Physical models.
Traceable evidence.

• Python 3.11+, NumPy core, import equipment_diagnostics, MIT original code; research Alpha 0.17.1 is available as source and a Zenodo archive. PyPI 0.17.1 and the GitHub release are now public; upstream records a fresh installed-package smoke check without new measured fault validation. Raw data, external engines and fitted corpus profiles are separate.
• Evidence and limits: 286 software/physics tests and seven scoped object integrations are reported upstream, not rerun here. An alarm is neither fault probability nor confirmed root cause. Reused heat-pump gains depend on a 0.3 K between-session difference; synthetic ±0.3 K normal offsets caused 99.28% alarms. Fault specificity, independent equipment transfer and several modules’ measured fault truth remain unresolved; no universal accuracy or certified safety operation is claimed.

Zenodo 0.17.1

• Python 3.11+, import plantdesign, GPL-3.0-only with required DWSIM-derived model notices. Version 0.11.0 is public; ordinary calculators require no third-party runtime dependency.
• Recorded verification reports 364 tests on Windows/Ubuntu Python 3.11/3.12 and installed-package checks across all 67 methods. These were not rerun here. The 44 organizing families are not additional calculators. Prices, coefficients and factors in examples are assumed inputs; source-qualified project data, detailed physical design and complete environmental boundaries remain necessary. No design approval, full LCA, permitting or investment/site qualification is established.

PyPI 0.11.0 · GitHub · Zenodo DOI

• Python 3.11+, MIT original code; Alpha 0.1.0a5 is publicly available. Raw data and trained rig-specific profiles are acquired separately; standards limits must be supplied by the caller.
• Evidence and limits: the 99.82% result was fault-only on the same apparatus/bearing/severities, with no new healthy test. New-bearing and independent-rig transfer failed; the strict API held all 580 new windows. This does not establish a zero false-alarm rate, field diagnosis, remaining-life prediction or automatic shutdown authority. Public software checks are distinct from the historical private 547-check count and were not rerun here.

PyPI 0.1.0a5 · GitHub · Zenodo 0.1.0a5 - 0.1.0a5 update: stricter profile/numeric/decision contracts, acquisition quality and scoped normal calibration, measured-speed angle experiments, bounded queues, detached stage/sink results, an optional pinned process-monitoring bridge, hashed profiles, power-conserving PSD history and reviewed caller rules. Life estimation now requires explicit experimental opt-in. Upstream reports 224 selected installed-wheel tests; no new measured classification or field-life qualification follows. A 48 kHz/four-channel burst still missed its 250 ms latency budget 17 times. See the release record.

• Python 3.12+, MIT original code, import pipeline_integrity. PyPI 0.1.0a1 and Zenodo archive are public; external engines/data retain separate requirements. The catalog maps 223 methods to 304 APIs, not 304 independently validated solvers.
• Recorded evidence: bounded independent physical profiles cover 198/223 methods; 25 remain pending and four empirical acceptance failures remain. The portable installed-wheel subset reports 276 passes. Full-style lint findings remain unresolved. These upstream checks were not rerun here and do not establish field safety or standards certification.

• Capabilities: thermodynamics, equipment and finite transfer, hydride/adsorption storage, electrochemistry and separation, material permeation/diffusion, release histories, and facility/safety interfaces. Optional adapters connect specialist engines such as CoolProp, Cantera, LH2DT, PSA-Dyna and HyRAM+; native dependencies and their licences must be acquired separately.
• Architecture and use: ordinary Python objects, functions and arrays let researchers replace parameters and providers without a web service. SI units, material/phase identity, energy references and aligned clocks are explicit; provider failures do not silently substitute different physics. The catalog has 82 entries and verification campaigns cover 78 scoped topic families, neither a count of independent equations nor proof of complete qualification.
• Recorded evidence: selected input, formula, boundary, regression and connection campaigns are documented. Four conditional measured-data holdouts cover effective permeance, normalized hydride reaction curves, PEM polarization voltage and fixed-coefficient nozzle flow. They use different metrics and cannot form one accuracy score; the nozzle comparison reports about 29.8% mean absolute relative error. Full physical qualification remains unestablished, with rejection-only and blocked native connections retained explicitly. These calculations were not rerun for this portfolio.
• Release and reuse: 0.2.0 Alpha is published on GitHub with wheel/source assets; PyPI 0.2.0 and the Zenodo DOI software archive are now public. Registry publication and DOI metadata were checked; installation smoke tests are reported upstream and were not rerun for this profile. Python 3.12+ and Apache-2.0 original code; external observations, engines and reference documents retain separate terms and are excluded from the distributions. See the quickstart, comparison report and publication status.

• Published extensions: 0.2.0 extensions add indoor accumulation and conditional ignition-delay peaks, FA/DDT screening, mechanical expansion and weak-wave loads, linear structural response, thermal cycles/exergy, solid redox equilibrium, electrode metrics and a coupled passive autocatalytic recombiner (PAR). These are included in the published 0.2.0 release; optional providers remain separately required. Upstream reports 139 distinct focused checks completed across runtime profiles, including installed-wheel checks; a separate broader replay retains 61 environment/artifact failures. These are scoped software checks, not new experimental or facility qualification, and were not rerun here.

• Capabilities: storage mass, energy and wall-temperature dynamics; a three-stage compressor with intercooling; three-bank cascade dispatch; valves, precooling and hose inventory; normal control and independent safety logic.
• Interface and integration: FastAPI simulation jobs, 3D monitoring, trends and incident exercises; outputs for 82 HAZOP sensor tags; SAGA-PY explanation APIs and a HyRAM+ consequence bridge.
• Evidence boundary: component and public-data evaluations are documented separately. Gas-detector tags include virtual zone proxies. This is a modelling and training prototype, not a certified physical controller or a claim of full-loop field validation.

• Capabilities: storage tanks, cryogenic heat transfer, valves, pipes, vaporizers, reliquefiers, pumps, compressors, vent stacks and instrumentation.
• Architecture: property packages, two-port components, links, networks and a sensor/HART observation layer; public-reference benchmarks, observation-loading and blind-validation documentation.
• Use: engineering studies, digital-twin prototyping and model-in-the-loop work. The distribution is lh2dt_models, with lh2dt and lh2dt_models import namespaces. Validation remains specific to each component and benchmark.

• Capabilities: scenario orchestration and a traceable ledger for operating mode, input origin, frequency evidence, release history, weather, ignition, population, protection and consequence output.
• Architecture: process and consequence models remain independent providers. Measured and predicted states are distinguished; missing accident-source information is not silently inferred from normal operating flow.
• Use and limits: the prism-risk package provides the prism CLI, standard-mode templates and public LH₂ backend documentation. A completed calculation chain is not proof of a complete facility QRA or regulatory certification.

• Calculation: pressurized gas release, delayed ignition, upstream isolation, residual downstream release, and point-source radiation within a declared assessment horizon.
• Outputs: an offline dashboard, time histories, raw sensitivity samples, convergence diagnostics, input/dependency hashes, and PNG/SVG figures.
• Boundary: one supported substance and scenario in v0.1.0. Conditional harm probabilities and sensitivity indices do not establish annual facility risk or experimental validation. The exported dashboard displays results; inputs are changed through JSON and the CLI.

• Physics and numerics: finite-volume packed beds, Ergun pressure-flow closure, competitive dual-site Langmuir equilibrium and LDF kinetics, gas/solid/wall energy balances, compressible valves, and optional Peng–Robinson properties. Multiple beds and cycle schedules, stiff integration, sparse Jacobian structure and checkpoint/restart support numerical research.
• Evidence: the public validation summary reports that the Wang single-bed breakthrough figure comparison passes the declared composition, temperature, mesh, time-step and conservation gates. This does not validate cyclic four-bed VPSA or an industrial PSA design; the S-015 candidate remains a numerical design screen, and the Boer biogas cycle track remains blocked by apparatus-response identification.
• Process integration: a guarded DWSIM contract carries cycle-average FEED, PRODUCT and TAIL streams with thermodynamic state, enthalpy and provenance. The interface is implemented, but automatic DWSIM release and industrial-design qualification remain closed. A single completed cycle is not cyclic steady state.
• Research continuity: the research handoff, machine-readable research state and S-015 applicability supplement distinguish completed work, failed approaches and open validation/design gates.
• Relationship: a reusable dynamic separation model, distinct from the psa-qae-resource-audit archive of a particular engineering probability event and hardware probe. See the physics, validation and DWSIM bridge documentation.

• Capabilities: rules, statistical and subspace methods, classical machine learning, temporal/graph/neural models and CoolProp residuals share an explicit signal contract. Causal windows, independent threshold calibration, validity masks and elapsed-time alarm policies preserve the meaning of an output.
• Evidence: the release documents 72 named configurations, six official UCI archives, 117 passing local tests and 31 exploratory benchmark runs. Configurations are not 72 independently invented algorithms, and dataset-specific scores are not a universal detector ranking. See the benchmark results.
• Integration and limits: adapters cover station, LH2DT, PRISM, SAGA and PSA boundaries, with actual object checks and pinned schemas distinguished from live-service tests. Invalid observations are not healthy operation; an anomaly is not a proven root cause or a calibrated accident frequency. Integration evidence records scope and known mismatches.

• Scope: 23 registered calculations cover source/inventory screening, passive Gaussian dispersion, fire/explosion effects, exposure/dose/probit and conditional risk. Optional native SLABx histories and a shared coupled 1D layer are separate routes; the latter is isothermal, fixed-width and vertically mixed, not a validated thermal or two-layer replacement.
• Reproducibility: calculation receipts preserve context, evidence, source hashes and dependency traces. Synthetic examples and validation records separate software conservation/refinement checks from field validation. See the manual.
• Availability: Zenodo 0.1.0 and PyPI 0.1.0 are published; the upstream record documents fresh-environment installation checks. The pinned optional SLABx 1.0.8 engine is not currently available from PyPI; the dense extra cannot be installed from PyPI alone. Native-engine skips in public CI do not reproduce the local native tests.
• Independence and rights: original work is Apache-2.0. This is not an official or endorsed CCPS/AIChE product; the legacy ccps_consequence import is a compatibility identifier. Purchased standards, books and external engine sources are excluded.

• Capabilities: document indexing and retrieval, evidence validation, term aliases, and Korean/English operation. Korean regulatory retrieval and English NREL/HIAD libraries use distinct language paths.
• Integration: provides an advisory explanation layer for calculated process and sensor context from the independent hydrogen-station simulator.
• Requirements and limits: grounded answers require a lawful local document corpus and provider credentials. The public release excludes private PDFs and indexes. Outputs support analysis and training; they do not control physical equipment or replace official regulations.

• Stack: Java 21, Spring Boot, Spring AI, Apache Tika, Lucene/Nori Korean retrieval, PostgreSQL/pgvector and Ollama, with HAZOP/LPG modules and a web interface.
• Contents: source, tests, Gradle wrapper, dictionaries and a PDF.js viewer. Private documents, databases and indexes are excluded.
• Relationship: its runtime and APIs differ from SAGA-PY; it is not a drop-in replacement or a validated live-station safety controller.

• Previously QESx; now published as PlumeSwift 0.1.0a4 on PyPI, GitHub and Zenodo. Python 3.11+, module/CLI plumeswift. QES-derived code is GPL-3.0-only; glibc-derived numerical components retain LGPL-2.1-or-later, with upstream and third-party notices. Acceleration, thermodynamics, GIS/GPU and visualization are optional and require compatible environments.
• Includes a packaged synthetic LH₂ example and optional offline 3D concentration views. The FFI workflow reads measurements locally from a user-supplied official report; the report and measured sensor records are outside the software distribution.
• Recorded evidence: seven FFI conditions and 210 sensors showed agreement among sequential and parallel Python concentration/mass-balance paths. Measured-concentration RMSE was 1.9843 percentage points, with only 2/9 sensors whose measured means were at least 4 vol% captured. These calculations were not rerun here. Complete upstream feature parity remains unfinished, and the reported timings used differing output workloads.

PyPI 0.1.0a4 · GitHub 0.1.0a4 · Zenodo DOI

• Capabilities: evaporating pools, horizontal and rising vertical jets, instantaneous releases, and the original plume-to-puff transition.
• Extensions: optional rainout, finite-rate evaporation, added mass, ground heat transfer and real-fluid properties are separated from the reproduction path.
• Evidence: compared with the original Fortran and evaluated against 38 field and wind-tunnel trials. Some observation data must be obtained separately; applicability follows the documented physical assumptions and benchmark conditions.

• Capabilities: cryogenic source handoff, thermodynamic extensions, ground interaction and routing checks; documents solid–vapour water treatment below the triple point and fusion-enthalpy corrections.
• Evidence: the current paper configuration uses seven FFI tests and 210 records. Those records are not 210 independent experiments, and reproducing them requires a separate research workflow and inputs.
• Relationship: complements the base slabx model and offers an optional LH2PoolX adapter. Package, archive and research configurations must be distinguished when reproducing results.

• Capabilities: steady/transient ground-level releases, pressurised jets, input-deck readers and jet-touchdown handoff to the ground model.
• Source contract: SourceLedger and SourceState preserve resolved source-plane mass flows, area and time history. Unresolved liquid fractions are rejected; upstream flash, deposition and pool physics must be supplied separately.
• Evidence and scope: reproduction checks use the EPA reference distribution and benchmark listings. DEGALI research extensions are excluded; an optional pool adapter does not change the original dispersion physics.

• Capabilities: steady/transient dispersion, receptor dose, jet-to-ground workflows, flashing, buoyant trajectories, air condensation/freezing, component enthalpy and ground interaction.
• Evidence: EPA/Fortran checks and PRESLHY/FFI-DNV assessments are recorded in separate evidence lanes. Peak and time-mean concentration operators are not merged into one accuracy score.
• Status: alpha research software. Experimental 3D and obstacle-related numerical work has explicit conservation, closure and convergence limits; numerical checks do not constitute independent field validation or regulatory qualification.

• Two routes: a release/pressure/declared-deposition calculation, or conditional evaporation from an observed or declared pool footprint.
• Physics: isenthalpic flash, liquid reaching the ground, conduction-limited evaporation with a critical-heat-flux cap, and an equilibrium-area balance.
• Boundary: does not predict unknown impact footprints, deposition fractions or transient pool inventory. Source assumptions, route, version and status must accompany the result.

• Capabilities: axisymmetric shallow-layer spreading, wet-contact evaporation, liquid inventory and an auditable mass balance, driven by an explicit DeclaredInflow history.
• Evidence: restricted Stage C component checks include conservation/refinement studies and JUEL/HSE comparisons. The full HSE expansion/retraction trajectory is not claimed as reproduced.
• Boundary: jet impact, splashing, slopes, drains, barriers, complex terrain and atmospheric dispersion are outside the qualified scope. The historical quasi-steady path is retained; substituting dynamic results into earlier FFI work requires a separate coupling study.

• LPG workflow: infer heat input from closed-vessel pressure rise before relief, estimate effective flame-contact angle/height/area, and calculate heat flux using pressure and wall-temperature histories.
• LH₂ workflow: ortho–para thermodynamics, state-point and pathwise composition screens, published conversion kinetics and a CoolProp HEOS nozzle closure.
• Evidence: original validation sources and comparison limits are documented. Experimental and validation datasets are not redistributed in the repository or package.

• Calculation chain: heat ingress → self-pressurisation → binary PSV opening/closing and vent history → flash/outlet thermal state → TNO-aligned jet-fire and dispersion routing indicators.
• Public tools: load published results, calculate flash-equivalent mass and conditional risk area, classify dispersion routes and regenerate five manuscript figures from saved ledgers.
• Boundary: the complete upstream tank and internal TNO solvers are not redistributed. This is a citable research/post-processing record, not a design-certified PSV sizing tool.

• Capabilities: uncertain damage-proxy comparisons for ten materials, explicitly declared material/weld/environment/inspection multipliers, and allocation of one initiating frequency across mutually exclusive protection paths.
• Reproduction: seeded NumPy sampling, evidence qualifications and source identifiers, existing API tests, and an archived seven-output Sobol summary.
• Boundary: material rankings are conditional proxies, not observed annual leak rates. Calibrated frequencies need exposure/failure/inspection evidence; the package does not contain a complete ammonia source, dispersion, consequence, or facility-QRA solver.

• Python 3.11+, MIT original code, import matintegrity; source metadata identifies research Alpha 0.1.0a1. PyPI 0.1.0a1, the GitHub release and Zenodo archive are now public; registry availability and DOI metadata were checked. Optional scientific/property providers and external data require separate installation and rights.
• Recorded limits: 757 registered APIs include data/workflow tools, with 63 core calculation groups; 230/232 audited core APIs computed and two await external evidence. Installed-wheel numerical checks report 59 passes; 885 Ruff findings and one full-example-coverage check remain unresolved. Counts overlap and are not independent physical validations. No automatic design approval, failure probability from deterministic damage, or field qualification is claimed; checks were not rerun here.

• Contents: 32 Quantum ESPRESSO input/output pairs, archived MACE-OMat-0 and CHGNet arrays, analysis tables, figure scripts and the exact pseudopotentials used.
• Reproduction: figures use archived tables and arrays without requiring external model weights. Registries and checksums identify model and pseudopotential provenance.
• Limit: the dedicated CHGNet NEB generation script was not retained. Archived paths, barriers, cell definitions and method records support inspection, but figure regeneration is distinct from rerunning every original calculation.

• Capabilities: exact weighted probabilities, classical Monte Carlo and ideal maximum-likelihood QAE comparisons at equivalent oracle-query budgets; Qiskit state preparation, Grover checks and Aer simulation.
• Reproduction: a NumPy-only core, CSV/CLI interface and optional circuit/IBM dependencies let examples run without a QPU.
• Boundary: the probability model, estimator, circuit validation and opt-in hardware execution remain separate. Ideal-estimator results are not a demonstration of hardware quantum advantage.

• Engineering evidence: corrected process topology, public experimental holdouts, independent integrator checks and DWSIM cross-checks; process performance and probability-discretisation errors are reported separately.
• Quantum audit: resource accounting for the final ten-bit arithmetic oracle and an IBM Kingston test of a seven-qubit Boolean circuit representing the same event at two-bit resolution.
• Inspection and artifacts: begin with EXECUTION_BOUNDARIES.md and the consolidated evidence. Inspecting archived counts does not require a new QPU submission; historical scripts can retain workstation-specific dependencies.
• Boundary: the hardware test covers all 64 basis inputs, not coherent QAE or the full arithmetic oracle. Manuscripts, results, circuits, raw counts and checksums are provided; synchronized validation of the final integrated apparatus remains outstanding.

• Physics and engineering: dipolar, bond-dependent and interlayer interactions, sample shape, public experimental evidence, multimodel analysis and N=12/N=18 finite-cluster transfer checks.
• Quantum methods: low-energy entropy estimation, physics-informed state preparation and a complete six-qubit IBM Kingston method benchmark.
• Evidence boundary: hardware resolves the entropy-correction sign but fails quantitative-magnitude qualification. Finite-cluster results are not a bulk solution and no quantum advantage is claimed. Source manifests, derived data, Python/Octave checks, figures and manuscripts are public.

• Capabilities: task DAGs, Bayesian observed-count updates, SPAR-H/HEART/CREAM/IDHEAS-ECA references, censored response models, recovery/dependency, repeated-event-safe FTA, conditional ETA, small Bayesian networks, bounded Petri nets and optional sensitivity tools.
• Evidence: the retained evidence catalog provides 1,334 records from seven numeric sources, including 578 observed context-group count records. The total includes method components, multipliers, intervals and rules; it is not 1,334 independent measurements or complete human-error probabilities. See the evidence guide.
• Version 0.5.0: adds original field-collection forms, independent demand-register reconciliation, unknown outcomes and frozen HEP forecasts with group-leakage and calibration checks. Twelve operational extensions cover executable actions, communication/permits, causal/risk workflows, coupled dynamics, progress/deadlines and assurance lifecycle. See the field collection manual and operational guide.
• Additional benchmark: 5,617 measured clinical specimen opportunities are a separate success/nonconformance aggregate, not additions to the HEP catalog. Individual human attribution and industrial transfer remain unqualified, and an unfavorable temporal calibration result is retained in the denominator validation record.
• Application boundary: SACADA evidence concerns nuclear training simulators. Chemical-plant field HEP calibration and observed annual error-frequency series are not claimed; calendar frequencies need compatible exposure and explicit transfer review. Integration traces do not establish that every receiving application has been implemented. Software is MIT; catalog evidence retains source-specific terms.

• Capabilities: typed calculation contracts cover damage, failure frequency, inspection, release, effects, Level 1/2 consequence composition, financial risk and planning. Optional process-isolated SLABx/SLABx-LH2 adapters retain source fingerprints and explicit capability status; API 520 sizing utilities occupy a separate companion namespace.
• Software evidence: the public release reports 509 typed calculation registrations, 861 source-section records, direct-test traceability and local/Nurion verification receipts. These are software verification records, not independent experimental evaluation or third-party API conformity certification. See the module catalog and development status.
• Standards and use: no API affiliation or endorsement is asserted. Standards and purchased source tables are excluded and must be obtained under the user's licence. Original work is Apache-2.0; outputs are not a fitness-for-service decision, inspection release or operating authorization.

• Workflow: feed characterization → direct 34-state ADM1 integration → gas/product/material ledger → CAPEX/OPEX, NPV/IRR/LCOB and partial GHG screen → grid/Pareto, robustness and sensitivity records with hashes. Original equation/input sources are preserved, while missing analysis interfaces were reconstructed from research notes.
• Recorded evidence: five light execution/bookkeeping tests and a small four-point demonstration grid are documented. These do not establish experimental validity, complete historical numerical reproduction or global optimization. See the reconstruction record and test report.
• Interpretation: environmental outputs are a partial historical GHG screen, not a full LCA or verified product footprint; the archived fertilizer-credit assumption can overstate credits. Descriptive sampled ranks are not the original converged global sensitivity analysis. DWSIM/Aspen and downstream hydrogen optimization are outside this archive.