Before trusting a Sharpe ratio: six reproducible pipeline invariants

[initial-d]

A new deterministic engineering audit is now available on main.

Run it without a market-data account:

git pull
python -m pip install -e ".[dev]"
make technical-audit

The maintained reference run checks six implementation invariants:

1. the complete factor catalogue produces a finite [180, 40, 213] tensor;
2. the synthetic panel is bitwise deterministic at a fixed seed;
3. poisoning masked cells with magnitude 1e9 causes zero drift at valid outputs for five core tensor primitives;
4. one-period forward labels stop at the sample boundary and use endpoint tradability masks;
5. the controlled backtest path matches weights[t-1] × returns[t] exactly;
6. cumulative transaction-cost drag scales 10× when the configured fee moves from 1 bps to 10 bps.

All six checks pass on commit 300e1ad. The new test is included in the full suite, which now passes 97 tests locally.

• Technical deep dive
• 中文技术文章
• Reference Markdown report
• Machine-readable JSON

This is an engineering audit on deterministic synthetic data. It does not establish market alpha, out-of-sample performance, or production readiness. Reproductions on other Python, PyTorch, CPU, and GPU environments are welcome.

[initial-d]

Independent audit runs now have a structured report path:

Open the technical-audit reproduction form

Please report the commit, environment, exact command, and unedited output whether all six checks pass or not. The most useful follow-up is a concrete disagreement with the reference run—non-zero mask drift, a label-boundary mismatch, execution-timing drift, cost-scaling drift, or an environment where the audit cannot complete. Reproducible discrepancies can be turned into regression tests or documented compatibility limits.