# Track 2 Magpie-lite feature-group ablation

**`prop02_success_claimed`: False.** Same N=5k setup as `structure_al_pilot.md`. Uncertainty arm only for LOGO.

## Question
Which Magpie-lite **structure-meta** groups drive the composition→composition+structure lift?

## Groups
| Group | Features |
|---|---|
| `struct_size` | log1p_nsites, nelements |
| `struct_geometry` | log1p_volume, density |
| `struct_thermo` | energy_above_hull, is_stable |
| `struct_electronic` | is_metal |

## Leave-one-group-out (uncertainty final recall @ budget 400)

| Variant | n_feat | mean ± std | CI95 | Δ vs full |
|---|---:|---:|---:|---:|
| `composition` | 84 | 0.2419 ± 0.0164 | [0.2275, 0.2563] | -0.0599 |
| `drop_struct_electronic` | 90 | 0.2781 ± 0.0172 | [0.2631, 0.2932] | -0.0237 |
| `drop_struct_size` | 89 | 0.2939 ± 0.0073 | [0.2875, 0.3003] | -0.0079 |
| `drop_struct_geometry` | 89 | 0.2832 ± 0.0169 | [0.2684, 0.2979] | -0.0186 |
| `drop_struct_thermo` | 89 | 0.2946 ± 0.0095 | [0.2863, 0.3030] | -0.0072 |
| `composition+structure` | 91 | 0.3018 ± 0.0100 | [0.2930, 0.3106] | +0.0000 |

Composition-only vs full lift: **+0.0599** absolute final recall.

## RF impurity importance (group sums, 1 AL seed final labeled set)

| Group | Summed importance |
|---|---:|
| `composition_frac` | 0.3968 |
| `composition_magpie` | 0.3127 |
| `struct_geometry` | 0.1332 |
| `struct_electronic` | 0.0846 |
| `struct_size` | 0.0726 |
| `struct_thermo` | 0.0000 |

### Structure features (individual)

| Feature | Importance |
|---|---:|
| `is_metal` | 0.0846 |
| `log1p_volume` | 0.0711 |
| `density` | 0.0621 |
| `log1p_nsites` | 0.0590 |
| `nelements` | 0.0136 |
| `energy_above_hull` | 0.0000 |
| `is_stable` | 0.0000 |

Figure: `figures/track2_structure_al_ablation.png`

## Limits
LOGO removes one structure group at a time from the full vector; not a causal claim.
Importance is impurity-based on one seed's final labeled set — supportive, not definitive.
Still not Prop-02 success.