-ready self-hosted software, not a certified payment product. Agentic Data Analyst tells you the sample CSVs are small on purpose.

That posture is the work.

Open source is not a distribution trick. It is a forcing function:

• If an invariant matters, it has a test.
• If a threat is out of scope, the threat model says so.
• If a number is from a single run, it is labeled as such.

Credibility is not a tone of voice. It is an artifact someone else can run.