May 4, 2026

A Self-Learning Loop That Cannot Approve Itself

Datarim turns repeated lessons into proposals and backlog work, while a human remains the authority for changing framework rules.

“The agent learned from the task” is usually impossible to audit. Datarim uses a narrower definition. A completed task may produce a concrete lesson; that lesson may become a proposal; an approved proposal may become a framework change or a backlog task.

The important word is may. Reflection does not grant an agent permission to rewrite its own rules.

Reflection produces evidence

At the end of a task, the workflow records what helped, what failed, and whether the same problem is likely to recur. A useful lesson names the failure shape and the smallest reusable response. “Be more careful” is not a reusable lesson; “a shell wrapper must prove that a failing child exits non-zero” is.

A proposal is not a rule

Small content corrections follow a normal review path. Changes to authority, routing, sources of truth, or safety boundaries require a design artifact before approval. A person can accept, reject, or reshape either class. Accepted follow-ups enter the backlog with their own scope and tests.

This separation prevents a single strange task from teaching the framework a universal rule. It also leaves a trail: reflection, decision, implementation, test, and evolution log are distinct artefacts.

Learning becomes a regression gate

The loop is complete when a recurring incident turns into a test that fails under the old behavior. Until then, the framework has recorded an opinion, not learned a capability.

Read the current Evolution skill reference for the approval classes and rollback rules.

Published retroactively from the framework archive. The reflection-to-backlog loop was documented on May 4, 2026.