RELIABLE is four primitives that stop AI coding agents from lying, spiraling, grading their own homework, and losing hard-won judgment. Extracted from a live AI operating system that runs a real business. Yours, free.
One zip. Five print-ready docs, two working Python scripts, four copy-paste templates. No signup wall.
Your agent reports "done" when nothing ran. This makes "done" a fact the world proves, not a claim the model makes.
A job logs "sent" while nothing arrived. Every recurring task gets a check that reads the outcome in the world, never its own log.
Builds spiral ten cycles deep with no "done." A four-line brief, thinnest-slice-first, and six guardrails that keep you one step from a working commit.
Hard-won lessons vanish every session. Turn judgment into a versioned asset a cheaper model can follow. Errors worth more than wins.
If it saves you one ten-cycle spiral, that's the whole point. The two load-bearing ones are the Proof-Gate and the Process-Check. Install those two and your agent's "success" stops being a sentence it generates and becomes something the filesystem confirms.
No email, no account. Built by one person with a full-time job and an AI that does the work. Built in public.
Download RELIABLE ↓