Choose a benchmark slice, inspect its size, and copy the exact dataset-generation code. Ground truth is generated once; observation masks are deterministic views applied afterward.
Optional: future model campaign planning
Future model campaign planning
Primary matched-mask comparison: every normal learned baseline gets an independent checkpoint for each PDE and each of the nine observation masks. The model trained on random 3% is reused only for the separate mask-transfer/OOD table; it never replaces matched-mask training in the primary IID table.
Open the complete observation-training protocol
Campaign preset
| Method row | Fit policy | Seeds | Observations | Preparation | Evaluations | Builder status |
|---|
2. Copy the generated code
The dataset YAML is the executable source of truth. Use the environment tab that matches your machine.
Loading benchmark contract...
3. Quality report for every PDE
Every generated sample receives finite-value, geometry, initial/boundary, and family-specific physics diagnostics. Dataset aggregation reports each selected PDE separately; the complete benchmark reports all seven losses together.
| PDE | Reported normalized loss | Interpretation |
|---|
Always produced
Gate levels
- Report: report scientific losses without applying an unfrozen PDE threshold; malformed array/geometry contracts are still quarantined.
- Strict: reject structural/BC/IC failures, log them to
*.quality-failures.jsonl, and apply a PDE limit only when calibrated. - Publication candidate: expert-only and intentionally blocked by this Builder until validated solver evidence and a per-stratum threshold table exist. It does not by itself set
publication_ready.