AUTONOMY NEEDS PROOF
Will your agent follow the rules?
Turn written policies into rules you can review. Test your agent against them before it acts.
Agent, on its own
“Refund
approved.”
$4,986 booking · simulated airline test
Compiled policy
Not eligible.
No cancellation condition met.
WATCH THE PRODUCT
Truthbound in under
two minutes.
A narrated walkthrough built from saved runs: compile a policy from several documents, review each rule, then test and check an agent. Simulated and sandbox tests are labelled on screen.
1:34 · captions on · Download MP4
WHY IT MATTERS
More autonomy.
More at stake.
Agents can issue refunds, approve requests and make promises. Test whether they follow your policy before you give them that responsibility.
From policy to action.
- 01Review the rulesTurn your policy into rules, even when it spans several documents. Have the policy owner check them.
- 02Test your agentGenerate a case for each outcome in the rules. See where the agent disagrees.
- 03Check each actionIn our demo, the rules block forbidden calls before they take effect.
SEE THE DEMO
A refund the agent
should refuse.
In a simulated airline test, an agent tried to refund an ineligible $4,986 booking. Truthbound’s policy check blocked the call.
Open the demo- Simulated refund call blocked
- $4,986
- Agent alone → with the check
- 4/7 → 7/7
Seven cases, one airline policy. Scores show final agreement with compiled rules.
About these results
GPT-4.1 nano matched the rules in 4/7 cases across three unguarded runs and finished at 7/7 in five guarded runs. Its first decisions were 5/7 to 6/7 before intervention. No forced overrides were needed. This is a prototype demonstration; a policy owner still needs to verify the rules. See methods and all attempts ↗
NEW · MULTI-DOCUMENT POLICIES
One rule set from
several documents.
Real policies are split across pages and tables. Truthbound compiles them together, cites the source lines, and stops when a required document is missing or two equally authoritative documents conflict.
- Policy documents in one rule set
- Up to 12
- Missing or conflicting source
- Stops
Tested on exact excerpts from two public Delta pages plus our scope notes, and on synthetic policies for the stopping cases.
About this test
We compiled Delta’s cancellation page and fee table, plus our scope notes, into one rule set. It matched all 10 test cases written before the run on eligibility, and 7 of 10 on the full answer, because unknown fees were set to zero. After two fixes it matched 10 of 10; that is a regression test, not a holdout. Separate synthetic tests with a missing required document, and with two equally authoritative documents in conflict, both stopped compilation. Claude Fable 5.1 compiled these policies. The feature is in our pipeline, off by default; this website shows saved results and does not run it.
THE RESEARCH BEHIND TRUTHBOUND
Built on
KnownLieBench.
KnownLieBench studies when agents misreport their actions under conflicting incentives, using 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?
We’re looking for design partners building agents for refunds, fees, credits or approvals. Bring a policy and a workflow to test.