Skip to content

METHODS

How OpenDeco reaches a result.

Source selection, data normalization, model reconstruction, validation, uncertainty, leakage control, release engineering and reproducibility are documented as part of the research rather than hidden behind the final chart.

Core methods

  1. How OpenDeco records source provenance, preserves immutable originals, normalizes pressure and time, handles missingness and distinguishes source labels from derived fields.

  2. How model family, version, implementation and configuration are separated, verified and versioned before a result is benchmarked.

  3. Benchmark protocol

    /methods/benchmark

    How datasets and models become eligible, how evaluation modes and metrics are frozen, and how calibration overlap and sensitivity are handled.

  4. Reproducibility and releases

    /methods/reproducibility

    How immutable inputs, run manifests, code identity, checksums and clean reruns connect a public result to the system that produced it.

Calibration-data leakage

A model evaluated on data that helped create or tune it has not been independently validated by that evaluation. OpenDeco records known development relationships and exposes overlap warnings at dataset, model and benchmark-result level.

Uncertainty and sensitivity

Decompression evidence is sparse, heterogeneous and often historically constrained. OpenDeco reports uncertainty where it can be estimated and tests whether conclusions survive reasonable changes in configuration, inclusion criteria and analytical choices.

Safety and interpretation

OpenDeco research describes evidence and model behaviour. It is not an operational dive plan, a certification standard or proof that one schedule would have produced a different clinical outcome on an individual exposure.