Research
What we're actually working on — measurements of real deployed code, the tools built to do it, and honest write-ups of what was found. Public code only; proof over claims.
Research isn't only measuring other people's code. The Digital Integrity Institute's v0.1 rubric scores whether an AI-produced record is actually defensible — independently authored, not something we wrote ourselves. The day it went live, we ran our own verification methodology against it and posted the real number publicly, gaps included. As far as the rubric's own author has confirmed, that made us the first public, reproducible self-score against the standard.
Running it once wasn't the end of it. Pushing the build past our own score surfaced a real bug — not in the thing we were verifying, but in the verification mechanism itself: a timestamp-format mismatch that let a hash check silently pass when it shouldn't have. That failure class — the proof apparatus having a defect the thing it's proving doesn't — is being folded into the rubric's v0.2 draft as its own criterion. Contributing the gap, not just passing the test, is the part we'd rather be known for.
Current self-score: 16/20 · floor criteria (the ones that decide whether a record survives a real challenge): 6/8 — every point earned, nothing rounded up.
See the full breakdown, criterion by criterion →