HelloWorldLTY
Add FDA nozzle Re=500 intermediate design brief
8af9ada
|
Raw
History Blame Contribute Delete
4.13 kB
---
pretty_name: OpenFOAM OSWorld Benchmark
language:
- en
tags:
- openfoam
- paraview
- osworld
- gui-agent
- scientific-computing
- benchmark
license: other
---
# OpenFOAM OSWorld Benchmark
This gated dataset contains OSWorld-style, long-horizon OpenFOAM and ParaView
tasks. Each release task combines deterministic CFD evidence and engineering
decision criteria with a substantive native ParaView state and scientific image
deliverable. CLI assistance is allowed, but the final GUI evidence is graded.
## Current release tasks
The current five-task set is:
1. `of_battery_coldplate_release_review_v4`
2. `of_fda_nozzle_re5000_validation_review_v4`
3. `of_hvac_thermal_comfort_acceptance_v1`
4. `of_porous_barrier_transient_assessment_v1`
5. `of_edge_compute_thermal_qualification_v1`
Task definitions are in `harness/tasks/`; matching guest assets are in
`harness/assets/`. The task JSON contains the locked SHA-256 for its asset.
The other task JSON and asset versions are retained as reproducible benchmark
history. They are not current release candidates. The `rejected/` directory is
an explicitly withdrawn evaluator experiment and must not be used as a task.
## Repository layout
- `harness/tasks/`: OSWorld-style task JSON files.
- `harness/assets/`: guest-visible task assets.
- `intermediate_design/`: bilingual review-stage scenario and GT-policy
materials; these are maintainer documents, not agent-visible task inputs.
- `reference_private/`: answer-side GT, deterministic oracles, source snapshots,
asset builders, and task generators. Do not expose this directory to an agent
during an evaluation run.
- `audit.py`: rebuild, evaluator-replay, leakage, PNG-contract, and lint audit.
- `HANDOFF.md`: detailed version history and harness handoff notes.
- `rejected/`: withdrawn research artifacts, clearly separated from releases.
## Evaluation boundary
During a benchmark run, the agent should receive only the selected task JSON,
its matching asset, and the configured VM snapshot. `reference_private/` is
included in this repository because it is needed by benchmark maintainers to
audit exact answers and regenerate artifacts; it is not guest input.
The instructions shown to agents are intentionally concise. Physical
definitions, acceptance thresholds, and blank deliverable templates are stored
inside each task asset. Method tutorials, command sequences, and answer values
are excluded from guest-visible assets.
## Answer-isolated public trial
`public_trial/` is the safe solver-facing distribution. It contains sanitized
task JSON files and the complete matching guest assets, but no evaluator,
expected values, private GT, oracle, score code, or answer-side source snapshot.
The directory is also available as
`public_trial/Openfoam-Osworld_Benchmark-public-trial.tgz` for direct
distribution. See `public_trial/manifest.json` for locked file hashes and guest
deliverables.
## Current edge-compute correction
`of_edge_compute_thermal_qualification_v1` uses
`net_exhaust_mass_flow_kg_per_s`, defined as the signed face-flux sum with net
outward flow positive. Its state evaluator accepts native ParaView files where
`ServerManagerState` is bare or wrapped by `GenericParaViewApplication` or
`ParaView`.
## Validation
From the repository root in the original Orion workspace:
```bash
/diskarray/home/tl688/orion/envs/orion/bin/python codex/openfoam/claude_handoff/audit.py
```
The audit verifies task/evaluator array consistency, generated JSON parity,
asset hashes, deterministic GT replay, perfect-submission scoring, neutral
identifiers, tutorial and answer leakage, native-state contracts, full PNG
decoding, adversarial blank/missing-panel rejection, and the harness's actual
`score.py` lint.
## Licensing note
The benchmark bundle contains derived and frozen evidence from multiple
scientific sources and OpenFOAM tutorial material. Consult the provenance in the
private builders and the applicable upstream licenses before redistribution or
commercial use. The repository metadata therefore uses `license: other` rather
than asserting a single license for every artifact.