COMPLETED WORKFLOWS
2 bounded passes
Airfoil v7 and internal-flow v2 produced completed solver runs and verified evidence manifests. Their observers kept design readiness at no.
AERO / LOCAL-MODEL OPERATOR BENCHMARK
Can a local model pilot a guarded CFD workflow?
Two public OpenFOAM seeds with retained proposal failures. These trials test case handoff and evidence reading, not aerodynamic accuracy or design readiness.
7B LOCAL RUN
Solver completed
The 7B model authored the accepted proposal and read the terminal state. Its strict observer grade failed on evidence-key aliases. A Codex operator prepared and started the run.
Retained attempts
Airfoil v1-v7 are iterative revisions, not independent samples. The 14B airfoil rubric and the 89/100 internal-flow review use different scoring methods and cannot be averaged.
| Trial | Model | Outcome | Solver? | Finding |
|---|---|---|---|---|
| Airfoil v1 | 7B | Proposal rejected | No | Conservation and wall-treatment criteria missing. |
| Airfoil v2 | 7B | Proposal rejected | No | Reference, mesh-independence and model-fidelity criteria missing. |
| Airfoil v3 | 7B | Proposal rejected | No | Lift, drag, mesh-independence and model-fidelity criteria missing. |
| Airfoil v4 | 7B | Schema rejected | No | Invalid assumptions field. |
| Airfoil v5 | 14B | Proposal rejected | No | Unresolved setup questions. |
| Airfoil v6 | 14B | Freeze rejected | No | Wall-y-plus request unsupported for the template. |
| Airfoil v7 | 14B | Workflow pass | Yes | Solver and manifest verified; strict observer passed; design-ready no. Recorded workflow rubric: 100/100. |
| Internal flow 7B | 7B | Solver completed; strict observer failed | Yes | Model-authored proposal and numerical pass; two evidence-key aliases. Operator-assisted setup and start. |
| Internal flow v1 | 14B | Schema rejected | No | Cross-backend fea_case field in an OpenFOAM proposal. |
| Internal flow v2 | 14B | Workflow pass | Yes | Corrected proposal, numerical pass, verified manifest, strict observer pass; design-ready no. Independent workflow review: 89/100. |
The engineering boundary
Aero's typed harness and operator approval decide what may run; OpenFOAM and retained evidence decide what happened. Numerical convergence is not validation. The successful internal-flow run still lacks an independent matched reference, demonstrated model fidelity, and mesh independence.
Was the 7B run fully solo?
No. The local model wrote the accepted proposal and assessed the solver evidence. Earlier proposal failures led to a clearer reusable example, and a Codex operator invoked the staged handoff and solver start. The receipt does not name that operator's model, so Luna's involvement in this specific run is unverified.
Read the complete report
The report records each retained attempt, the 7B model's role, the operator's role, strict observer limits, and audit hashes.