Arbor

Arbor — Autonomous Optimization via Hypothesis Tree Refinement

Overview

This skill runs an Autonomous Optimization (AO) loop: starting from an existing artifact and a measurable objective, improve it through many rounds of experiment and evaluation — without step-by-step human supervision and without overfitting to the feedback signal. It's the right tool when the bottleneck isn't writing one good change, but organizing dozens of trials so that lessons accumulate instead of evaporating.

It implements Hypothesis Tree Refinement (HTR) from Arbor (Jin et al., 2026). The key idea: keep the research state in a persistent hypothesis tree rather than in conversation history. Each node binds a hypothesis, the distilled insight it produced, and a pointer to the artifact version that realizes it. You play the long-lived coordinator that owns this tree and decides where to search; short-lived executor subagents test one hypothesis each in isolated git worktrees and report back. A held-out merge gate admits a change only when it improves on a test evaluator the search never optimized against. This is what turns trial-and-error into cumulative, auditable research.

Use the scripts/tree.py state manager for all the bookkeeping (creating nodes, writing evidence, propagating insights, pruning, the merge gate, the Observe projection). It keeps the state consistent and frees you to spend judgment on what the evidence means.

When to use this skill

Reach for Arbor when the task is iterative improvement of a concrete artifact under an evaluator:

• Model training: optimizer/architecture/recipe changes to lower loss or hit a target in fewer steps.
• Harness/agent engineering: raising pass rate or accuracy of an agent loop, search harness, or tool-use scaffold.
• Data synthesis: improving a generation/filtering pipeline judged by downstream model behavior.
• Benchmark optimization: MLE-bench / Kaggle-style "improve the submission" tasks.
• Prompt/system optimization where you can score outputs automatically.

The distinguishing signals: there's an artifact you can modify, an objective, a way to score candidates, and you expect to run many experiments. If the user only wants a single fix or a one-shot answer, this is overkill — just do the work directly. If they want open-ended ideation with no evaluator, use hypothesis-generation or scientific-brainstorming instead.

The AO setup — pin this down first

Before any experiments, establish the task tuple (M_0, O, E_dev, E_test). Getting this right matters more than any later decision, so confirm it explicitly:

• M_0 — initial material: the artifact to improve (a repo, a script, a config, a prompt). Make sure it's under git and currently runs.
• O — objective: the natural-language goal and the metric direction (maximize accuracy? minimize loss/steps?).
• E_dev — development evaluator: a command you can run freely during search to score a candidate. Fast, repeatable.
• E_test — held-out test evaluator: a separate evaluator (different seeds, different split, or a larger run) used only at the merge gate. It must not be used as a search oracle — that's the whole point.

If the user hasn't given you a clean dev/test split, construct one and say so. The dev/test separation is the mechanism that catches overfitting: a candidate that wins on dev but not on test isn't a success, it's a warning that you're exploiting the feedback signal. Without it, autonomous search reliably overfits.

Initialize the run:

python scripts/tree.py init \
  --objective "Improve BrowseComp answer accuracy on the search harness" \
  --dev-eval "python eval.py --split dev --n 50" \
  --test-eval "python eval.py --split test --n 300" \
  --material "." --metric-direction max --branching 3 --max-depth 2 --budget 12

--branching is how many sibling hypotheses you propose per parent; --max-depth 2 keeps directions at depth 1 and concrete interventions at depth 2 (the paper's default); --budget is the number of coordinator cycles. Start small (10–20 cycles) — structured search beats brute force, and you can extend if progress is still being made.

The coordinator loop

You run repeated cycles of six steps. This is the heart of HTR; do not collapse it into ad-hoc editing. Run python scripts/tree.py cycle once per cycle to track the budget.

1. Observe

Begin every cycle by re-grounding in the tree, not in your memory of the conversation:

python scripts/tree.py observe

This prints the objective, global insights, the active frontier (selectable hypotheses), executed nodes with their evidence, pruned lessons (negative constraints), and the current best artifact. Treating the tree as the source of truth is what keeps you coherent over a long run, after context compression has thrown away the details.

2. Ideate

Pick a promising parent and propose a few child hypotheses under it. Condition on the tree's evidence — this is the difference between Arbor and random search:

• Validated insights are assumptions you can build on.
• Pruned nodes are dead ends to avoid.
• A "half-right" result is a starting point for a sharper hypothesis, not a reason to abandon the direction.

Each hypothesis should be a falsifiable claim about how changing the artifact will move the metric, not a vague intention. Depth-1 nodes are broad directions ("the search harness loses correct answers it already retrieved"); depth-2 nodes are concrete, executable interventions ("run K=5 independent rollouts and aggregate by evidence dossier instead of majority vote").

python scripts/tree.py add-node --parent n0 --hypothesis "Verification, not retrieval, is the bottleneck: candidates are found but discarded"
python scripts/tree.py add-node --parent n4 --hypothesis "Decompose the question into atomic constraints and verify each independently"

3. Select

Choose which pending leaves to run next. Selection is not pure score-maximization — pick a hypothesis because it has strong prior evidence, because it would resolve an ambiguity its siblings exposed, or because its failure would clarify an important assumption. Frontier control under delayed feedback rewards informative experiments, not just promising ones.

4. Dispatch

Run each selected hypothesis as an executor subagent in an isolated worktree (use the Agent tool with isolation: "worktree", or have the executor create one with git worktree add). Isolation matters: parallel experiments must not clobber each other or the current best, and exploratory changes stay quarantined until they pass the merge gate.

Dispatch siblings in parallel (multiple Agent calls in one message) when they're independent — comparative evidence within one direction is exactly what makes later pruning and abstraction possible.

Give each executor a tight, hypothesis-bound brief. See references/executor-brief.md for the full template. The contract that makes HTR work: the executor may not change the hypothesis when the metric stalls. It repairs its own code and reruns, but h_n is fixed — otherwise the returned score is no longer evidence about the assigned node and the tree's semantics break. The executor returns exactly four things:

• dev_score — the dev evaluator result (for selection);
• result — a factual summary of what happened;
• insight — the distilled, reusable lesson (why the result supports, weakens, or bounds the hypothesis);
• branch_ref — the git branch/commit/worktree path holding the artifact.

Mark a node running before dispatch (tree.py set-status --node n5 --status running) so the Observe projection stays accurate.

5. Backpropagate

When an executor returns, write its report into the node, then abstract the lesson upward:

python scripts/tree.py set-evidence --node n5 --dev-score 70.0 \
  --result "K=5 dossier aggregation recovers answers in minority rollouts" \
  --insight "Correct answers often appear in a minority of rollouts; aggregation beats majority vote" \
  --branch-ref "wt/n5"

python scripts/tree.py propagate --node n5 \
  --insight "Candidate coverage, not verification, limits this direction" --to-root

This is the step that makes the tree more than a log. A leaf-level observation ("data-interface mismatch") should become a direction-level constraint and, if it generalizes, a global prior that shapes future ideation. Insight propagation is the component that drives most of HTR's gains — in the paper's MLE-Bench Lite ablation, a tree without insight feedback scored even lower than a flat experiment queue with no tree at all (54.5% vs. 63.6% any-medal, against 81.8% for the full system). Hierarchy alone isn't enough: the semantic memory is what matters. So spend real thought on the abstraction; don't just copy the leaf insight upward verbatim.

6. Decide

Decide what to do with the new evidence: keep expanding a direction, prune a falsified subtree, or attempt to merge a candidate.

• Prune dead ends, recording why — the reason becomes a negative constraint:

  python scripts/tree.py prune --node n7 --reason "search-augmented judge overfits dev questions; no test transfer"

• Merge gate — promote a candidate to the new best only if it improves on `E_test`. Run the test evaluator in a fresh worktree (not the dev worktree, to avoid leakage), then:

  python scripts/tree.py merge --node n5 --test-score 67.67 --branch-ref "wt/n5"

If the gate rejects it, that's informative: a high-dev / low-test candidate is evidence the direction may be exploiting the dev signal rather than producing a transferable improvement. Record that lesson; don't quietly promote it anyway.

Repeat until the budget is spent, the frontier is exhausted, or progress has clearly stalled.

Finishing the run

When you stop, produce a short report (see references/report-template.md) covering:

• the final best artifact, its test score, and its delta over M_0;
• the tree (python scripts/tree.py status) as the audit trail of what was tried;
• the main hypothesis shifts — how task understanding deepened across the run (early nodes test broad mechanisms; later nodes find their limits; ancestor insights compress these into the constraints behind the final design);
• merged vs. explored: many nodes improve dev, far fewer pass the test gate — report that gap honestly rather than overstating dev wins.

Always leave M_best as a real, runnable artifact on a named branch, and tell the user how to check it out.

Principles that make this work (not rote rules)

These come from the paper's analysis; understanding why matters more than following them mechanically.

• The tree is the memory; conversation is not. Over a long horizon your context gets compressed. Re-Observe each cycle so decisions rest on durable evidence, not a lossy summary.
• Structured search, not more sampling. Arbor's gains come from how the budget is organized — maintaining competing hypotheses, comparing siblings, carrying lessons forward — not from spending more tokens. Don't fan out aimlessly; each experiment should be conditioned on what the tree already knows.
• Dev guides, test admits. Use dev feedback freely to steer exploration, but never let a dev win into the final artifact without test confirmation. The dev/test disagreement is itself a signal worth reading.
• Executors are hypothesis-bound. Local engineering flexibility (edit, debug, rerun) is fine; silently changing the hypothesis to chase a better number is not — it destroys the meaning of the evidence.
• Failures are constraints, not noise. A falsified hypothesis tells you what the solution must avoid. Pruned-with-a-reason is more valuable than pruned-and-forgotten.

Reference files

• references/htr-methodology.md — deeper explanation of HTR, the node structure, the six steps, and the paper's empirical lessons (ablations, transfer, cost). Read when you want the rationale behind a design choice.
• references/executor-brief.md — the template for the brief you hand each executor subagent.
• references/report-template.md — the final-report structure.
• references/arbor-upstream.md — how to install and run the standalone arbor CLI from RUC-NLPIR/Arbor instead of orchestrating it natively, and when to prefer each.

Running the standalone Arbor CLI (upstream tool)

This skill normally runs HTR natively — Claude is the coordinator and subagents are executors. That's the recommended path: no extra install, no separate API keys, and you stay in the loop to read evidence between cycles.

But the paper's authors also ship a full implementation as a CLI. Use it instead when the user explicitly wants to run the published system (e.g. to reproduce paper results), wants Arbor to run fully unattended for many hours via its own live dashboard, or wants its built-in report/web-UI tooling.

Source: https://github.com/RUC-NLPIR/Arbor

Install

Requires Python ≥ 3.10 and Git.

git clone https://github.com/RUC-NLPIR/Arbor.git
cd Arbor
python -m venv .venv && source .venv/bin/activate
pip install -e .
arbor doctor      # verify install, PATH, git, API keys

Configure provider/model/keys

arbor setup       # writes ~/.arbor/config.yaml (provider, model, base URL, keys)

Supported backends: Anthropic, OpenAI / OpenAI-compatible Responses API, and LiteLLM (DeepSeek, Gemini, Qwen, vLLM, Ollama, local gateways). Keys can also be set via environment variables.

Run

1. Prepare a benchmark directory: an initial artifact under a clean git repo plus an evaluation script (your E_dev / E_test). 2. Author a project research_config.yaml (task description, coordinator settings — max cycles, depth, merge thresholds — executor max turns, UI mode). See examples/research_config.example.yaml in the repo. 3. Start the interactive session:

   arbor

Arbor runs an intake conversation, forms a Research Contract, then a live dashboard takes over. Each experiment runs in an isolated git worktree; verified improvements merge into a per-run trunk. 4. Outputs land in .arbor/sessions/ with REPORT.md, the event log, and results. Re-render a past session's report with arbor report <session>.

Key CLI commands

Codebase map (for the curious / for debugging)

The implementation lives under src/ (a src-layout; the CLI installs as arbor). The package directories are:

• core/ — ReAct loop, tools, LLM providers, context management
• coordinator/ — coordinator agent, the tree, orchestrator, coordinator tools
• executor/ — executor agent and CLI
• cli/ — intake, live dashboard, setup, doctor, config
• events/ — typed event bus and payloads
• report/, webui/ — report generation and read-only run monitor
• search_agent/ — the minimal ReAct search harness (the M_0 for the

BrowseComp / search-agent tasks)

• plugins/ — domain plugins (e.g. mle_kaggle.yaml)
• skills/ — on-demand markdown playbooks

(top-level src/ also has dashboard.py, run.py, review.py.)

Naming note: the paper and this skill call the persistent state the hypothesis tree; the tool's code and dashboard call the same structure the Idea Tree. They are the same thing. The depth convention also matches the native skill: root/depth 0 = objective + global insights, depth 1 = research directions, depth 2+ = concrete tested methods.

Native vs. upstream — quick guide

• Native (this skill): best default. Lower setup, transparent, you read and

steer between cycles, reuses your existing Claude Code session and worktrees.

• Upstream CLI: choose for paper reproduction, long unattended runs with the

official dashboard, or when the user specifically asks for the arbor tool.

Executor brief template

Each executor is a short-lived subagent that tests one hypothesis in an isolated git worktree and returns structured evidence. Dispatch it with the Agent tool (use isolation: "worktree" so it gets its own copy of the repo, or instruct it to run git worktree add itself). Dispatch independent siblings in parallel — multiple Agent calls in one message.

Fill in the bracketed parts. Keep the brief tight: the executor needs the hypothesis, the context that lets it implement well, and a crisp contract for what to return — nothing more.

---

You are an Arbor executor. Test ONE hypothesis in an isolated git worktree and
return structured evidence. Do not change the hypothesis — your job is to give
the coordinator clean evidence about THIS claim, even if it turns out false.

HYPOTHESIS (h_n):
  [the falsifiable claim, e.g. "Aggregating K=5 independent rollouts by an
   evidence dossier recovers correct answers that majority vote discards."]

CURRENT BEST ARTIFACT (M_best):
  [path or git ref of the current best, e.g. branch `arbor/best` — start from this]

RELEVANT INSIGHTS FROM THE TREE (assume these; build on them, don't re-litigate):
  [ancestor + sibling insights, e.g. "Verification is not the bottleneck;
   candidate coverage is. Search-augmented judging overfits dev questions."]

OBJECTIVE & METRIC:
  [O and direction, e.g. "Maximize BrowseComp answer accuracy."]

DEVELOPMENT EVALUATOR (E_dev) — run this to score your candidate:
  [exact command, e.g. `python eval.py --split dev --n 50`]

WHAT TO DO:
  1. Create/confirm an isolated worktree from M_best so you don't touch other
     experiments or the current best.
  2. Implement the MINIMAL change that realizes the hypothesis. You may edit,
     debug, and rerun freely to get a working implementation — but keep the
     change bound to this hypothesis. If the metric stalls, fix YOUR code; do
     not pivot to a different idea.
  3. Run E_dev and record the score. Run it more than once if it's noisy.
  4. Commit the artifact on a clearly named branch.

RETURN EXACTLY THIS (your final message IS the data the coordinator reads):
  - dev_score: <number from E_dev>
  - result:    <1-3 sentences of factual outcome — what the change did>
  - insight:   <the reusable lesson: WHY this result supports / weakens /
               bounds the hypothesis. This is the most valuable output —
               make it a constraint future experiments can use, not a restatement
               of the score.>
  - branch_ref: <git branch/commit/worktree path holding the artifact>

Do NOT run the held-out test evaluator — that is the coordinator's merge gate.

---

After the executor returns, the coordinator records it with:

python scripts/tree.py set-evidence --node <id> \
  --dev-score <n> --result "..." --insight "..." --branch-ref "<ref>"

then abstracts the lesson upward with tree.py propagate.

Hypothesis Tree Refinement (HTR) — methodology and evidence

Background reference for the arbor skill. Source: Toward Generalist Autonomous Research via Hypothesis-Tree Refinement (Jin et al., 2026, arXiv:2606.11926; code: github.com/RUC-NLPIR/Arbor). Read this when you want the reasoning behind a design choice in the main loop.

The problem: Autonomous Optimization (AO)

AO is the operational core of autonomous research. An agent starts from an initial artifact and a research objective, then improves the artifact through experimental feedback without step-level human supervision. Formally a task is a tuple P = (M_0, O, E_dev, E_test):

• M_0 — mutable initial material (usually a codebase + its data).
• O — objective: what "better" means, as a metric direction over the

artifact's output.

• E_dev — development evaluator the agent may use freely during search.
• E_test — held-out test evaluator. Same objective, different evidence.

The goal is to return M* = argmax over candidates of S_test(M'), subject to the constraint that hypotheses and implementation decisions are made without using `E_test` as an exploration oracle. A candidate that exploits dev-split idiosyncrasies may raise S_dev but is not a successful AO solution unless the gain also transfers to S_test.

Why this is hard: feedback is delayed, experiments are expensive, and failed attempts contain information that should guide later search. If an agent treats each trial as an independent local attempt, it loses the structure of the research process — what was tried, what evidence came back, how each result reshapes the space of future hypotheses.

The three design requirements

HTR is built to satisfy three requirements that ordinary agentic tool use does not:

1. Branching with coherence. Multiple competing hypotheses can be plausible at once, so exploration must branch — but unrestricted branching degenerates into an unstructured log. The frontier must keep competing directions organized, comparable, and actionable. 2. Global strategy with local execution. Strategic decisions depend on evidence across the whole run; implementing one hypothesis is short-horizon code editing. Separate the two so low-level traces don't obscure the global state, and outcomes stay attributable to the hypotheses that produced them. 3. Exploration with held-out admission. Dev feedback guides search; artifact-level progress is admitted only when it transfers beyond that feedback. The system must distinguish exploratory dev improvement from verified test improvement.

The hypothesis tree as research state

A rooted tree T = (V, E). Each node is a research unit n = <h_n, iota_n, mu_n>:

• Hypothesis `h_n` — a verifiable/falsifiable claim about how changing the

material improves the objective. Granularity tracks depth: nodes near the root are broad directions; deeper nodes are concrete interventions an executor can implement and evaluate. This organizes exploration as progressive refinement rather than a flat sequence of independent trials.

• Insight `iota_n` — the reusable interpretation of evidence. For an

executed leaf: what was tried, what happened, and why the result supports, weakens, or constrains the hypothesis. For an internal node: an abstraction over its children's insights — the current understanding of that direction. It is not an execution transcript; it is compact semantic memory for later ideation and selection.

• Metadata `mu_n` — connects the semantic hypothesis to executable evidence:

node status, dev score, factual result, implementation reference (git branch or commit), optional background. The material itself is not duplicated in the tree — only references to external artifact states produced in isolated worktrees. This keeps the state compact while every hypothesis stays grounded in a verifiable implementation.

Internal nodes hold abstract directions and accumulated lessons; leaves hold candidate interventions to dispatch. After a leaf executes, its score, result, artifact ref, and insight are written back, and the insight is propagated upward along the path to the root. Through this abstraction, local outcomes become direction-level lessons and eventually a compact global understanding.

The tree therefore plays three roles at once: a search frontier (which directions are active/validated/pruned), a long-term memory (reusable evidence from successes and failures), and an auditable record (each artifact change linked to the hypothesis and evidence that motivated it).

The coordinator–executor split

• A persistent coordinator owns the shared tree and decides where to expand,

which evidence to trust, what to prune, and when to merge. It sees the whole frontier but does not perform every low-level implementation step.

• Short-lived executors are invoked to test one hypothesis each. An executor

gets h_n, relevant ancestor insights, and the current best artifact; it creates an isolated git worktree, implements the minimal change h_n requires, evaluates on E_dev, repairs its own broken/inactive code, and returns structured evidence.

The boundary is the point: exploratory code changes stay isolated until they pass the merge gate, and the tree records only decision-relevant evidence (scores, factual outcomes, artifact refs, distilled insights) rather than a raw log of tool calls. This is how transient execution traces become persistent research state.

Executors are hypothesis-bound (and why)

An executor's local loop may involve many edits and reruns, but it stays bound to the assigned hypothesis: h_n is fixed. If an executor were allowed to change the hypothesis when the metric stalls, the returned score would no longer be evidence about the assigned node, and ancestor insights built from it would become impossible to interpret. Keeping executors hypothesis-bound preserves the semantic meaning of every tree update while still allowing local engineering flexibility.

The six-step cycle (Algorithm 1, HTR)

Each coordinator cycle is a controlled mutation of the tree through a narrow interface:

1. Observe — re-ground in a structured projection of the tree (frontier, root/global insights, ancestor insights, current best). Makes the tree the authoritative state after context compression, instead of relying on lossy conversation history. 2. Ideate — under a chosen parent, propose k child hypotheses, each a refinement/alternative/correction. Ideation is conditioned on tree evidence: validated insights are assumptions to build on, pruned nodes are negative constraints, recent reports suggest what's feasible or under-tested. 3. Select — choose pending nodes to execute. Balance expected utility against the evidence already accumulated around ancestors and siblings. A node may be selected because it has strong prior evidence, because its siblings exposed an unresolved ambiguity, or because its failure would clarify an important assumption. Selection is frontier control under partial, delayed feedback — not raw score maximization. 4. Dispatch — selected hypotheses go to independent executors in fresh worktrees. Parallel sibling execution yields comparative evidence within one direction, which feeds later pruning and abstraction. 5. Backpropagate — write each executor's evidence into its leaf, then update insights along the path to the root. The propagated signal is not just a scalar: it includes causal attributions, applicability conditions, and reusable lessons. A leaf-level data-interface mismatch can become a direction-level constraint and then a global prior. 6. Decide — continue expanding a direction, prune a falsified subtree, or attempt a merge. Promotion is guarded by the held-out merge gate: the candidate is evaluated on E_test in a fresh worktree and merged into M_best only if it improves under O. This separates exploratory success on E_dev from verified artifact-level progress.

Empirical lessons (use these to prioritize effort)

From the paper's experiments across six AO tasks (model training, harness engineering, data synthesis) plus MLE-Bench Lite:

• Insight feedback is the dominant component. Ablating insight propagation

while keeping the tree caused a larger drop than removing the tree entirely (on MLE-Bench Lite: full 81.82% any-medal vs. 54.54% w/o insight feedback vs. 63.64% w/o tree). Hierarchy alone is not enough — a tree without propagated lessons organizes experiments syntactically but provides no semantic memory. Invest your judgment in the abstraction at Backpropagate, not just in generating more hypotheses.

• Structured search, not a bigger budget. Arbor used a comparable token

budget to single-trajectory baselines (~20–43M tokens) yet got larger held-out gains. The win is in how the budget is organized: maintaining competing hypotheses, isolated execution, comparison, and an updated frontier.

• The dev/test split exposes overfitting. Across tasks, many nodes improved

dev but only a subset passed the test gate. On Terminal-Bench, the highest-dev candidate was not the best on test. Always report the merged-vs-explored gap honestly; a high-dev/low-test result is evidence of feedback exploitation.

• Refinement deepens task understanding. Early nodes test whether a broad

mechanism holds; later nodes localize where it stops working; ancestor insights compress these into the constraints the final design must satisfy. Successful proposals are usually evidence-conditioned responses to earlier failures, not fresh guesses.

• Lessons transfer. A harness optimized only on one task's dev feedback

improved unrelated held-out tasks, indicating HTR discovers generally useful design changes rather than fitting the source benchmark — when, and only when, the merge gate is enforced.

• *What HTR does not fix.* Arbor is strongest at a sequence of concrete

refinements once a runnable solution exists. It is weaker when progress requires a genuinely new high-level formulation only weakly connected to the current tree — that still leans on good human task design (the choice of M_0, evaluator, metric, and interface).

Final report template

Produce this when the run ends (budget spent, frontier exhausted, or progress stalled). The point is an honest, auditable account — not a victory lap. Keep it concise and grounded in the tree.

# Arbor run report: [objective]

## Result
- **Best artifact**: [git branch/ref of M_best, e.g. `arbor/best`]
- **Test score**: [S_test of M_best] vs. initial [S_test of M_0]  →  delta [Δ]
- **How to check it out**: `git checkout [ref]`
- One-line summary of the change that won.

## What was tried (audit trail)
[Paste `python scripts/tree.py status` — the tree shows every direction,
which were pruned, which merged, with dev/test scores.]

## How understanding evolved
2-4 bullets tracing the main hypothesis shifts: which early nodes tested broad
mechanisms, what they confirmed or ruled out, and how that reshaped later
hypotheses. The story should explain *why* the final design looks the way it
does — i.e. the constraints the run discovered.

## Dev vs. test (overfitting check)
- Nodes that improved **dev**: [count]
- Nodes that passed the **test merge gate**: [count]
- Comment on the gap: were there high-dev / low-test candidates? What did
  rejecting them tell you? An honest gap here is more trustworthy than a clean
  "everything worked".

## Open directions
What you'd explore with more budget, and any direction that seemed to need a
new high-level formulation rather than further refinement (HTR's known weak
spot — flag it for the human).

Always leave M_best as a real, runnable artifact on a named git branch.