A field profile gives the engine domain-aware reviewer weighting (loaded via --field physics, see docs/ENGINE.md §2.2). It does not relax any universal rule — it only re-prioritizes which branches the 12 framings explore first. Every list stays short and concrete; a physics reviewer weighs unit consistency and error budgets, not vibes.

What does a senior ApJ/MNRAS/PRD reviewer reliably attack first?

• Dimensional/unit consistency: h⁻¹ Mpc vs Mpc, comoving vs physical distances, stray factors of h in masses and power spectra.
• Error budget quotes only the statistical σ while the dominant systematic (PSF modelling, shear multiplicative bias m, additive bias c, baryons) is unbudgeted.
• Significance quoted without a look-elsewhere / trials correction for the scanned parameter range.
• Covariance estimated from too few mock realizations with no Hartlap (or Sellentin–Heavens) correction, biasing the inverse covariance.
• Method validated only on the simulation family it was tuned on, with no independent N-body/hydro suite.

What does the field take for granted that a result might quietly lean on?

• Gaussian likelihood for two-point statistics — breaks at small scales / low S/N where the covariance is non-Gaussian and the estimator is skewed.
• Photo-z posteriors are approximately Gaussian — breaks on catastrophic outliers whose true redshift sits in a secondary mode.
• Shape noise dominates the shear covariance — breaks at large scales / high source density where sample (cosmic) variance takes over.
• A fixed nonlinear P(k) fitting formula is adequate — breaks at k ≳ 0.1–1 h/Mpc where baryonic feedback shifts power at the 10–30% level.
• Intrinsic alignments are a subdominant additive term — breaks for luminous-red / low-z lens bins where IA rivals the lensing signal.

Concrete, recurring ways the work goes wrong in practice.

• An off-by-h or comoving/physical mismatch mislabels the x-axis while the code runs clean.
• A Fisher forecast reports unrealistically tight constraints because it fixed nuisances (m, photo-z Δz, IA amplitude) that should be marginalized.
• Best-fit χ² looks acceptable because the model was fit and evaluated on the same scales, with no held-out check.
• Shear bias calibrated on one image-sim blend density, then applied to data with different blending.
• MCMC chains quoted as final while Gelman–Rubin R̂ is still far from 1.

• Strong: reproduction on an independent dataset/survey/sim suite, a blinded analysis frozen before unblinding, and full-covariance error bars that include the dominant systematics on converged chains.
• Weak (insufficient on its own): a single-realization σ, visual agreement of two curves by eye, χ² on the tuning set, or a forecast with nuisance parameters held fixed.