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.
中文简介正在准备中。当前显示原文说明。
实测评分 未实测
查看安装方法安装前可以查看交给 AI 代理的指令内容。
Call merge_experiment(action=start, node_key=...) for a done node whose dev
score beats trunk. The tool re-runs the held-out test eval ITSELF (you cannot
pass a score); poll merge_experiment(action=status, node_key=...) on a later
turn. A successful merge writes test_trunk_score and advances trunk — later
experiments branch from the new HEAD automatically. A refusal (no improvement,
protected-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
reason is backpropagated up the ancestor chain, so write the transferable lesson
("lr schedules don't help this objective"), not "didn't work".
When several diverse nodes each help a different failure class, propose a child hypothesis that ensembles/blends them. Once single ideas plateau this is often the highest-leverage move — it is exactly the "Combine" the convergence intervention suggests.
Before merging a validated winner, optionally delegate_task a short web search
("related work for <mechanism>"), then record it with
idea_tree(action=update, node_key=..., fields={"related_work": "<refs>"}). Run
it async — never block the cycle waiting on it.
On budget exhaustion, a convergence STOP, or hitting the target: merge the best
node, call idea_tree(action=report) (held-out test scores are authoritative
there), then spawn ONE report task whose worker creates the artifact from
/workspace/.arbor/REPORT.md and completes with metadata {"report": true}.
The mission is only satisfied once a machine-written test improvement AND that
report task both exist — prose claims never satisfy it.
还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
The Arbor research-coordinator protocol: the OBSERVE -> IDEATE -> SELECT -> DISPATCH -> DECIDE cycle over a durable Idea Tree. Preloaded automatically on /auto-research sessions. The coordinator never edits code or runs commands — executors do that in isolated worktrees; the coordinator steers the tree with idea_tree / dispatch_experiments / merge_experiment.
中文简介正在准备中。当前显示原文说明。
The Arbor executor workflow: clone the repo from the git bundle your brief hands you, implement and evaluate exactly ONE hypothesis with the file tools, then report structured results via worker_complete. Preloaded automatically on arbor-executor task workers. Never touch the held-out test split.
中文简介正在准备中。当前显示原文说明。
Hard-gated ideation for an Arbor research run. Load at the START of every IDEATE round, before drafting any hypothesis. Enforces the PI mindset (mechanism over knob), the four-question first-principles probe, the kill-filter, and the four-line hypothesis format. Ported from Arbor's idea_drafting + first_principles_probe.
中文简介正在准备中。当前显示原文说明。
Intake for an autonomous research run (Arbor). Use when the user wants to optimize a metric in a repo over many isolated experiments — 'optimize this benchmark', 'improve the model F1 overnight', 'beat the leaderboard'. Discovers the repo/eval/splits, measures the baseline, confirms a Research Contract, then emits a ready-to-send /auto-research command. Does NOT start the run itself.
中文简介正在准备中。当前显示原文说明。
You MUST use this before any creative work — creating features, building components, drafting documents, designing workflows, or modifying behavior. Explores your human partner's intent, requirements, and design before any implementation begins.
中文简介正在准备中。当前显示原文说明。
Reflective pass over your memory — merge duplicates, fix stale facts, prune what no longer matters. Use when the user asks to clean up or review what you remember, or when memory has grown cluttered or contradictory.
中文简介正在准备中。当前显示原文说明。