Skip to content

Global Decompression Benchmark

Compare decompression models against evidence, not reputation.

The benchmark evaluates documented models, tables, procedures and configurations against standardized historical exposure and outcome data. It is designed to show where evidence is strong, where performance changes with configuration and where the available data are not sufficient to decide.

What the benchmark evaluates

The benchmark separates model family, version, implementation and configuration. A result is attached to the exact system that produced it, rather than to a broad model name that may hide material differences.

Evaluation is constrained by the evidence available for each exposure. OpenDeco distinguishes calibration data from independent validation data and does not treat a counterfactual schedule as an observed outcome.

Explore the current release

Empty

No public release yet.

Draft and internal research runs are never exposed as released evidence.

Views in the explorer

  1. 1Overview
  2. 2Models
  3. 3Datasets
  4. 4Results
  5. 5Failure regions
  6. 6Sensitivity
  7. 7Methods
  8. 8Downloads
  9. 9Changelog

How to read a benchmark result

A benchmark score is not a universal safety ranking. Performance depends on the dataset, outcome definition, model configuration, calibration history and question being asked. OpenDeco therefore publishes the result together with its scope, uncertainty and failure regions.

Identity

Check the exact model version, implementation and configuration before comparing results.

Evidence

Check whether the evaluated dataset helped create or calibrate the model. Training-data performance is not independent validation.

Failure regions

Average performance can conceal exposure regions where a model behaves differently. OpenDeco keeps those regions inspectable.

Reproduce the release

Every public benchmark release carries the dataset versions, model identities, code commit, run manifest, protocol version and artifact checksums needed to understand what was executed. Where redistribution or proprietary restrictions prevent exact reproduction, the limitation is stated explicitly.