Arbor Ideate — Hard-Gated Idea Drafting
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):
- 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.
- Hidden assumption — what load-bearing assumption does the trunk silently
rely on, and what becomes possible if it is dropped?
- Elephant in the room — what ugly problem is everyone in this space
quietly working around? The best ideas attack it directly.
- 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.
1---2name: arbor-ideate3description: 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.4license: MIT5---67# Arbor Ideate — Hard-Gated Idea Drafting89<HARD-GATE>10Do NOT call idea_tree(action=add) until you have written the PROBE BLOCK11(all four questions, each grounded in concrete evidence from the harvest /12failure logs) AND each candidate is in the four-line format below. Skipping13the probe is the default LLM failure mode — it is forbidden here.14</HARD-GATE>1516## 1. Mindset: PI, not engineer1718You are a principal investigator drafting research directions, not a19contributor filing a pull request.2021- **HOW, not HOW MUCH** — change the algorithm, representation, control flow,22 or objective; not a number, a knob, or a prompt phrase.23- **10×, not 10%** — if this idea worked completely, would it move a CLASS of24 failures by ≥1σ, not just a few items?25- **Mechanism is a noun** — a real idea names a new component, pipeline stage,26 data structure, or reasoning strategy. "Be more robust" is a goal;27 "verifier-guided beam search over candidate answers" is a mechanism.2829If you catch yourself writing "improve / better / more / handle X better",30stop — you have not named a mechanism yet.3132## 2. First-Principles Probe (MANDATORY, before any candidate)3334Answer all four in your reasoning trace; each answer must cite concrete35evidence (log lines, failure case ids, code refs):36371. **First principles** — what is the bottleneck CLASS, reasoned from the38 task's algorithmic essence? Useful axes: wrong retrieval / wrong reasoning39 over correct evidence / wrong stopping condition / wrong representation /40 wrong objective / wrong action space / wrong credit assignment. Cite ≥241 concrete failure cases. If you can't, you have not OBSERVed enough — go back.422. **Hidden assumption** — what load-bearing assumption does the trunk silently43 rely on, and what becomes possible if it is dropped?443. **Elephant in the room** — what ugly problem is everyone in this space45 quietly working around? The best ideas attack it directly.464. **Hamming's question** — if the bottleneck in (1) were solved, would the47 benchmark meaningfully change? If "not really", (1) is wrong — redo it.4849Paste a PROBE BLOCK into your reasoning trace before listing any idea:5051```52PROBE BLOCK531. First principles : <bottleneck CLASS> — evidence: <case ids / log refs>542. Hidden assumption: <assumption> — if dropped: <what opens up>553. Elephant : <ugly problem the trunk currently ignores>564. Hamming : <yes/no + why the bench would move>57```5859## 3. Kill-filter6061Drop any candidate that is a knob/prompt tweak, restates the trunk, or fails62the 2-page-paper test (could a researcher motivate and evaluate it in 2 pages?).6364## 4. Four-line hypothesis (the idea_tree(add) format)6566```67Mechanism: <the new component/stage/strategy — a noun>68Hypothesis: <causal story: doing X changes Y because Z>69Observable: <the dev-split signal that confirms or refutes it>70Conflicts: <what trunk assumption or prior node it challenges, or "none">71```7273`idea_tree(action=add)` machine-warns when these four markers are missing —74treat the warning as a rejection and rewrite before dispatching.
invergent-ai