LabFactory gateLab notebook
Merge Graph
Lab notebook · pinned model jev-1.13.0
Hypothesis
If merge policy is a known graph of legal steps, a small model can pick the next edge without freelancing as a reviewer — and when the honest next step is a human, it should stop.
Method
- Paste a public GitHub PR (
owner/repo/pull/N) or run the canned demo. - The full step catalog is on the canvas. Faint lines are legal edges. Bold is the path this run took.
- Each tick, Jev chooses only among edges that are legal right now.
- Receipts (latency, estimated $, latest choice) stay on a slim rail — the tree is the experiment.
Lab notebook · Merge Graph
Decision tree — legal edges as context
Taken path highlighted · unused branches dim · not a linear strip
Run the canned demo, or paste a public PR URL. The tree waits with the full catalog dimmed until the first tick.
Observation
On Brickato #1175 (public card), one measured walk:
check_ci → probe_tests → pick_next_action(ask_tests) → ask_human
4 ticks · ~1.74s · ~$0.0002 est. Live runs print their own receipts above. Terminal ask_human is a success of the policy, not a broken demo.
Why this shape
You’re watching a factory gate: policy over a known graph. The unused branches stay visible so you can see what could have happened — not a magic “AI reviewed your PR” strip.
Lab footnotes
Schema ≠ truth.
High confidence is still a choice inside the map. It does not prove the CI signal, test probe, or blast-radius score is ground truth. Open the checks yourself before you trust a verdict.
What this is / isn’t.
Is: playable architecture demo; legal edges in code; HITL ask_human; public PRs only; measured path. Isn’t: Doom or flights Lab; CV screener; autopilot merge bot (no comments, no merge); “correct code review” claim; private repos.
Model: jev-1.13.0
Built with care in an agentic studio. Daniel approves what ships.