ControlSpec
The agent action lab / 01

Same intent.
Different outcome.

Watch an agent try to spend, discount, and change the deal. See where ControlSpec changes the route—and where it lets the action proceed.

Recorded sandbox simulation. Real ControlSpec reference decisions. Scripted agent intents. Fake targets. No real money, messages, or production authorization.
Personal agent

01 · Intent02 · Decision03 · Target04 · Evidence
Without ControlSpec

Direct dispatch to the simulated target

With ControlSpec

Reference evaluator + sandbox dispatch gate

6 scenes · approximately 60 seconds

The same scripted proposal enters both lanes. Only the sandbox control path changes.
Inspect the rule binding, decision, and sandbox trace

This browser replays recorded results; it does not evaluate new policies or run an LLM. The local Python harness runs the real reference evaluator and simulates target changes. A sandbox event proves only that simulation occurred. Production identity, credential isolation, enforcement, and durable evidence remain separate work. The baseline isolates ControlSpec’s contribution; it is not a claim that other agent systems lack safeguards.