The Pledge

Most AI-generated content starts from a carefully engineered prompt — persona definitions, tone instructions, guardrails, output formatting. The goal is to make the AI say something specific. That's still a human writing tool, just automated.

We do the reverse. We present our current best practices to the AI and ask it to brainstorm. No prompt engineering. No context. No personas. The AI is pushed to its limits, unconstrained, and we publish what it tells us verbatim, no matter where it lands.

If the AI fabricates a non-existent CVE, we publish that. If it finds a genuine blind spot, we publish that. The whole point is to remove ourselves from the feedback loop. Steering the output would defeat the exercise.

"No context, no memory, no personas" is easy to promise and hard to actually enforce. We found that out directly: an early test ran the judgment prompt through what was supposed to be a blank-slate AI agent, and it still knew who we work for and quietly saved a note about the topic to an internal knowledge base before answering. The response itself was fine. The isolation wasn't real.

The problem is structural, not a one-off mistake. Coding agents and IDE assistants attach user-level rules, memory tools, and identity context to every session they start — including a subagent spun up specifically to "start fresh." Sandboxing the session doesn't fix it, because the rules are declared to apply to any agent running under that account, not to a particular runtime instance.

So the judgment for every post is generated by a small script that skips agent runtimes entirely: one stateless HTTPS request, no system message, zero prior conversation turns, no tools except an optional web search so "go find it" can actually look something up without also opening a door to memory or file access.

A clean request isn't enough on its own — the machine running the script still has to actually be clean. So the script ships as a public Docker image, built and published automatically by a GitHub Actions workflow anyone can inspect. Every judgment runs in a fresh docker run --rm container pulled from that image — new filesystem, no volumes, destroyed the moment it exits. There's no machine state for one post's generation to leak into the next.

Model selection routes through Requesty, a gateway sitting in front of 300+ models from different providers, so which model judges a given post is a single flag, not a separate credential or code path per vendor.

The script, the Dockerfile, and the workflow that publishes it are all public in the repo. You can read exactly what gets sent, pull the exact image that sent it, and check the build logs yourself.