| name | autolab-research-planner |
| description | Plan the most informative next directional-perps experiment from a small workspace brief. Use workspace search/open and a minimal evidence tool set when the default files are insufficient. Emit a mostly free-form research note. Never emit candidate JSON. |
Research Planner
Goal: decide the most informative next experiment, not the easiest local tweak.
Default context:
RUNBOOK.md
TASK.md
WORKSPACE_INDEX.md
current/SESSION_STATE.json
current/frontier_brief.md
current/market_brief.md
current/parent_card.md
Workflow:
- Read the default files first.
- If the answer is still unclear, browse the workspace with
search_workspace and open_file.
- If you do not fully understand a feature or need a semantically diverse alternative set, use
search_features, inspect_feature, or suggest_feature_set.
- Run probes only when they would discriminate between competing explanations.
- Write one research note in normal markdown.
Requirements:
- Prefer discriminating experiments over threshold tweaks.
- If evidence is contradictory, investigate before deciding.
- Cite only relative workspace
evidence_paths.
- Keep
tools_used truthful.
- Never emit candidate JSON.
- You may propose novel feature formulas; aliases are not the full limit.
- Only compose new features from the family manifest's listed aliases, raw series, and formula operators.
- The family manifest includes alias definitions; prefer existing aliases when they already express the intended signal.
- The note should make the intended family, named features, gate dimensions, and next test obvious enough for a deterministic extractor and the writer to preserve them.
- Make
must_answer effectively a concrete falsification question with a yes/no structure and a success criterion tied to pre-audit return, validation, or drawdown, even if you state it in prose.
- If you name a specific feature or gate dimension, say it explicitly in the note.
- If the workspace state says a non-regime axis of variation is required, do not answer with another regime-only carry variant.
- For
perp_multi_asset_carry, remember the family is a cross-sectional ranked long/short execution family. Carry features may rank the book, but realized returns can still be mostly price-led.
- For
perp_multi_asset_carry, treat feature mix, book structure, long/short counts, concentration, and regime suppression as the primary structural levers.
- Do not treat cross-sectional
trade_style as a primary evaluator-side lever for perp_multi_asset_carry unless the evidence packet explicitly shows it mattered.
- When switching families or trade styles, say so explicitly in the body.
- If the next experiment depends on a regime gate, include a
## Suggested gate spec section with validator-legal shapes only:
- string expression, e.g.
ge(pair_corr_72h,0.9)
- dict with
expression and optional min / max, e.g. {"expression":"market_volatility_168h","max":0.0085}
- do not invent keys like
op, condition, threshold, or active
- If you include tiny thresholds, prefer plain decimals like
0.000015 instead of scientific notation.
Output shape:
- Use the template in
templates/research_note.template.md.
- The body should usually include:
## Diagnosis
## Evidence
## Competing explanations
## Proposed next experiment
## Suggested gate spec when gates are central to the idea
## Risks
Read references/workspace_guide.md only if you need a quick reminder of the workspace contract.