The same tool call, the same declared values, ten times. Nothing varies between runs.
Every verdict is shown: none filtered, none retried.
press R to run
biject — Lean 4 kernel proof
Same declared values → same proposition → same verdict.
No run yet.
LLM-as-Judge (representative architecture)
Reasoning-based verdict over full context, re-evaluated per call — the pattern shared by most current agentic guardrail systems.
No run yet.
Fifty calls, with state accumulating
One event enters the rolling window per call. Both systems answer the same question each time:
is this action still within the declared ceiling? Scored against what the compiled policy decides.
press R to run
time to verdict
LLM-as-Judge—
biject—
agreement with the compiled policy
No run yet.
recent calls
—
The same error rate, at production volume
Act 2's measured rate at a call volume you choose. A false ALLOW is an action the compiled policy
refuses and the reviewer permitted: an authorization your audit trail cannot account for.
LLM-as-Judge (representative architecture)
Reasoning-based verdict over full context, re-evaluated per call — the pattern shared by most current agentic guardrail systems.
—
false ALLOWs per day
biject — Lean 4 kernel proof
Decided, not estimated. Same declared values, same verdict, at any volume.
0
false ALLOWs per day, on this policy
0 × any volume = 0.
Scope: this is the false-ALLOW count for this compiled predicate over the declared parameter values it
was given. It is not a claim about policies that were never written, values that were never declared, or
conduct outside the tool call.