# YETKI — Known Limitations

## Controlled deployment environment

The external staging/production deployment target is a stateful simulator. It is intentionally used to provide deterministic hidden-world ground truth for ambiguous side-effect experiments. It does not reproduce every property of a real cloud or CI/CD provider.

## Authority calibration

EBAC v0.2 is a transparent, versioned engineering policy. Its weights and thresholds are not claimed to be universally optimal and the conservative lower bound is not a probability that an agent is “safe”.

## Certification coverage

Authority is only as strong as the scenarios and evidence types used to produce it. Unseen failure modes remain possible.

## Rollback evidence scope

The current `ROLLBACK_VERIFICATION` certification evidence verifies code-artifact backup/restore integrity in the Sandbox. The Action Proof separately requires a declared and verified deployment rollback plan. A future real deployment adapter should verify rollback against the actual target provider.

## Domain scope

The current prototype focuses on autonomous software-engineering workflows. Generalization to other agent domains has not been established.

## Blind-world proof scope

The current leak test provides structural evidence that hidden-world labels and expected decisions are absent from the runtime implementation and timeout payload. It is not a formal information-flow proof.

## Safety

YETKI does not guarantee autonomous-agent safety and does not eliminate production incidents. It measures, constrains and progressively expands operational autonomy based on observed evidence and explicit policies.