The problem
Investment ideas can look convincing after the fact. Individual investors need a way to test hypotheses without future-data leakage, silent parameter drift, or an automatic path from research to trading.
What I built
AI Investment Research & Strategy Lab is an audit-first personal finance research tool. A frozen experiment contract binds the question, dataset schema, SHA-256 inputs, decision gates, and safety boundaries before a run begins.
The working demo:
- rejects feature rows that were not available before the sample date;
- keeps 2025 validation and 2026 out-of-distribution windows sealed;
- compares high and low feature tails only inside equal-date momentum groups;
- produces a supported or rejected verdict instead of optimizing for an attractive backtest;
- independently rebuilds the result and checks inputs, code, lineage, portability, and closed execution surfaces;
- renders self-contained HTML, SVG, PNG, JSON, and Markdown reports.
The bundled synthetic margin-quality experiment intentionally ends with hypothesis_rejected. That negative result is the product: it is explicit, reproducible, and safe to review.
A sanitized Round23 case study demonstrates the same control pattern on aggregate real-system evidence. It found no provider-invariant margin signal, left validation and OOD sealed, and stopped before the next round.
Try it
PYTHONPATH=src python3 -m ai_investment_lab demo
PYTHONPATH=src python3 -m ai_investment_lab verify
python3 -m unittest discover -s tests -v
Python 3.9+ is required. There are no runtime dependencies. Synthetic sample data and generated artifacts are included.
How Codex and GPT-5.6 were used
Codex with GPT-5.6 audited the existing private research system and selected the smallest safe vertical slice for a judgeable package. It implemented the frozen contracts, deterministic experiment, independent verifier, report renderer, adversarial tests, and documentation.
During review, Codex caught and fixed two fail-closed gaps: a manifest could reopen execution without failing verification, and a delivery receipt did not bind its own safety fields. It also traced a wheel metadata failure to an old macOS build frontend and moved the single metadata source to setup.cfg for compatibility.
The final package was verified with 10 adversarial and regression tests, 13 independent checks, byte-identical output on Python 3.9 and 3.12, a standard wheel build, and clean path and credential scans.
Safety and privacy
This is research software, not investment advice. It contains no holdings, balances, credentials, licensed raw provider data, broker integration, order execution, production promotion, or automatic next round. Every output is manual-review-only.
Log in or sign up for Devpost to join the conversation.