Parity and correctness evidence¶
Measurement environment
The tables below summarize a harness baseline measured in 2026-10 with
PyTorch 2.13. Results on the supported PyTorch 2.14 lane come from the
torchsky-harness CANFAR evaluation campaigns; see the
CANFAR guide.
Scientific correctness is evaluated against trusted references in
torchsky-harness, not in this
repository. This page records what that evidence covers, where it lives, and
what it does not prove. The tables summarize the
baseline record linked below; they are reported by the harness, not re-measured here.
What is compared, and against what¶
| Domain | Reference | Gate | Reported result |
|---|---|---|---|
| WCS celestial transforms | Astropy (WCSLIB) | 38 cases across 30 projections plus reference-latitude-offset cases | Angular parity max 6.1e-7 asec; sky round trip 1.28e-6 asec; pixel round trip 1.27e-6 px |
| WCS throughput | Astropy | Same cases, same machine | Forward 0.46–2.65x, inverse 0.88–4.32x |
| HEALPix indexing and queries | healpy | 12 operations at nside 512 | Zero index/angle mismatches; max 1.14e-13 deg; throughput 1.02–2.19x |
| Spherical and spin SHT | healpy | Transform and spectrum parity | Gates pass; throughput ratios recorded, not gated |
| Catalog crossmatch | Astropy SkyCoord |
Separation parity | Max difference 1.13e-10 asec |
| CUDA WCS | Astropy | 17 cases including offset cases | All pass; max angular parity 8.2e-3 asec |
Source: the harness baseline record in
torchsky-harness/docs/releases/0.1.0/baseline.md.
That document carries the subject artifact identity, the environment, and each
gate's status.
What the gates mean¶
- Parity means the same operation on the same inputs agrees with the trusted implementation to the stated tolerance — not that the two packages share an API or internal representation.
- Round trips check
pixel → world → pixeland the inverse where the transform is invertible. They catch sign, origin, and projection-boundary errors that a single forward comparison can miss. - Throughput ratios are context for the parity run. They are machine- and size-specific and are not a performance claim for other workloads.
What is not covered here¶
- The bounded tests in
tests/(adjoints, gradients, device contracts, API regressions) are necessary but are not a substitute for the reference comparisons above. - Large-data, real-survey, and CANFAR campaign results live in the harness release bundle. They are not copied here because their inputs, versions, and environment must travel with them.
- The reference benchmark tables on this site measure torchsky's own runtime, not agreement with another package. See reference benchmarks.
Reproduce a gate¶
From a harness checkout with a subject build installed:
pixi run torchsky-harness list --domain wcs
pixi run torchsky-harness subject
pixi run torchsky-harness run bench_wcs_parity -- --cases all
The harness records the imported subject and distinguishes process success from valid scientific results. Follow its methodology page rather than invoking comparator packages directly.