# Abyssi decompression engine numerical study

**Study snapshot: 2026-11-14.v1** · 14 November 2026

Article: https://abyssi.io/research/decompression-engine-study/

The authors are the Abyssi development team. They develop the evaluated engine and have a commercial interest in Abyssi. This is a self-authored software evaluation. It is not independently reviewed, peer reviewed or clinically validated. The reference-project maintainers did not author or endorse the study.

## What can be reproduced publicly

This package publishes numerical inputs, reference outputs, frozen Abyssi outputs, a profile matrix, methods and file hashes.

Run `python3 methods/verify_data.py` from the extracted package to verify hashes and recompute arithmetic comparisons between the published records. This checks the dataset; it does not execute Abyssi.

With Git and Rust/Cargo installed, run `bash methods/reproduce_reference.sh` to fetch the pinned open-source Rust implementation and run the original reference harness using Cargo's locked dependencies. The rerun writes a separate output file and does not replace the published capture.

**The private Abyssi engine, internal tests and DecoTengu adapters are not distributed.** Their recorded hashes identify the authors' frozen implementation but do not independently verify it. An outside reader cannot completely regenerate all 33 comparisons from this package. This is a public data-and-method audit with a separately rerunnable native reference, not a claim of full reproducibility.

## Dataset and denominators

The automated DecoTengu-reference matrix contains **33 entries and 31 distinct normalized profiles**:

- **12 original entries:** square profiles, gas changes, freshwater, supplied altitude pressure and no-decompression cases.
- **11 extended entries:** four 6 m final-stop cases, three complete multilevel cases, and two repetitive first/second-dive chains. Two antecedent first dives reuse original profiles; they are required carryover checks, not new distinct experiments.
- **10 phase-rate entries:** profiles evaluated with an adapter for initial/decompression/final ascent rates of 18/10/9 m/min.

`profile-matrix.csv` and `study-data.json` enumerate every entry. Duplicate profiles identify their original IDs. `data/inputs-and-reference.json` records exact segment durations, gas fractions, GFs, rates, stop geometry, pressure settings, residual tissues and comparison policies. `data/abyssi-captures.json` records the authors' frozen actual outputs. The three `data/decotengu-*-reference.json` files preserve the reference fixtures.

All 33 stop sequences and stop oxygen fractions agree. The runtime tolerance is 1e-8 minutes. Each of 16 nitrogen and helium final compartment pressures is compared at 1e-10 bar. The observed maximum final-pressure difference is recorded in the structured results; it is not a physiological uncertainty estimate.

A **separate four-profile matrix** compares unchanged upstream dive-deco 6.1.2. All 96 coefficients and 128 pressure-matched bottom-tissue values agree within the stated criteria. Three decompression schedules remain unequal; one no-decompression case agrees to floating-point precision. Both positive and negative results are retained. These four differential cases are not counted as four additional exact schedule matches.

The authors' internal engine run passed **322 tests in 29 suites**. Some tests cover multiple schedule entries; others cover boundaries, warnings or unrelated behavior. This is not a count of independent profiles, external validations or physiological trials. The private suite cannot be rerun from this package.

## Normalization and reference adapters

The DecoTengu contract uses the 1a first-compartment convention, 16 nitrogen/helium compartments, 1.01325 bar sea-level surface pressure, gravity 9.80665 m/s², water vapour at 0.0627 bar and initial dry-air nitrogen fraction 0.7902. Saltwater density is 1025 kg/m³; the specified freshwater case uses 1000 kg/m³. Descent is 18 m/min. Original and extended ascents use 9/9/9 m/min. The grid and final stop are 3 m unless a 6 m final stop is explicitly selected. Runtime includes descent.

The acclimatized 1500 m case supplies a normalized surface pressure. It checks scheduling at that pressure, not the altitude formula, travel acclimatization or altitude safety.

Extended multilevel and surface-interval inputs are loaded through the external model API before its native ascent scheduler runs. In repetitive tests, Abyssi computes the first dive and carries its actual returned surface tissues into the second. Its production initialization applies the surface interval. Starting tissues are compared with the independent post-interval state before the complete second schedule is checked.

DecoTengu exposes one ascent rate. The mixed-rate adapter changes it at the decompression-phase boundary and rescales final transit while upstream retains stop, GF, gas and tissue decisions. Before mixed-rate generation, all ten adapter captures at 9/9/9 are asserted identical to stock DecoTengu at 9 m/min. This is a tested adapter assumption within one reference implementation, not native upstream three-phase support or corroboration from ten separate algorithms. Pinned private APIs are used; adapter source is not distributed in this release.

### Six-metre oxygen-policy exception

Pure oxygen at 6 m has a partial pressure of 1.616358975 bar under the normalized saltwater model. The reference honors the requested switch; Abyssi's default 1.6 bar cap would move it shallower. To compare that same physical switch, extended cases explicitly permit 1.62 bar and mixed cases 1.7 bar. Exposure warnings remain recorded.

These are numerical-test acquisition settings, not oxygen-limit recommendations or operational dive plans. Default-cap parity and new exposure or gas-budget validation are not established by these cases.

### Second-reference normalization and negative findings

Rust uses depth/10 bar, 1013 mbar surface pressure and 0.79 initial-air nitrogen. The authors' pressure-matched capture maps physical depth, stop spacing and rates and supplies matching initial tissues without changing either engine. Bottom-tissue mathematics agrees, but the three decompression schedules still differ.

Rust's cached unrounded GF anchor, integer-percentage GF and one-second/current-ceiling stop decisions differ from Abyssi's rounded first-stop anchor, continuous GF, destination-GF departure check and whole-minute minimum holds. No one-factor ablation assigns every runtime delta to one policy. Longer or shorter software schedules do not demonstrate relative safety.

## Freeze and audit limits

`provenance.json` records the study snapshot ID, actual engine calculation-revision label, source-unit and aggregate hashes, reference pins, upstream-source hashes, internal generator hashes, fixture hashes and toolchains. The aggregate source digest identifies the dirty working tree actually evaluated; a repository commit alone would not do so. Private source contents and paths are not included.

`SHA256SUMS` covers every distributed content file except itself and the ZIP archive. The ZIP contains the same public files. No private engine/application source, proprietary bundled source, commercial screenshots, credentials, usernames or local absolute paths are included.

The original acquisition plan enumerates 145 case IDs, not its former total of 148. There are 105 later uncaptured cases in each original commercial acquisition track. The independent matrix in this study does not fill those missing vendor rows. The article distinguishes unsupported finer stop resolution, vendor-specific oxygen policies, air breaks, extended/accelerated-stop policies, transient-altitude uncertainty and excluded CCR. This public release provides no commercial ranking or proprietary implementation analysis.

All cases were selected by the authors. The sample is neither random nor exhaustive. Important gaps include independent approval of the scheduling contract, arbitrary stop geometry, broader repetitive/multigas coverage, transient altitude, any future CCR support and clinical outcomes. No clinical validation, certification, regulatory approval, prediction of individual illness risk or fitness for operational reliance is established.

## Acknowledgments and upstream sources

- **DecoTengu 0.14.1**, Artur Wroblewski and contributors; GPL-3.0-or-later. Source/version: https://pypi.org/project/decotengu/0.14.1/ . Documentation: https://wrobell.dcmod.org/decotengu/ . This package distributes numerical outputs, not the library or GPL-dependent adapters. The source-distribution checksum is in the provenance manifest.
- **dive-deco 6.1.2**, its contributors; MIT. Pinned source: https://github.com/KG32/dive-deco/tree/7f70ce0feba1b907c8baa18d7f3227c5e57050fe . The original reference harness calls its public API. Upstream tracked source was unchanged in the authors' capture.

Project names identify reference software; no affiliation, endorsement or independent review is implied. See `NOTICES.md` for the limited, file-specific helper license. The private Abyssi engine is not licensed by this package.
