AUTONOMY NEEDS PROOF
Test policy-governed agents before release.
Truthbound compiles a written policy into a reviewable check, generates a test case for each outcome the check distinguishes, and checks the agent's consequential tool calls against the approved rules while it works.
In our demo, a low-cost support agent issued a cancellation-and-refund call on a $4,986 simulated reservation that the policy says is not eligible. The same check blocked the call.
Agent, on its own
“Refund
approved.”
cancel_reservation call issued · simulated airline backend · $4,986 reservation · economy, no insurance, booked a month ago
Compiled policy
Not eligible.
Unmet: not within 24 h · airline did not cancel · not business class · no insurance
WHY IT MATTERS
More autonomy.
More at stake.
An agent's decision can trigger a refund, a release or a promise to a customer. A written policy only protects you if the agent follows it, case after case, and you can show which rule applied and why.
What Truthbound does, before
and after an agent release.
- 01Compile the policy for reviewPolicy text becomes an executable check over a finite set of conditions, validated for syntax, totality and rule conflicts. A policy owner still reviews it for fidelity to the document.
- 02Test each outcome it distinguishesEnumerate the combinations the compiled check can tell apart, generate one realistic case per outcome, grade the agent against the check.
- 03Check actions at runtimeIn the demo integration, refund and transfer calls and the final answer are checked against the compiled rules before they take effect.
SEE IT ON A REAL POLICY
A $4,986 refund call,
caught twice.
We compiled an airline cancellation policy, generated a case for each of the seven outcomes the compiled check distinguishes, and ran a low-cost support agent with real tools against a simulated airline backend. It issued a refund call the policy forbids. Testing each outcome found it; the runtime check blocked it. With the check on, the same cheap model finished 7/7 on those seven cases. GPT-5.5 and Claude Opus 5.5 each scored 7/7 unguarded in one run, at roughly fifty times the inference price.
Open the demo- Simulated refund call blocked at runtime
- $4,986
- Same low-cost model: unguarded → final outcomes with enforcement
- 4/7 → 7/7
- Inference list price per seven cases vs GPT-5.5 or Claude Opus 5.5
- ≈2%
- Combinations of six inferred conditions → distinct outcomes
- 240 → 7
GPT-4.1 nano: three unguarded runs; five guarded runs, with first decisions at 5/7 to 6/7 before interventions and zero forced overrides. Frontier models ran once each. The 2% excludes the $0.14 compile-and-case-generation run and all operating costs. One policy was developed against; three more were attempted, with mixed results, and every attempt is listed in the demo. Not a benchmark.
THE RESEARCH BEHIND TRUTHBOUND
Built on
KnownLieBench.
Policy compliance is the first release test we built. The research behind it, KnownLieBench, studies when agents misreport what they did under conflicting incentives, in controlled customer-service simulations.
- Models evaluated
- 18
- Grounded cases
- 112
- Multi-turn interactions
- 18,144
LET'S BUILD TRUST IN AGENTS
Testing an agent that
must follow written policy?
Research stage. Looking for two or three design partners: a support or operations agent that makes a policy-governed, high-value tool call (refunds, credits, approvals), a policy owner who can review compiled rules, and a scoped pilot.