science-core

name	science-core
description	Cross-cutting scientific discipline for the astroai/sfabbro astronomy stack — scientific correctness, math rigor, statistical principles, data-driven validation, science impact
metadata	{"domains":"astronomy, astrophysics, machine-learning"}

Science core discipline

Applies to every repo in the ecosystem (torchsky, torchfits, uspm, cosmodist, xmatch, cfhtcast, weightmask, torchregress, torchz, zensus). Read the repo's AGENTS.md for repo-specific rules; this skill is the shared scientific standard. When repo guidance conflicts, AGENTS.md wins.

Scientific correctness

• Exact, well-founded methods over heuristics. No ad hoc thresholds or data-dependent branching without a numerical, algorithmic, or benchmark justification.
• Verify against trusted baselines: upstream libraries (astropy, healpy, reproject, CFITSIO), analytic ground truth, or closed-form solutions.
• State units, coordinate frames, and physical conventions explicitly at boundaries; never mix radians/degrees or frames silently.
• Fail closed on invalid states; assert shape/dtype invariants at function boundaries.

Math logic

• Derive before coding: write the math (loss, Jacobian, propagation, adjoint) down and check it against a reference before implementing.
• Precision: float64 for astronomy numerics unless measured; preserve differentiability and device behavior in public torch paths.
• Numerical stability: log-space where needed, stable softplus/logsumexp, avoid catastrophic cancellation; test edge regimes (zero variance, empty bins, NaN handling).
• When adding a fast path, keep a correct reference path and a parity test.

Statistics-principled

• Always quantify uncertainty: error bars, predictive distributions, or coverage — never bare point estimates for science claims.
• Calibration over confidence: validate that stated coverage matches empirical coverage (conformal, quantile, or histogram checks).
• Use proper scoring rules and honest metrics (CRPS, interval score, NLL, coverage), not just MSE.
• Document priors, weights, and missing-data handling; make assumptions explicit and testable.
• No p-hacking or cherry-picking: pre-register comparisons, report all trials, and state significance honestly.

Data-driven

• Validate on real survey data and realistic simulations, not toy examples alone. Real-data or upstream-fixture checks are required for user-facing behavior.
• Benchmarks on real workloads with honest methodology: fixed sizes, exact parity, throughput AND memory, committed baseline gates.
• Data are sacred in this stack: keep data in native frames where the design says so; never silently resample or transform for convenience.

Science impact

• Prioritize work by expected scientific and operational impact, then correctness risk, then performance at survey scale.
• Reproducibility is non-negotiable: every experiment traceable (session logs, configs, seeds, versions), results archivable, claims backed by evidence.
• Deliver publishable, skeptical-audience-grade results: interpretable models, documented physical/statistical motivations, honest comparisons against the state of the art (e.g. SED fitting for photo-z).