Forward Model

Boundary

euclid_dsps.model is the only native DSPS boundary. MAP, MCLMC, synthetic-closure generation, and amortized inference all pass through the same theta-vector and model-evaluation contract.

FENIKS Closure Model

The active FENIKS closure configs use the 18D diffsky_basic parameter vector:

z_obs
log10_stellar_mass
diffstar_lgmcrit
diffstar_lgy_at_mcrit
diffstar_indx_lo
diffstar_indx_hi
diffstar_lg_qt
diffstar_qlglgdt
diffstar_lg_drop
diffstar_lg_rejuv
diffmah_logm0
diffmah_logtc
diffmah_early_index
diffmah_late_index
diffmah_t_peak
log10_stellar_metallicity
dust_av
dust_delta

The 14-band production setup uses:

LSST u,g,r,i,z,y + Roman F062,F087,F106,F129,F146,F158,F184,F213

During generation, FENIKS truths are decoded with local DSPS to write flux_true_<band>. The configured m5_depth error model then writes fluxerr_<band> and noisy flux_<band>. Validation recomputes the true fluxes from the stored truth vector.

FS2 Comparison Model

configs/fs2_gpu.yaml keeps the Euclid comparison path. It uses ten bands:

LSST u,g,r,i,z,y + Euclid VIS,Y,J,H

and a 16-parameter PopCosmos-like vector with compressed stellar, gas, and AGN assets. FS2 is useful for comparison and regression checks, not for defining the FENIKS learned-prior target.

HLTDS Debug Model

The active HLTDS configs are dataset/download configs, not MAP experiment configs. They preserve the low-z HLTDS photometry, error model, provenance, and projected-truth fields so that FENIKS can be compared to an external reference.

SSP Assets

FENIKS and HLTDS use the HLTDS SSP assets:

Data/diffsky/raw/hltds_cosmos_260215_04_14_2026/diffsky_hltds_cosmos_260215_04_14_2026_ssp_data.hdf5
Data/diffsky/raw/hltds_cosmos_260215_04_14_2026/diffsky_hltds_cosmos_260215_04_14_2026_ssp_basis_k64_coeff16.hdf5

FS2 uses the local compressed FSPS assets:

Data/popcosmos_chabrier_stellar_ssp_basis_k64_coeff16.h5
Data/popcosmos_chabrier_gas_grid_basis_k64_mixed16.h5
Data/popcosmos_chabrier_agn_component_basis_k12_fagnlinear_coeff16.h5

Likelihoods

The active configs use flux-space Student-t likelihoods with explicit fluxerr_* columns plus configured fractional floors. For FENIKS, the error columns are generated by the closure pipeline and are part of the controlled dataset contract. For HLTDS, the error columns are materialized during dataset preparation and should be treated as debug/reference metadata.