Science Assessment ================== Active Model ------------ The active science workflow is intentionally small: * fit ``z_obs`` from photometry with no photo-z prior; * fit ``log10_formed_mass_msun`` as the SED amplitude; * fit lognormal SFH shape parameters ``sfh_t_peak`` and ``sfh_tau``; * fit scalar stellar metallicity ``log10_metallicity``; * derive current SFR from fitted mass plus fitted SFH shape; * inject COSMOS proxy dust columns when ``dust_model: cosmos_proxy_fixed`` is selected; * use ``ssp_flux`` as the SSP spectral table, with separate SSP emission-line tables used only for diagnostics. ``log10_sfr`` and ``dust_av`` are not free parameters in the main science config. ``log10_sfr`` is only a fixed internal SFH scale before mass normalization. With ``dust_model: cosmos_proxy_fixed``, COSMOS dust columns drive attenuation and ``dust_av`` is marked inactive, not inferred. Dust catalogue columns are proxy-only. Even if a future config frees ``dust_av``, the FS2/COSMOS dust proxy should not be plotted as ground truth. Redshift -------- The science preset does not initialize ``z_obs`` from ``phz_median`` and does not use a photo-z prior. ``redshift.initial: fixed`` is only the MAP starting value; fitted ``z_obs`` is free within configured bounds. Redshift multi-start MAP was removed, so posterior runs should be used when redshift degeneracy is scientifically important. PHZ interval priors were removed. Redshift validation must compare fitted ``z_obs`` against ``z_true_gal``. Photometric Errors ------------------ The active 10-band config uses continuum flux columns as fit targets and per-band catalog uncertainties from ``*_el_model3_ext_odonnell_ext_error`` for LSST and Euclid. Those uncertainties are converted from flux units into local AB-magnitude errors for reporting, but the default MAP and MCMC likelihood is Gaussian in ``Fnu`` cgs flux space. Set ``fit.likelihood_space: mag`` only for legacy diagnostics. Truth And Proxies ----------------- Catalog truth columns are diagnostics, not likelihood terms. ``log_stellar_mass`` is converted from catalog ``h`` units for mass comparison. The fitted DSPS amplitude is ``log10_formed_mass_msun``. This is formed mass, not necessarily surviving stellar mass, so an offset relative to catalog stellar mass can be semantic rather than a pure conversion bug. ``log_sfr_true`` is compared to derived current SFR. ``metallicity_true`` is treated as a proxy only: gas-phase oxygen abundance is not the same observable as DSPS stellar metallicity. Fit-vs-catalog plots only include comparable active/free or derived parameters. Fixed inactive quantities such as ``dust_av`` under COSMOS proxy dust are audited but not plotted as inferred parameters. Dust is deliberately excluded from fit-vs-catalog truth metrics because the catalog value is not a DSPS attenuation truth. SED Diagnostics --------------- COSMOS-template SED reconstruction is pseudo-ground truth. It helps inspect whether fitted DSPS SEDs have plausible broad-band shapes. It is not physical SPS truth. Photometric markers in SED plots are model-anchored ratios. They are visual residual diagnostics, not an independent rest-frame spectrum. SED diagnostics also include a flux-space residual panel showing ``(F_obs - F_model) / sigma_F`` per band. Batch MAP runs split SED diagnostics between worst-fit and best-fit galaxies. COSMOS proxy SED CSVs and manifests include the scalar normalization factor and normalization-band residuals. Large COSMOS/DSPS height mismatches should be debugged as normalization, filter, dust, redshift, or template-unit issues before being interpreted as physical disagreement. Non-Detections And Calibration ------------------------------ In flux-space mode, finite non-positive fluxes with valid errors are kept by ``selection.nondetection_policy: gaussian_flux``. ``upper_limit`` is reserved but not implemented. ``band_calibration.mode: fixed_offsets`` can test fixed per-band zero-point offsets. Offsets are applied to the model during likelihood evaluation, never to both data and model. Goodness Of Fit --------------- ``reduced_chi2`` is the DOF-corrected value ``chi2 / dof``. The number of valid bands, effective free parameters, and DOF are written per object. The old ``chi2 / n_bands`` diagnostic is still available as ``chi2_per_band``. Priors ------ Current named prior sets: * ``weak_physical``: broad stabilizing priors used by ``configs/fs2_phz1_science.yaml``; * ``flat_debug``: flat priors inside configured fit bounds; * ``popcosmos_like``: reserved and intentionally unavailable until exact POP-COSMOS parameter mapping, units, and selection treatment are implemented. Current priors are not POP-COSMOS priors yet. Do not label results as POP-COSMOS-like until that mapping exists. Out Of Likelihood ----------------- Emission-line catalog columns remain diagnostics only. The active fit uses ``ssp_flux`` as the spectral table. The newer SSP asset also provides ``ssp_emline_luminosity``, ``ssp_emline_wave``, and line names, when available. The pipeline reads those datasets and writes diagnostic line/filter crossings near fitted-redshift attractors, but it does not add the line table to the likelihood. Adding it naively could double-count lines already present in ``ssp_flux``. DSPS Parameter Count -------------------- The active science MAP fit has five free parameters: * ``z_obs``; * ``log10_formed_mass_msun``; * ``sfh_t_peak``; * ``sfh_tau``; * ``log10_metallicity``. This is not "the DSPS paper parameter count". DSPS is a differentiable SPS framework. The paper demonstrates differentiability through SFH, metallicity, dust, and nebular choices, but does not define one canonical Euclid+LSST photo-z fit. Any future paper-like preset must state the exact SFH model, metallicity model, dust model, nebular model, active parameter list, units, and priors.