METHOD / MODELS
Reconstruct the system that actually produced the answer.
OpenDeco separates a named decompression model from the version, implementation and configuration that turns it into executable behaviour.
CautionBoth reference models run on invented parameters
Model identity
A family name is not enough. Each runnable record identifies the family, version or coefficient set, implementation, configuration and supported exposure capabilities.
In this repository today
A runnable record declares which of the 15 capabilities spec 125 names it supports. The call gateway checks that declaration before the adapter is reached, so a trimix profile sent to an implementation that declares no helium is refused outright. The alternative is an answer that looks ordinary and is wrong.
Primary-source implementation
OpenDeco prefers primary equations, official source code, original tables and authoritative technical documentation. Secondary descriptions may guide discovery but do not silently replace a primary source that is available.
In this repository today
No published model has been reconstructed here. What exists is a record of the sources that would be needed. Every parameter group names the kind of source that must supply it. That source also has to state the unit, the pressure reference frame and the population a fit was estimated against.
No entry names a work, an author, an identifier or a year. A test fails the build if a gap record acquires a DOI or a URL, because an invented citation is worse than a stated gap.
Recreated versus original
A recreated implementation is labelled as recreated even when it matches published examples. OpenDeco does not imply that reconstructed code is the original author's implementation.
Numerical verification
Model adapters are tested against published examples, golden vectors, table values or other defensible reference outputs. Numerical tolerance and platform differences are documented instead of hidden.
In this repository today
Verification runs V0 to V3. A harness computes the level the evidence supports and ignores the level an adapter declares for itself, then names each unmet requirement in a shortfall.
Both reference models sit at V1. V2 needs agreement with an authoritative implementation, published examples or verified device output. Neither model exists outside this repository, so no authoritative version of either one can be obtained. Neither carries a cross-check record, neither carries an independent review, and the metadata type refuses to let either claim more.
Configuration
Settings such as gradient factors, ascent constraints, gas options and conservatism parameters are versioned as configuration, not promoted into separate model families.
In this repository today
The configuration sweep planner refuses to plan a grid whose combination count passes its ceiling, and above a declared cap it counts combinations instead of enumerating them. A sweep is therefore bounded before anything is queued.
Proprietary and black-box systems
Where lawful observation is possible but source code is not, OpenDeco records the system as black-box and limits claims to observed behaviour. It does not invent an internal algorithm from output resemblance.
Versioning
A scientific change to an adapter creates a new implementation version. Released benchmark results remain tied to the previous version rather than changing in place.
In this repository today
A digest in a diff is a prompt to review, and a developer can regenerate a golden file and commit its new digest in one motion. The acceptance ledger makes that motion fail. Each accepted baseline records who accepted it, when, why and where the review is written down. Entries form a chain, so deleting or reordering one breaks the ledger.