AI Alpha Research Lab

A formula-alpha sandbox with hard rejection rules. The point is not to make a demo look profitable; it is to keep the research loop honest when a candidate fails after time-ordering, out-of-sample checks, turnover, and costs.

Showcase

• Rejection Gallery: leakage, broken time order, cost erosion, turnover overload, and rejection accounting.

Related repos

• AI-Alpha-Research-Lab: formula search, evaluation, and rejection gates.
• Paper-Alpha-Replications: replication notes with claim ceilings.
• Quant-Research-Toolkit: reusable checks for factor panels and diagnostics.
• Strategy-Game-Agents: repeated-choice experiments and baseline agents.

What is here

• Allow-listed formula language: rank, delta, mean, and arithmetic.
• Point-in-time evaluation on stock-date panels.
• Promotion gate for OOS IC, turnover, and cost-adjusted return.
• Rejection-gallery helpers that turn failed candidates into inspectable evidence.
• Deterministic synthetic-data demo that rejects the candidate.
• Tests for parsing, future-field rejection, rejection accounting, and promotion logic.

Run

python -m pip install -e .
python -m ai_alpha_lab.demo
python -m unittest discover -s tests -v

Evidence boundary

The first public record is intentionally conservative. An AutoAlpha-style study ran through the search and validation pipeline, but the frozen protocol produced zero final candidates. See evidence/validation-summary.md, reports/rejection-gallery.md, and research_state.json. The public disclosure boundary is recorded in DISCLOSURE.md, with a tiny synthetic candidate-log fixture in sample_data/.

Layout

src/ai_alpha_lab/  public research core
tests/             regression and safety checks
evidence/          short evidence records
reports/           recruiter-readable public summaries
research_state.json current claim ceiling and reopen condition

Not included

• private data, employer code, credentials, or proprietary research assets
• tradable-performance claims
• automatic capital deployment