Output formats

`simulate-pheno` outputs

Two files are written:

<out>: A whitespace/tab-delimited phenotype table (#IID + the phenotype column, named per --pheno-name), directly usable as the --pheno argument to baseline, attribute, or pal.
<out with .causal.txt suffix>: One variant ID per line, the union of all SNPs referenced by any --additive/--dominant/--recessive/--interaction2/--interaction3 term. This is the ground truth used to score locus-discovery accuracy in the Quickstart.

For example, --out outputs/simulated.pheno produces outputs/simulated.pheno and outputs/simulated.causal.txt.

`attribute` output (`.glm`-style TSV)

A PLINK-.glm-flavored tab-separated table, one row per variant:

Column	Meaning
`#CHROM`, `POS`, `ID`, `REF`, `ALT`	Copied from the input `.pvar`. Kept identical to PLINK2’s column names/order so the file can be joined against real GWAS summary stats on these five columns.
`A1`	The tested allele; currently always equal to `ALT`.
`TEST`	Model/attribution identifier, e.g. `MLP_integrated_gradients`.
`IMPORTANCE`	Mean absolute Integrated Gradients attribution for that SNP across held-out validation samples.
`P_PERM`	Two-sided p-value from a one-sample t-test of the per-sample signed attribution values against zero (not a permutation test despite the column name — see `integrated_gradients_importance()`).
`P_CORRECTED`	`P_PERM` after Bonferroni correction across all tested variants.
`N`	Number of samples the attribution was computed on (the held-out validation split).

`pal` output (TSV)

Column	Meaning
`#CHROM`, `POS`, `ID`, `REF`, `ALT`	As above.
`MU`	Mean Model Attribution Score (MAS) across all real-label models (`mu` in `PALResult`).
`AMAS`	Aggregated MAS: `MU` down-weighted by how consistently the variant (or an LD-linked neighbor) was independently flagged across models. See Concepts and methodology for the full definition.
`IN_PAL_COMMON`	`True` if the variant exceeded its per-model threshold in every real-label model.
`IN_PAL_AMAS`	`True` if the variant’s `AMAS` score exceeds the global threshold — the primary PAL discovery set.
`P_VALUE`	Bootstrap-estimated p-value against the fitted null distribution, populated only for `IN_PAL_AMAS` variants (`NaN` elsewhere).
`N_MODELS`	Number of real-label models trained (`--n-models`).

Both writers validate that the per-variant array lengths match the number of rows in the .pvar-derived variant table before writing, and raise ValueError on mismatch.