Vortex Research

Field Notes

Short write-ups of experiments we run on AI models, including the ones that came back negative, newest first.

• Field Notes · 2026-09

  Does the most capable model catch the most?

  Three models premortem the same database spec, on the same commit, with the same prompt. The two findings that mattered came from one of them, and it wasn't the most capable.

  Read the note →
• Field Notes · 2026-09

  Which model actually premortems properly?

  Four AI models premortemed the same engineering spec. 17 of 27 problems were found by only one of them. Then our own citation check got it wrong, and we nearly published the mistake.

  Read the note →
• Field Notes · 2026-07

  Does "be careful" make a model think harder?

  Applied-AI research from Vortex Computation. We ran 140 blind agents across two models to test whether framing a task as "careful" vs. "casual" changes how deeply a model reasons, and published the null result.

  Read the note →

Vortex Computation™ · Applied-AI research · vortex.llc/research