Blog
Research, results, and what we found when we looked.
Research2026-04-14
Marcus Is Right About Neurosymbolic AI. He's Wrong About What It Looks Like.
Claude Code's print.ts is plumbing, not symbolic AI. Here's what principled neurosymbolic verification actually looks like — and what it catches.
Audit2026-03-15
We Ran Assay on Next.js. 601 Claims. 90 Bugs. 17 Failures.
601 claims verified across 6 server modules. 17 confirmed failures including unstable_cache crash, dead code in edge runtime, and preview mode reading env vars instead of cookies.
Launch2026-02-22
We Verified Code from 4 AI Platforms. Average Score: 40/100
Bolt 42. Lovable 42. Replit 44. Claude 35. 21 bugs across 4 projects. 0 passed.
Research2026-02-22
Why AI Hallucination Is Mathematically Inevitable
Four independent proofs from four research groups. Every future model will hallucinate. This changes how you build.
Experiment2026-02-22
We Tried Training Models to Verify Themselves. It Made Them Worse.
120 curated pairs: 91.5%. 2,000 pairs: 77.4%. More data caused catastrophic collapse. Verification stays external.