The previously private revision modules are now available as public downloads. This publishes source and documentation—not access to the running private service, private engineering corpus, conversations, machine credentials or employer data.
Download a revision
| Revision | What it contains | Download |
|---|---|---|
| 2.0 | Historical deployment-profile and boundary reference. Partial snapshot, not a complete installable application. | Rev 2.0 baseline record |
| 2.1 | AnyJev Routing Lab source, generic examples, frozen routing contract, page and tests. Alternatives, scores, chosen route and observable selection reasons are auditable. | Rev 2.1 module |
| 2.2 | Cumulative Routing Lab and Engineering Validity Gate modules, deterministic checks, engineering contracts, hypotheses, approval controls and synthetic regression tests. | Rev 2.2 cumulative source |
| 2.3 | Cumulative modules plus controlled Physics-AI witnesses, isolated DoMINO reference adapter, provenance, applicability checks and tests. Research-only model; not validated engineering inference. | Rev 2.3 cumulative source |
Latest source: Rev 2.3 Transolver addendum (cumulative; preserves the original Rev 2.3 archive). These are source module packages, not complete one-click installations of the private Aero application. The application shell, personal integrations, solver installations and model weights are excluded. Public availability is not evidence of engineering qualification.
Pull and verify
Latest research update: Transolver reference execution and checkpoint audit. The NVIDIA adapter executed; checkpoint-specific accuracy remains unestablished. New archive identity · Manifest · Verification.
Rev 2.3: Release and archive hash · SHA-256 manifest · Verification · Implementation report · Reference results.
Historical Rev 2.0–2.2 archives and their original manifests remain unchanged:
Release index and archive hashes · SHA-256 manifest · Verification record · README
Download the desired ZIP without a login, verify its SHA-256, and extract into a new directory. Each archive includes an interior source manifest with original and published file hashes. Personal edge URLs are replaced with an example origin; text line endings are normalized. Runtime source is left unchanged.
In an isolated Python 3.12 environment:
python -m pip install -r requirements-test.txt
python -m pytest Aero/tests -qRev 2.0 contains no test suite. Rev 2.1–2.3 software tests use controlled synthetic evidence and model/solver test doubles. Model inference requires its separately configured environment and checkpoint. Conventional CFD/FEA execution requires its existing worker integration. Do not mistake passing software tests for validation of a physical model or a real engineering result.
Release verification: 16 extracted-archive tests passed for Rev 2.1; 54 for Rev 2.2; seven publication/custody checks passed. The initial attempt used a Python environment without Flask and failed collection; the successful runs used the existing Aero Python environment. Test record.
What changed, and why
Rev 2.0 separated the personal local Aero agent from the corpus-free portable deployment. The original historical snapshot was scoped, not a full runtime backup.
Rev 2.1 made model-selected workflow choices inspectable. It retains what the choices were, the model’s scores, the selected route and the observable decision rule. It does not claim to expose hidden model reasoning. Its small-model pilot does not authorize autonomous physical changes or engineering acceptance.
Rev 2.2 added an engineering investigator’s control loop. A frozen contract and pre-run analytical baseline challenge numerical results; numerical and physical validity stay separate. Discrepancies produce competing hypotheses and bounded experiments. Physical changes still require approval. Missing evidence cannot be inferred into existence, and earlier failed qualifications remain failed.
Rev 2.3 adds an optional supporting-witness capability: neural physics models with applicability checks, checkpoint provenance, isolated execution and explicit limits on authority. A surrogate cannot grade its own work, replace a required solver run or turn agreement into validation. The existing private application remains Rev 2.2-local; this publication does not expose or restart it.
Rev 2.3: what actually ran
The pinned DoMINO automotive surface checkpoint ran on a 753,234-cell DrivAerML reference geometry through PhysicsNeMo in an isolated Linux/CUDA worker on an RTX 3090. The adapter invocation took 119.478 seconds, including setup; this is a single reference execution, not a hardware benchmark. Inputs and runtime remain on the approved D: storage, outside Aero’s main Python environment.
| Reference coefficient | Prediction | Reference | Relative discrepancy |
|---|---|---|---|
| Drag Cd | 0.306863 | 0.303512 | 1.10% |
| Side Cs | 0.035471 | 0.047668 | 25.59% |
| Lift Cl | 0.433089 | 0.067728 | 539.45% |
Pressure, shear and force exactly reproduced the first reference execution. Sensitivity gradients did not: maximum absolute difference was 0.0007534 in the upstream sensitivity units. No strict-determinism claim is made.
Independent force integration agreed within 0.0000842 N. A constant pressure-offset check did not explain the lift discrepancy. Two upstream preprocessing paths use different COM-offset scaling; that is an unresolved compatibility hypothesis, not a proven cause. Native CFD/STL field correspondence is not established.
The correct outcome is REFERENCE_INFERENCE_COMPLETED_NOT_VALIDATED. Applicability remains NOT_ESTABLISHED, authority RESEARCH_ONLY, and shared OpenFOAM training lineage means LOW independence from OpenFOAM. No neural result can satisfy a required engineering qualification gate.
The software tests and actual model execution are separate evidence. See the Rev 2.3 test record. At the original release, Transolver remained a stub; PDEBench/RealPDEBench execution, PFEM warm starts, autonomous surrogate screening and training are not implemented. This is a bounded source release, not a claim that every planned Physics-AI feature is done.
Version boundaries
September 24 workflow addendum
The development follow-up adds identity-aligned field benchmarking, review-only candidate ranking and an optional-experiment hook in the existing campaign loop. A frozen contract can allow a traceable, applicable surrogate witness to defer a specifically optional numerical experiment. It cannot eliminate a physical hypothesis, waive a required solver run or override an engineering gate.
113 local tests and 110 extracted-package tests pass. These test software contracts, not neural accuracy or demonstrated CFD savings. The new harness also scores retained DoMINO coefficients descriptively, with no post-hoc acceptance limits. Native-field correspondence and model validation remain unestablished.
Implementation and limitations · Test evidence · Retained-model comparison · Archive identity · SHA-256 manifest · Verification.
The original multi-model milestone is not fully complete. At the workflow-addendum milestone, Transolver was blocked pending weights, reference data and usage review. PDEBench/RealPDEBench dataset reproduction and PFEM execution are still future work. No private service was restarted, no additional model downloaded and no CFD run launched. This addendum does not silently promote the private application or rewrite any historical release.
September 24 Transolver addendum
The NVIDIA Transolver adapter now executes a pinned 200,000-cell automotive reference, retaining exact native sample IDs. 125 extracted-package tests pass. The 2.404-second inference phase is one execution, not a performance benchmark. Pressure discrepancy is 56.2% after dimensional conversion; shear discrepancy is 82.8% with a conditional units caveat. Checkpoint training/preprocessing compatibility and held-out status remain NOT_ESTABLISHED. Authority stays RESEARCH_ONLY.
The source audit found eight-versus-twenty-layer documentation disagreement and different legacy/modern coordinate conventions. No inputs were tuned to improve the answer. A later NVIDIA benchmark split lists run 1 in its 436-row training file and not in its 48-row validation file, but that split was committed after the checkpoint release and is not linked to these weights. The checkpoint recipe and run 1’s held-out status remain NOT_ESTABLISHED. The questions and evidence matrix are published as an on-site research update; no GitHub issue or external inquiry will be submitted. No sign-in is required to read the report or download the source. Earlier archives and historical results remain unchanged; the current website-only direction supersedes the historical inquiry-submission instructions.
Full report, results, source download and next-step status.
These are Aero source revisions, not the public presentation’s revision numbers, the V2/V2.1 convergence contracts, or the benchmark report’s release number. The running private application retains authentication and its existing release until a new runtime revision is deliberately tested and deployed.
Aero is a personal project. Public release grants no employer-data ingestion authority or employer endorsement. Upstream models and frameworks retain their own licenses; no model weights are redistributed here.