Structural Admission Released to Verify Dependency Structure Before Training

When agents train in multi-step environments, it is crucial to verify whether the observed learning curves or 'emergence' phenomena are actually due to the environment's causal structure. Structural Admission is a Python-based standard library tool that, before training begins, verifies whether the dependency structure of a designed task matches the actual observation and action interfaces.

This tool uses task adapters and script-based Oracles designed by researchers to perform the following rigorous verification process:

• Data Integrity: Separation of calibration seeds and task execution seeds, CI measurement restricted to pre-registered phases.
• Verification Logic: CMI (Conditional Mutual Information) threshold-based verification, evaluation under uniform random and scripted oracle policies.
• Leakage Prevention: Pre-disclosure leakage checks on declared observation field groups.
• Reproducibility: Content-hashed reports and byte-level reproduction of deterministic artifacts.

Verification results are reported as Admitted, Rejected, or Inconclusive. In an actual case, a specific relationship was expected to be non-operative, but the CMI measurement exceeded the set threshold, leading to the rejection of the task before proceeding with the training experiment, thereby identifying a structural flaw.