Does Governance Slow AI Down?
The short answer
No. In our tests, a governance decision takes between a few hundredths of a millisecond and a few milliseconds. The AI model it governs typically spends hundreds of milliseconds to several seconds deciding what to propose. Checking that proposal against every applicable authority adds roughly 0.1% to 1% to the time already spent on the action.
A common assumption is that layering authorities (federal, state, institutional, local), each with its own veto, must create a bottleneck. We measured it. The numbers below are what we found, and the last section explains what they don't show.
What we measured
We ran each test on our reference implementation of the decision engine. We built it from the same written specification as Cornerstone, and the two agree on every outcome, reason, and deciding authority across our published test scenarios. The heaviest case ran 20,000 times, and we report the slowest run, not only the typical one.
Question
Configuration
Result
Does a bigger rulebook make decisions slower?
10 rules up to 10,000 rules
No. About 0.022 ms at every size
What does each added authority cost?
Authorities added one at a time
About 0.012 ms per authority, growing in a straight line
A typical layered decision
5 authorities
About 0.06 ms
The heaviest case we test
64 authorities, every check sending the action up for review
Typical 1.3 ms; 99 in 100 under 2.7 ms; slowest of 20,000 runs 4.9 ms
All 20,000 runs returned the same decision, reason, and binding authority.
Compared with what it governs
The governance decision is by far the smallest step in the chain.
Step
Typical time
Governance decision, 5 authorities
about 0.06 ms
Governance decision, heaviest case tested
about 1–5 ms
A person's fastest reaction to a signal
about 200–250 ms
An AI model deciding what action to propose
hundreds of ms to several seconds
In the time it takes a person to react once, the engine can make the five-authority decision about 4,000 times. Speed isn’t why it exists, though. The point is that every authority applies its limits the same way, every time, and records why.
Why these numbers matter
Adding rules costs no time. An institution can encode every applicable regulation and policy without slowing its systems. The size of the rulebook did not move the clock.
The cost is known before deployment. Decision time depends on how many authorities are involved, not on what the AI proposes. A buyer can work out the worst case in advance instead of discovering it in production.
The same request gets the same answer. Given identical inputs and the same prior state, the decision is identical. No AI model is involved, so you can re-run and check a past decision.
What these numbers do not show
Desktop measurements, not production hardware. We measured these on a desktop computer running the current implementation. Production figures will differ, and we will publish them when we have them.
Engine time only. The figures exclude network travel between the AI system and the engine, and time spent fetching live facts such as a user's role. In a real deployment, those steps usually take longer than the decision itself.
Not a hard real-time guarantee. The slowest run we observed is a measurement, not a proven upper limit. Machinery that must decide within a fixed deadline hundreds of times per second needs further engineering and testing on its own hardware. We do not claim that today.
Human review takes human time. When an action is sent to a person for approval, the wait depends on how long that person takes. The figures above cover only automated decisions.