docs/consultant_reference.md

# Consultant research reference

This compact reference is distilled from the supplied WorldQuant BRAIN consultant
documents. It is intentionally kept as text so the Hugging Face Space can ingest
it without Git LFS/Xet binary storage.

## Ordinary consultant submission gates

- Delay 1: target Sharpe above 1.58 and Fitness above 1.0.
- Delay 0: target Sharpe above 2.69 and Fitness above 1.5.
- Regular candidates should have turnover above 1% and below 70%.
- A simulation passing local metrics is not a platform submission; BRAIN must
  accept the submission checks and show the alpha as submitted/active.

## Research quality rules

- Prefer a documented economic hypothesis tied to the selected dataset.
- Use a small, valid operator tree and verify every operator's arity before
  spending a simulation.
- Avoid repeated horizon-only mutations when a formula family repeatedly misses
  Sharpe or Fitness. Move search budget to a different data/operator family.
- Preserve the seed's native region, universe, delay, neutralization, decay, and
  truncation when performing controlled mutations.
- Treat Sharpe, Fitness, turnover, and simulation validity as a joint score;
  fitness alone is not evidence of a useful research parent.

## Power Pool separation

Power Pool is a separate themed campaign with its own universe, data, correlation,
and submission requirements. Its lower/alternate thresholds must not be applied to
ordinary consultant research. Keep the Power Pool engine and ordinary research
engine as separate campaign loops.

## Generation guardrails

- `group_neutralize` takes one expression and one literal group field.
- Time-series operators must receive exactly the supported number of arguments.
- Do not emit undocumented operators or computed group fields.
- Keep document context bounded so the model receives the actual target settings
  and grammar constraints in every request.