Lineage

Lineage is version control for a self-improving AI agent. Every edit the agent makes to its own instructions is treated like a commit, a jury of Meta models acts like CI, blame works like git blame, rollback works like git revert, and a locked test set works like a release gate. In practice it is an issue-triage agent that learns a real project’s conventions from the corrections people make, and it never keeps anything it learns until that change has passed on data it never trained on. I built it because teams want agents that learn from their corrections, but nobody ships one that rewrites its own instructions unsupervised, since it can learn the wrong lesson from a single bad batch and afterwards nobody can say why it made a call. Lineage demonstrates on 1,269 real kubernetes issues, using the project’s own kind and sig labels as ground truth. Across two full runs with seven lineages each, the gate rejected every edit learned from a deliberately relabeled batch of bugs while every unsupervised agent absorbed it. One unsupervised agent fell from 66 percent to 55 percent on the locked test the generation it learned the bad rule and never recovered, which Lineage found with a git-bisect style tool that scores every historical playbook. Agents that edited themselves without the gate ended up 41 percent different from each other on identical data, compared with 23 percent for gated agents and a measured noise floor of 4 percent. And no agent, gated or not, beat the frozen model on data it never saw, which the harness reported honestly instead of hiding behind rising batch curves. A positive control run with a learnable convention came back inconclusive and is reported as such. The dashboard has a triage view a team would actually use, curves, divergence, every self-edit as a diff with the jury’s verdict, a human review queue, blame with rule accountability, rollback, bisect, and a live cost meter. The harness also writes a one-page audit report for every run. Live dashboard on the real run at https://romaisafatima1.github.io/lineage/dashboard/?run=../runs/full2/events.jsonl