rlat-learn

To capture durable world facts into a knowledge model, this Skill mines user statements for durable facts, standing constraints, and tried-and-falsified findings, then filters them through four validation gates before prompting user confirmation to store the approved items.

Extracting standing constraints in-session without storing personal data requires privacy safeguards that scope learning to world-relevant statements only. This Skill applies a world-only scope gate to prevent personal data leakage while identifying durable constraints.

The land-and-confirm loop for adding facts to an rlat file works by prompting user confirmation before landing each candidate fact. Approved items are committed to the project-scale knowledge model via rlat capture-attribute.

The four gates used to filter candidate facts before landing them enforce durability, user-statement provenance, premise relevance, and world-only scope. This framework ensures only verified, world-relevant statements are captured in the knowledge model.

You do not need an API key to teach the rlat model its world during a session. The workflow operates entirely key-free in-session, relying on user confirmation rather than external authentication to populate the knowledge model.

You can use this approach to store tried-and-falsified findings for later analysis by identifying them from user input. The Skill applies its four gates to these findings and lands them into the project-scale knowledge model after explicit user confirmation.