The research question
How do quantum reservoir configurations compare with classical forecasting methods under a controlled evaluation? The study considered synthetic chaotic systems including Hénon, Lorenz-63, Mackey–Glass, and Lorenz-96, alongside classical ESN, NG-RC, and ELM methods.
The engineering problem underneath
Large experimental sweeps are easy to run badly. A preprocessing leak, a stale cached feature, or a partially written result can invalidate a comparison. I built the benchmark around traceability, repeatable configuration, and recovery from interrupted work.
FROM CONFIGURATION TO EVIDENCE
What I implemented
- Chronological splits and training-fitted preprocessing to reduce temporal leakage.
- Deterministic seeds and configuration identities for repeatable comparisons.
- Content-addressed feature caching with atomic writes.
- Restartable sweeps that distinguish completed work from missing results.
- Parquet artifacts and DuckDB queries for inspecting experiment outputs.
Testing the experiment infrastructure
Completed cells skipped during an unchanged 12-cell rerun.
Cells re-executed when resuming a controlled interrupted run.
Injected fault classes detected before final results.
These are outcomes from the recorded engineering checks, scoped to those experiments. They demonstrate the intended caching, recovery, and validation behaviour rather than a guarantee about every future run.
The finding
No evaluated quantum configuration beat the strongest classical comparator in this study. The value of the project is in a reproducible comparison and an honest result, including when that result does not support the more exciting hypothesis.
Why it matters to my engineering work
The same habits transfer to applied AI: preserve provenance, compare against a credible baseline, test failure cases, and make it possible to trace a headline number back to its inputs.