Agents you can let run
What they did is visible afterwards, and the irreversible parts are gated.
File 07 · converge
An agent saying “I followed the rules” proves nothing.
Compliance you can trust comes from structure, not instruction. Separate the maker from the checker, and make every check leave something you can look at.
An agent saying “I followed the rules” proves nothing. That's performative compliance.
Compliance you can trust comes from structure: what the agent can see, what it's allowed to do, what it must produce, and which contexts are kept apart.
Separate maker and checker. Only the work crosses the wall. The checker stays blind to the maker's reasoning.
Not a claim that it was checked: the check itself, something you can look at, on file.
Always allow “nothing found.” A clean result goes on file the same as a kill.
On anything high-stakes, the checker comes from a different model family. Same family, same blind spots: the check passes what it should catch.
Above a stakes threshold
Move the stakes:
CHEAP AND REVERSIBLE.A full outside check can cost more than the mistake would.
Before the act
A perfect receipt can document a leak after it happened. Irreversible actions get a gate before the act, not just a receipt after it.
Can the act be undone?
CAUGHT AFTER, IT CAN STILL BE WALKED BACK.The receipt after the act does the job.
Honest numbers
A probability needs a reference class, a settle condition and an evidence tier, or it's decoration.
Evidence quality caps how sure you may be.
Evidence tier
NO DENOMINATOR, NO NUMBER.Without its three parts, a probability is decoration.
What it allows