Skip to content
KIRRA
Mission Console
OP

Doer Bound

KIRRA bounds any planner behind the seam · real #131-checker verdicts

Simulated dataintent: GoTo(40, 0)checker: fail-closed

The invariant bound

same Mick intent · same world · two doers

The planner is a swappable doer behind one generic seam. Mick proposes the intent GoTo(40, 0); the world (corridor, obstacle, envelope) comes from perception. Whatever the doer proposes, KIRRA is the bound: the careful planner (Occy) is admitted, and a reckless / learned stand-in that drives straight through the obstacle is rejected — then the always-available safe-stop takes over. No unsafe trajectory reaches the actuator, whoever authored it.

Blocked road

stopped car at x=25 m · ego at x=5 · goal at x=40

occy · admittedreckless · rejected
goalcarego

Occy — reference planner

careful, safety-aware

Accept
reach
14.6m
vs obstacle
stops short

Stops short of the car well before reaching it — KIRRA admits the plan.

RecklessDoer — black-box stand-in

ignores obstacle / corridor / lane

MRCFallback
reach
28.9m
vs obstacle
drives through

Drives straight through the car. KIRRA rejects it — the same bound that admitted Occy.

On rejection → safe-stop

the always-available fallback

Accept

When a proposal is vetoed, the architecture falls back to a controlled safe-stop the checker accepts. The unsafe intent never actuates.

Clear road — the bound is precise

the very same RecklessDoer, no obstacle

Accept
goalego

With nothing to hit, the reckless doer is admitted — KIRRA rejects unsafe outputs, not the doer itself. That precision is what makes “swap in a learned planner, the safety case is unchanged” a real claim rather than blanket paranoia.