# Prop-02 Option B — diversity + chemsys leakage (fingerprint secondary)

**Framing:** fingerprint Magpie-lite track is pre-registered **secondary** (not support alone).
**`prop02_success_claimed`: False.** GNN remains primary for official support.

## Setup (match pilot scale)
- N=5000 stratified al_pool; hits=558
- seed=100; batch=25; budget=400; seeds=5
- Arms: random, uncertainty (RF vote-std), diversity (farthest-point in scaled fingerprints)
- Leakage proxy: fraction of acquired (post-seed) sharing a `chemsys` with the seed
- Control: `ban_seen_chemsys` — prefer querying unseen chemical systems

## Results

| ban_chemsys | arm | final recall ± std | CI95 | AULC ± std | frac share seed chemsys ± std |
|---|---|---:|---:|---:|---:|
| False | random | 0.1072 ± 0.0099 | [0.0985, 0.1159] | 0.0627 ± 0.0050 | 0.003 ± 0.002 |
| False | uncertainty | 0.2925 ± 0.0136 | [0.2805, 0.3044] | 0.1524 ± 0.0106 | 0.007 ± 0.003 |
| False | diversity | 0.0516 ± 0.0074 | [0.0451, 0.0581] | 0.0316 ± 0.0022 | 0.001 ± 0.001 |
| True | random | 0.0982 ± 0.0063 | [0.0927, 0.1037] | 0.0565 ± 0.0030 | 0.000 ± 0.000 |
| True | uncertainty | 0.2950 ± 0.0105 | [0.2858, 0.3042] | 0.1514 ± 0.0104 | 0.000 ± 0.000 |
| True | diversity | 0.0516 ± 0.0074 | [0.0451, 0.0581] | 0.0316 ± 0.0022 | 0.000 ± 0.000 |

Figure: `figures/track2_optionb_diversity_leakage.png`

## Limits
Chemsys is a coarse leakage proxy (not structure-graph similarity).
Exploratory Option B only — does not constitute Prop-02 support.