how it worksfor the curious
Models read. Rules decide.
Holt keeps only the evidence it can trace to a real GitHub page, then runs the same written rules on every repo.
Six steps. One of them decides.
AclassifyWhat kind of repo is this?model
BopportunityIs there a way in for outside work?model
CoutcomesWhat happened to outsiders who tried?model
DverifyOpen every citation. Drop whatever doesn't check out.no model
→verdict.pyWritten rules pick the verdict from what's leftno model
EnarrateExplain an answer it cannot changemodel
Typical input: 642 pieces of evidence from 200 pull request threads.
Don't take our word for it.
No verdict is right every time, so the evidence stays on the page. Open any citation and check it against the GitHub thread yourself.
- 55 / 55same verdict on all three runs
- read-onlynever writes to GitHub
- every claimlinks to a real GitHub page
That 55/55 is from Holt's competition days. Treat it as history, not a promise. [ reproduce it → ] [ full evaluation ]
Prefer the terminal? Same engine.
Same rules, on your machine, with your own GitHub token. Good for scripts, and for people who never leave the terminal.
uv tool install holt-cliholt / terminal app● read-only
