Evidence-first replications of quantitative-research papers. Every replication card separates the paper's claim, available implementation evidence, local observations, and unresolved deviations.
| Paper | Status | Claim ceiling |
|---|---|---|
| AutoAlpha (2020) | Method reconstruction complete; strict promotion record closed | implemented, not deployable |
| DoubleEnsemble | Quality-filtered model smoke comparison recorded | implemented, smoke only |
| OpenFE | Controlled feature-combination search smoke recorded | implemented, smoke only |
| SpatioTemporal Representation | Event/entity context layer smoke recorded | implemented, context only |
| FactorMAD | Multi-agent proposal/audit workflow smoke recorded | implemented, workflow only |
python -m unittest discover -s tests -v
python scripts/verify_cards.pyThe checks validate that each public card carries source, scope, deviation, and claim-boundary sections, and that each machine-readable paper card links to existing evidence. They do not validate trading performance.
The disclosure boundary is recorded in DISCLOSURE.md, and
sample_data/evidence_matrix_sample.csv is a synthetic schema fixture.
- Do not upload employer code, private data, tokens, or non-public experiment logs.
- Do not report a replication as successful without a runnable command and an expected-versus-observed comparison.
- Record negative or partial results as first-class evidence.
reproductions/ human-readable replication cards
paper_cards/ machine-readable claim and evidence metadata
schemas/ lightweight schema contracts for paper cards
scripts/ static verification commands
tests/ regression tests for the evidence contract