Decompression engine
numerical study
Agreement in tissue mathematics does not establish agreement between complete schedules—or diver safety.
Public data audit; partial reproduction. The private Abyssi engine and internal fixture adapters are not distributed. The package cannot independently regenerate all Abyssi comparisons. A native open-source reference can be rerun separately.
Abstract
We evaluated a frozen version of Abyssi’s Bühlmann ZH-L16C engine against two open-source implementations. The selected matrix covers square, nitrox and trimix profiles; decompression-gas switches; freshwater and supplied altitude pressure; 6 m final stops; fractional multilevel inputs; repetitive tissue carryover; and uniform or phase-dependent ascent rates. These are numerical test vectors, not operational dive instructions.
Thirty-three DecoTengu-reference fixture entries, representing 31 distinct normalized profiles, agree on stop sequences and gases. Runtime differences remain below 10−8 minutes and every final nitrogen/helium compartment pressure remains within 10−10 bar. Two antecedent dives are reused in repetitive chains and are not counted as new distinct profiles. The observed maximum final-tissue difference is 1.02e-12 bar.
A separate, unchanged dive-deco 6.1.2 capture corroborates all 96 coefficient values and the bottom-tissue mathematics of four pressure-matched profiles. Its three decompression schedules do not exactly match Abyssi. The no-decompression profile agrees to floating-point precision. Material differences in gradient-factor anchoring and stop decisions remain; no experiment apportions every runtime difference to one policy.
The supported conclusion is numerical agreement within a declared, selected model contract. This study does not establish an externally approved scheduling policy, individual decompression-illness risk, comparative safety, or fitness for operational reliance.
Methods
Implementations and freeze
The evaluated private engine identifies its calculation semantics as abyssi-engine-v2. Because the evaluated working tree contained uncommitted work, a repository commit alone would not identify it. We freeze the actual source through hashes of its 21 Swift units and an aggregate digest in the provenance manifest. Source contents are not published, and the digest is an identity record rather than an independent review of that code.
The first reference is DecoTengu 0.14.1, an experimental Python decompression library licensed GPL-3.0-or-later. The second is dive-deco 6.1.2, MIT licensed, pinned to commit 7f70ce0…7050fe. Versions, distribution/source hashes, fixture hashes, generator hashes and recorded toolchains are included. No reference code is linked into the Abyssi application by these development comparisons.
Pressure and input normalization
The DecoTengu contract uses the 1a first-compartment convention, 16 nitrogen/helium compartments, 1.01325 bar sea-level pressure, gravity 9.80665 m/s², water-vapour pressure 0.0627 bar and initial dry-air nitrogen fraction 0.7902. Saltwater density is 1025 kg/m³; the freshwater case specifies 1000 kg/m³. Descent is 18 m/min. The grid is 3 m, with a 3 m or explicitly selected 6 m final stop. Exact segment durations and gas fractions are preserved in the input manifest.
Runtime includes descent. A fully acclimatized 1500 m case supplies a normalized surface pressure; it checks calculation at that pressure, not the altitude formula, travel acclimatization or altitude safety. The model variant and pressure history must be specified before interpreting another planner as an equality oracle.
Three comparison families
- Original baseline: 12 entries. Complete constant-rate open-circuit schedules, with coefficient/tissue checkpoints, selected gas changes, freshwater, supplied altitude pressure and no-decompression cases. Original fixture values were preserved.
- Extended: 11 entries, nine newly distinct profiles. Four native 6 m final-stop cases, three complete multilevel cases, and two repetitive chains containing a first and second dive. In each chain the engine computes the first dive, carries its actual returned surfaced tissues forward, applies the surface interval, compares the post-interval starting tissues, and then compares the entire second ascent.
- Phase rates: 10 entries. An adapter changes DecoTengu’s single ascent rate at the decompression-phase boundary and rescales final transit for 18/10/9 m/min initial/decompression/final rates. Upstream still makes stop, gas, GF and tissue decisions. Before mixed-rate generation, all ten cases assert complete capture equality between the adapter at 9/9/9 and unmodified DecoTengu at 9 m/min. This is an adaptation of one reference, not native upstream support or a second independent implementation.
Extended multilevel and surface-interval inputs use the external model’s loading API and its native ascent scheduler. This involves pinned private reference APIs, which may change in future versions. The authors’ adapters import no Abyssi code or outputs. Their hashes are recorded, but their source is not distributed in this public release.
Endpoints and criteria
The full-schedule criteria are exact stop depth/duration sequences and stop oxygen fractions; absolute surfaced-runtime difference below 10−8 minutes; exact surface arrival; and an absolute error below 10−10 bar for each of the 16 nitrogen and helium final tissue pressures. Internal suites also check gas-switch depths, schedule continuity and selected warning behavior. These small tolerances compare compatible numerical calculations; they do not describe physiological precision.
Results: compatible reference contract
All 33 recorded entries meet the stated stop, gas, runtime and final-tissue criteria. The maximum observed runtime difference is 1.14e-13 minutes. Rows marked “reused” are antecedent dives already represented in the original family. The CSV includes pressure, water, GFs, gas fractions, interval and oxygen-cap metadata; the JSON files retain complete vectors and outputs.
| Entry | Family / profile | Max depth | Ascent m/min | Final stop | Max final-tissue error, bar |
|---|---|---|---|---|---|
| original-01 | square-30m-33minUniform-rate baseline | 30 m | 9/9/9 | 3 m | 6.82e-13 |
| original-02 | square-40m-20minUniform-rate baseline | 40 m | 9/9/9 | 3 m | 6.82e-13 |
| original-03 | square-30m-33minUniform-rate baseline | 30 m | 9/9/9 | 3 m | 6.82e-13 |
| original-04 | square-60m-25minUniform-rate baseline | 60 m | 9/9/9 | 3 m | 9.66e-13 |
| original-05 | square-30m-33minUniform-rate baseline | 30 m | 9/9/9 | 3 m | 3.95e-13 |
| original-06 | square-60m-25minUniform-rate baseline | 60 m | 9/9/9 | 3 m | 7.82e-13 |
| original-07 | square-40m-30minUniform-rate baseline | 40 m | 9/9/9 | 3 m | 9.38e-13 |
| original-08 | square-30m-33minUniform-rate baseline | 30 m | 9/9/9 | 3 m | 4.55e-13 |
| original-09 | square-30m-33minUniform-rate baseline | 30 m | 9/9/9 | 3 m | 3.98e-13 |
| original-10 | square-30m-35minUniform-rate baseline | 30 m | 9/9/9 | 3 m | 3.39e-13 |
| original-11 | square-12m-20minUniform-rate baseline | 12 m | 9/9/9 | 3 m | 1.14e-13 |
| original-12 | square-12m-40minUniform-rate baseline | 12 m | 9/9/9 | 3 m | 1.14e-13 |
| extended-01 | six-airGeometry, multilevel & repetitive | 30 m | 9/9/9 | 6 m | 5.68e-13 |
| extended-02 | six-air-gasesGeometry, multilevel & repetitive | 30 m | 9/9/9 | 6 m | 3.64e-13 |
| extended-03 | six-trimixGeometry, multilevel & repetitive | 60 m | 9/9/9 | 6 m | 7.96e-13 |
| extended-04 | six-trimix-gasesGeometry, multilevel & repetitive | 60 m | 9/9/9 | 6 m | 8.86e-13 |
| extended-05 | multilevel-airGeometry, multilevel & repetitive | 30 m | 9/9/9 | 3 m | 5.68e-13 |
| extended-06 | multilevel-nitroxGeometry, multilevel & repetitive | 18 m | 9/9/9 | 3 m | 5.68e-14 |
| extended-07 | multilevel-trimix-switchGeometry, multilevel & repetitive | 60 m | 9/9/9 | 3 m | 6.71e-13 |
| extended-08 | repetitive-shallow-firstGeometry, multilevel & repetitive · reused original-11 | 12 m | 9/9/9 | 3 m | 1.14e-13 |
| extended-09 | repetitive-shallow-secondGeometry, multilevel & repetitive | 30 m | 9/9/9 | 3 m | 4.55e-13 |
| extended-10 | repetitive-deco-firstGeometry, multilevel & repetitive · reused original-02 | 40 m | 9/9/9 | 3 m | 5.68e-13 |
| extended-11 | repetitive-deco-secondGeometry, multilevel & repetitive | 30 m | 9/9/9 | 3 m | 9.09e-13 |
| mixed-01 | air-30-33Phase-rate adapter | 30 m | 18/10/9 | 3 m | 9.09e-13 |
| mixed-02 | air-40-20Phase-rate adapter | 40 m | 18/10/9 | 3 m | 1.02e-12 |
| mixed-03 | air-low-gfPhase-rate adapter | 30 m | 18/10/9 | 3 m | 6.82e-13 |
| mixed-04 | trimix-60-25Phase-rate adapter | 60 m | 18/10/9 | 3 m | 7.96e-13 |
| mixed-05 | air-multigasPhase-rate adapter | 30 m | 18/10/9 | 3 m | 4.53e-13 |
| mixed-06 | trimix-multigasPhase-rate adapter | 60 m | 18/10/9 | 3 m | 5.60e-13 |
| mixed-07 | air-six-metre-multigasPhase-rate adapter | 30 m | 18/10/9 | 6 m | 4.52e-13 |
| mixed-08 | trimix-six-metre-multigasPhase-rate adapter | 60 m | 18/10/9 | 6 m | 5.00e-13 |
| mixed-09 | air-no-decoPhase-rate adapter | 12 m | 18/10/9 | 3 m | 5.68e-14 |
| mixed-10 | nitrox-no-decoPhase-rate adapter | 12 m | 18/10/9 | 3 m | 2.84e-14 |
Reported final-pressure errors aggregate 32 compartment values per entry. They are numerical differences, not estimates of diver risk. The sample is selected by the engine’s developers and is neither random nor exhaustive.
The authors’ full internal engine run passed 322 tests in 29 suites. That is a regression-suite count: some tests evaluate multiple profiles and others address unrelated boundaries or warnings. It is not 322 independent profiles, 322 external validations, an iOS-device acceptance result or clinical evidence. That private suite is not independently rerunnable from this public package.
A second implementation: agreement and disagreement
We built and executed unchanged upstream dive-deco at the pinned commit. All 16 × 6 coefficients agree exactly. For four profiles, every nitrogen and helium bottom-tissue pressure agrees within 10−10 bar after matching pressure history and initial tissues; maximum observed difference is 4.42 × 10−13 bar. These 128 bottom-tissue comparisons are separate from final surfaced-tissue comparisons and the internal test-suite count.
Rust uses depth/10 bar, integer-millibar surface pressure and 0.79 initial-air nitrogen. The comparison harness maps physical depth, grid spacing and rates and supplies matching initial tissues without modifying either engine. Both native and normalized outputs are retained. Normalization removes the tested bottom-tissue discrepancy; it does not remove the decompression-schedule differences.
| Synthetic profile | dive-deco, min | Pressure-matched Abyssi, min | Abyssi − reference, min |
|---|---|---|---|
| air-30-33 | 58.6833 | 56.3333 | -2.3500 |
| air-40-20 | 41.0111 | 40.4444 | -0.5667 |
| trimix-60-25 | 182.5667 | 175.6667 | -6.9000 |
| air-no-deco | 21.3333 | 21.3333 | ≈ 0 |
Three decompression profiles remain unequal; one no-decompression profile agrees to floating-point precision. Negative deltas mean shorter computed runtime in this selected software comparison. They do not demonstrate that either schedule is safer. Full stop sequences, coefficient arrays and matched bottom tissues are available in the differential results and native reference capture.
Source inspection identifies materially different policies. Rust caches an unrounded GF-low anchor, truncates sloped GF to integer percentages, and uses one-second/current-ceiling stop decisions. Abyssi anchors to rounded first-stop pressure, interpolates GF continuously, checks departure against the destination GF without future-transit off-gassing credit, and uses whole-minute holds with a default one-minute minimum. These differences exceed display rounding. We did not perform a one-factor ablation, so no exact share of each runtime delta is assigned to an individual policy. See the pinned GF implementation and scheduler.
Assessment of Abyssi’s algorithm
Strengths supported by this evidence: exact coefficient agreement; corroborated pressure-matched tissue mathematics from a second implementation; complete compatible-reference schedules extending beyond square dives to multilevel loading, gas switches, surfaced-state carryover and phase rates; retained negative findings; and stable original expectations rather than self-rebaselining.
Limits supported by the same evidence: exact agreement with a second complete scheduler is not established. Most exact schedule evidence comes from one experimental reference and its adapters. The chosen stop/GF contract still requires independent expert review. A selected finite matrix cannot exclude common-mode errors, omitted scenarios, adapter mistakes or model limitations, and private engine source prevents an independent full rerun of this release.
The next work is to obtain external review of the explicit scheduling contract, broaden independently generated boundary and repetitive/multigas cases, investigate additional geometry and transient-altitude behavior, and use controlled policy ablations if compatibility with another scheduler becomes a product requirement. This study does not justify changing coefficients or replacing external expectations with Abyssi outputs merely to achieve schedule equality.
Scope and exclusions
An earlier acquisition plan contained 145 unique case IDs; its former total of 148 was an arithmetic error. The selected independent matrix here is complete within its stated criteria, but the original acquisition campaign is not. There are 105 later uncaptured cases in each original commercial acquisition track. Independent normalized profiles do not fill those missing vendor rows. This public release makes no commercial ranking and includes no proprietary software internals or screenshots.
| Disposition | Case IDs |
|---|---|
| Supported OC features; exact vendor acquisition open | 121 |
| Supported repetitive features; exact vendor acquisition open | 4 |
| Feature-off baseline aliases | 2 |
| Transient-altitude research open | 1 |
| Unsupported finer stop-time resolution | 2 |
| Unsupported exact vendor oxygen-limit policy | 4 |
| Unsupported extended / accelerated-stop policies | 3 |
| Unsupported air-break policies | 3 |
| CCR excluded from the evaluated release scope | 5 |
| Total | 145 |
Unresolved areas include arbitrary stop geometry, broader repetitive and gas/exposure coverage, transient altitude, any future supported CCR, clinical outcomes and external approval of scheduling policy. The synthetic 6 m oxygen cases do not validate default oxygen-policy equality. The present results neither assess individual risk nor authorize operational reliance on a test vector.
Public data, methods and provenance
The versioned package contains numerical inputs and reference/actual outputs, all 33 rows, positive and negative second-reference results, original public Rust capture code, arithmetic audit code, source pins, toolchain records, file hashes and notices. It contains no private engine/application source, proprietary bundled code, commercial screenshots, session identifiers or credentials.
- Version 2026-11-14.v1 · data and methods ZIP
- Methods appendix and reproduction limits
- Structured study results · JSON / complete profile matrix · CSV
- Frozen implementation identities, pins and hashes / distributed-file SHA256 checksums
- Published-data arithmetic audit · Python / native Rust reference rerun
- Acknowledgments, licenses and file-specific notices
After extracting the package, python3 methods/verify_data.py verifies file hashes and arithmetic comparisons between the published records. It does not execute Abyssi. With Git and Rust/Cargo available, bash methods/reproduce_reference.sh fetches the exact pinned upstream commit and reruns only the native Rust reference. The private engine and undistributed adapters prevent a complete independent rerun of the 33 comparisons; this limitation is explicit rather than a claim of full reproducibility.
Frozen private-engine source digest771658613f38e2c6d5658de1ebb3dd93c045ad1ebdb1ac03c9579ed8d61f1c2d
Corrections or independent replication observations can be sent to hello@abyssi.io. A later evaluation should receive a new dated version and preserve the present dataset and negative results.