# Structure AL leakage / near-duplicate diagnostic

**Exploratory (Prop-02 02a / Option B secondary). `prop02_success_claimed`: False.**
Fingerprint-only ≠ support. Same N=5k Magpie-lite uncertainty setup as the structure pilot.

## Definitions
- Fingerprints: FLUX ENGINE Magpie-lite, `StandardScaler`d.
- Near-dup: max cosine similarity of a point to the **seed** labeled set.
- Thresholds: [0.99, 0.95, 0.9]
- Filtered U: at each batch, drop unlabeled with max cosine to current labeled ≥ threshold (fallback if too few).

## Pool vs seed baseline (unlabeled pool)

- mean max-cosine to seed: **0.6278** ± 0.0048
- p95 max-cosine: **0.8852**
- frac cos≥0.99: 0.0007 ± 0.0003
- frac cos≥0.95: 0.0074 ± 0.0022
- frac cos≥0.90: 0.0370 ± 0.0074

## Uncertainty runs (mean ± std; CI95 over 5 seeds)

| variant | final recall | CI95 | frac share seed chemsys | frac cos≥0.95 to seed |
|---|---:|---:|---:|---:|
| `U_unfiltered` | 0.3018 ± 0.0100 | [0.2930, 0.3106] | 0.003 ± 0.002 | 0.003 ± 0.002 |
| `U_filter_cos_0.99` | 0.3025 ± 0.0114 | [0.2925, 0.3125] | 0.003 ± 0.002 | 0.002 ± 0.001 |
| `U_filter_cos_0.95` | 0.3039 ± 0.0089 | [0.2961, 0.3118] | 0.003 ± 0.003 | 0.000 ± 0.000 |
| `U_filter_cos_0.90` | 0.2961 ± 0.0098 | [0.2874, 0.3047] | 0.001 ± 0.001 | 0.000 ± 0.000 |

Figure: `figures/track2_structure_al_leakage.png`

## Read
If filtered recall collapses toward random while unfiltered U looks strong, treat fingerprint U gains as leakage-sensitive.
Chemsys overlap is a coarse chemistry proxy alongside cosine near-dups.

## Limits
Not structure-graph similarity; not GNN primary path; not a Prop-02 support decision.