Concordat turns an assistant’s approval rules into executable state transitions. Milad Sikaroudi’s Python library offers only the actions permitted at the current conversational stage. Illegal calls are rejected before they change anything.
The library also reserves human decisions for an application-supplied approver and subtracts delegated spending allowances from the parent’s budget. Its experiment tests a narrower airline confirmation adapter.
- Simulated customers resist the first confirmation request, then agree if asked again. The study uses fifty airline tasks.
- With GPT-4.1, the standard assistant makes unconfirmed changes in thirteen of fifty tasks; Concordat records zero. Task passes change from twenty to 22.
- Across 128 paired Haiku task-runs, violations fall from 29 to zero, while task passes change from 112 to 105. Neither success-rate difference is statistically significant.
- In the first Haiku pressure run, eight of thirteen violating tasks still receive full task reward. A correct final booking can conceal skipped consent.
The gate and violation metric use the same keyword rule. Zero violations follows from that definition, which can accept ambiguous consent such as “yes, but not that one.” A deployment needs structured approval tied to the actual change.
These are author-run tests with simulated customers. The prompt changed between models, so their different task-success patterns cannot isolate the instruction change’s effect. The library and experiment code are public.