sven-mirana/sublation

Sublation is a governance framework for managing skills used by AI agents. The name comes from a philosophical concept (Aufhebung in German) that means preserving something while also changing it. Applied here, it describes a process where agent skills are improved over time based on observed failures, but never by having the agent directly edit its own live skills. The core problem it addresses is control: if an AI agent can freely modify its own instructions or capabilities, there is no audit trail and no human checkpoint. Sublation solves this by treating all active skills as read-only. When an agent encounters a failure during a skill execution, it files an observation describing what went wrong. From that observation, a candidate is created, which is a sandboxed proposed change. The candidate goes through a 22-point audit covering correctness, safety, and consistency checks. It then requires cross-agent review and user approval before being promoted to replace the active skill. After promotion, there is a safety window where the change is monitored in production. If something goes wrong, a rollback mechanism is available. The framework tracks provenance and license information for any content that gets incorporated into skills. The framework supports four types of changes: patches to the skill specification, enhancements to scripts, infrastructure fixes, and tooling improvements. It also handles merging multiple skills together, which the README notes was used to reduce 32 skills down to 11 in one consolidation effort. The project is described as being in maintenance mode at version 1.0, meaning the framework is considered complete and future changes will only come from real usage revealing problems, not from adding features speculatively. It is built in Python and uses command-line scripts to drive each stage of the pipeline.