# Model Card

## Model summary

This repository selects between a multinomial logistic-regression pipeline and a random
forest using five-fold stratified cross-validation on the training partition. The selected
pipeline is fitted on all training rows. A stratified 20% holdout is reserved from fitting
and cross-validation model selection, then used for final evaluation and interpretation.

## Inputs

The only model inputs are four dereddened SDSS colors:

- `u_g = dered_u - dered_g`
- `g_r = dered_g - dered_r`
- `r_i = dered_r - dered_i`
- `i_z = dered_i - dered_z`

Spectroscopic measurements, redshift, coordinates, identifiers, database classifications,
targeting fields, magnitude errors, and morphology measurements are excluded from the
feature matrix.

## Output

One of three labels: `STAR`, `GALAXY`, or `QSO`.

## Intended use

- Educational astronomy and machine-learning portfolio work.
- Exploring how celestial populations overlap in optical color space.
- Demonstrating reproducible data acquisition, leakage controls, model selection, and
  multiclass evaluation.

## Out-of-scope use

- Scientific catalog production or candidate discovery.
- Real-time telescope classification.
- Population-frequency estimation from predicted probabilities.
- Classification of data from surveys with different bandpasses or selection functions.
- High-stakes or autonomous scientific decision-making.

## Evaluation boundary

All metrics describe a same-survey holdout drawn from an intentionally balanced,
spectroscopically targeted SDSS sample. They are not expected deployment metrics for the
full photometric sky.