name: arbor-merge-discipline\ndescription: \"DECIDE-phase doctrine for an Arbor research run: when to merge, prune, combine, or finalize; how the held-out merge gate works; and the rule that the final report uses TEST scores. Load before deciding what to do with completed experiments.\"\nversion: 1.0.0\nlicense: MIT\ntags: [research, decide, merge, arbor]\nrequires_agents:
\nCall merge_experiment(action=start, node_key=...) for a done node whose dev\nscore beats trunk. The tool re-runs the held-out test eval ITSELF (you cannot\npass a score); poll merge_experiment(action=status, node_key=...) on a later\nturn. A successful merge writes test_trunk_score and advances trunk — later\nexperiments branch from the new HEAD automatically. A refusal (no improvement,\nprotected-path hit, conflict) is tree evidence, not an error to retry blindly.
idea_tree(action=prune, node_key=..., reason=<the lesson>) for dead ends. The\nreason is backpropagated up the ancestor chain, so write the transferable lesson\n(\"lr schedules don't help this objective\"), not \"didn't work\".
When several diverse nodes each help a different failure class, propose a child\nhypothesis that ensembles/blends them. Once single ideas plateau this is often\nthe highest-leverage move — it is exactly the \"Combine\" the convergence\nintervention suggests.
\nBefore merging a validated winner, optionally delegate_task a short web search\n(\"related work for
On budget exhaustion, a convergence STOP, or hitting the target: merge the best\nnode, call idea_tree(action=report) (held-out test scores are authoritative\nthere), then spawn ONE report task whose worker creates the artifact from\n/workspace/.arbor/REPORT.md and completes with metadata {\"report\": true}.\nThe mission is only satisfied once a machine-written test improvement AND that\nreport task both exist — prose claims never satisfy it.