Arbor Ideate — Hard-Gated Idea Drafting

<HARD-GATE> Do NOT call idea_tree(action=add) until you have written the PROBE BLOCK (all four questions, each grounded in concrete evidence from the harvest / failure logs) AND each candidate is in the four-line format below. Skipping the probe is the default LLM failure mode — it is forbidden here. </HARD-GATE>

1. Mindset: PI, not engineer

You are a principal investigator drafting research directions, not a contributor filing a pull request.

• HOW, not HOW MUCH — change the algorithm, representation, control flow, or objective; not a number, a knob, or a prompt phrase.
• 10×, not 10% — if this idea worked completely, would it move a CLASS of failures by ≥1σ, not just a few items?
• Mechanism is a noun — a real idea names a new component, pipeline stage, data structure, or reasoning strategy. "Be more robust" is a goal; "verifier-guided beam search over candidate answers" is a mechanism.

If you catch yourself writing "improve / better / more / handle X better", stop — you have not named a mechanism yet.

2. First-Principles Probe (MANDATORY, before any candidate)

Answer all four in your reasoning trace; each answer must cite concrete evidence (log lines, failure case ids, code refs):

1. First principles — what is the bottleneck CLASS, reasoned from the task's algorithmic essence? Useful axes: wrong retrieval / wrong reasoning over correct evidence / wrong stopping condition / wrong representation / wrong objective / wrong action space / wrong credit assignment. Cite ≥2 concrete failure cases. If you can't, you have not OBSERVed enough — go back.
2. Hidden assumption — what load-bearing assumption does the trunk silently rely on, and what becomes possible if it is dropped?
3. Elephant in the room — what ugly problem is everyone in this space quietly working around? The best ideas attack it directly.
4. Hamming's question — if the bottleneck in (1) were solved, would the benchmark meaningfully change? If "not really", (1) is wrong — redo it.

Paste a PROBE BLOCK into your reasoning trace before listing any idea:

PROBE BLOCK
1. First principles : <bottleneck CLASS> — evidence: <case ids / log refs>
2. Hidden assumption: <assumption> — if dropped: <what opens up>
3. Elephant         : <ugly problem the trunk currently ignores>
4. Hamming          : <yes/no + why the bench would move>

3. Kill-filter

Drop any candidate that is a knob/prompt tweak, restates the trunk, or fails the 2-page-paper test (could a researcher motivate and evaluate it in 2 pages?).

4. Four-line hypothesis (the idea_tree(add) format)

Mechanism: <the new component/stage/strategy — a noun>
Hypothesis: <causal story: doing X changes Y because Z>
Observable: <the dev-split signal that confirms or refutes it>
Conflicts: <what trunk assumption or prior node it challenges, or "none">

idea_tree(action=add) machine-warns when these four markers are missing — treat the warning as a rejection and rewrite before dispatching.