Simulation algorithm research — grading contract¶
Status: Active
Audience: numerics_researcher / autoresearch in MD and chemistry vertical plan loops
Methodology: benchmarks numerics research-methodology
Axes¶
| Axis | Primary evidence | Hard gate? |
|---|---|---|
| Validity | Tier-2 verify-results, composable smoke, checksum parity, registry honesty | Yes — todo cannot complete if fail |
| Performance | bench-package.sh --timing, ratio vs cpp/native | No — document tradeoff |
| Memory | sim-bench-memory.sh peak RSS | No — document tradeoff |
| Security | No new unsafe FFI; scoped li-tests security when touching common/*_core.c | Fail when native touched and security slice fails |
| Numerical stability | tier-0 rows, energy drift, no NaN blow-up | Tied to validity for physics kernels |
| Size scaling | ≥3 problem sizes in study table (N, grid, basis count) | Required in study deliverable |
Validity is never traded for speed or memory unless a human approves in the study doc with explicit locked axes.
Tradeoff documentation (required in every study)¶
Each docs/numerics/studies/YYYY-MM-DD-<slug>.md must end with:
## Grade matrix
| Axis | Result | vs prior | Notes |
|------|--------|----------|-------|
| Validity | pass/fail | … | … |
| Performance | … | … | … |
| Memory | … | … | … |
| Security | pass/skip/N/A | … | … |
| Stability | … | … | … |
| Size scaling | table attached | … | … |
## Tradeoffs
- Locked: validity (+ stability for MD integrators)
- Improved: …
- Regressed (if any): … — justified or rejected
Vertical scope¶
| Vertical | Package | Benches / smokes |
|---|---|---|
| md | li-sim-scientific | md_lennard_jones, heat_equation_2d; md_neighbor_cell_list when present |
| chem | composable + future chem | import_chem_dft_smoke when present; survey todos may pass on study-only |
Enforcement¶
Writes data/sim-<vertical>-research-loop/grade.json for the live agents canvas.