Repository navigation
What changed
Version 1.1 preserves the five frozen v1.0 compatibility cases and adds three BigBear.ai cases covering the complete knowledge-state transition: previously reported, withdrawn from reliance, and restated.
- 8 accession-aware PIT cases
- 2 evaluation tracks: evidence-restricted retrieval and knowledge-contamination control
- 16 machine-readable prompts and reference responses
- required abstention during the BigBear.ai non-reliance interval
- deterministic scorer with a temporal-integrity hard-failure rule
- byte-identity checks for every frozen v1.0 artifact
The bundled reference fixture passes 16/16 responses. The deliberate future-value leak in the adversarial fixture is rejected. These results validate the fixtures and scorer; they are not a result for any third-party model or investment strategy.
Permanent citation and audit paths
- Frozen version DOI: https://doi.org/10.5281/zenodo.22289017
- All-versions concept DOI: https://doi.org/10.5281/zenodo.22279743
- Human-readable research explanation: https://ahasignals.com/research/point-in-time-backtest-validation-benchmark/
- GitHub is the public code, version and verification entry point.
- Zenodo is the permanent archive and canonical citation object.
Exact archive
ahasignals-pit-backtest-validation-benchmark-v1.1.0.zip
SHA-256: a7c755dee3b7b800505274f6c680075dd856196d6c02775462a4ac874187bc85
Interpretation boundary
This benchmark evaluates data-time integrity. It does not establish factor alpha, information coefficient, portfolio return, Sharpe ratio, drawdown or general model intelligence. Research and education only; not investment, trading, legal, accounting or tax advice.