physics-check

name	physics-check
description	Numerical correctness check on saved result arrays (npz, HDF5, torch, etc.). Use after plot-check or when validating metrics without a new figure. Domain laws, bounds, and thresholds live in the project's CLAUDE.md, not here.

Physics check

Read the numbers behind a figure or analysis output. This skill does not replace plot-check (visual + point-count completeness); run both when a script just wrote a plot and arrays.

Principle: asserts on arrays catch wrong normalizations, sign errors, and silent drops that VLMs miss.

When to run

• After plot-check passes (or when only arrays/metrics changed, no new figure).
• Before citing quantitative claims in text or captions.
• When metrics jump by orders of magnitude or conservation seems violated.

Step 1 — load and summarize

For each array file the result depends on:

import numpy as np

d = np.load("path.npz")  # adapt for .npy, HDF5, torch, Parquet, etc.
print(sorted(d.files))
for k in d.files:
    a = np.asarray(d[k])
    print(k, a.shape, a.dtype,
          float(np.nanmin(a)), float(np.nanmax(a)),
          int(np.isfinite(a).all()))

Step 2 — generic asserts

assert np.isfinite(a).all(), f"{k}: non-finite values"
assert a.size > 0, f"{k}: empty array"
if a.ndim == 1 and k in ("x", "y", "ell", "bin_centers", "values"):
    n_bad = int(np.sum(~np.isfinite(a)))
    assert n_bad == 0, f"{k}: {n_bad} non-finite values will not plot"
if "ratio" in k or "residual" in k:
    assert np.nanmax(np.abs(a)) < 1e12, f"{k}: ratio/residual implausible"

Step 3 — project asserts (required)

Apply rules from the repo's CLAUDE.md (or a project-local validator skill):

• Conservation, symmetries, sign constraints, normalization conventions.
• Physical bounds (e.g. densities ≥ 0, |δ| ≪ 1 where expected).
• Cross-series consistency (same grid, units, sample definition).

This shared skill does not encode domain physics — only generic numerics.

Step 4 — provenance

• Sidecar metadata states units, normalization, seed when stochastic.
• Caption claims ("observed", "simulated", sample N) match arrays and config.
• Compared series share grid/units or document the transform → else BLOCKING.

Step 5 — verdict

Exactly one of:

• OK: <one-line summary with dominant numerical result>
• WARN: <summary>; <investigate>
• BLOCKING: <reason>; <fix>

BLOCKING → do not ship in manuscript, slides, or external report until fixed. Record in research-figure-manifest (physics_check field or combined validator).

Integration

Skill	Role
plot-check	Visual + point-count pass first when a figure exists
vlm-figure-audit	Print QA after numeric pass
research-figure-manifest	Store verdict with run id
results-check	Orchestrates plot-check then physics-check

Repo-local physics-check tables in CLAUDE.md extend Step 3; do not duplicate them here.