PROTON+: all results
Every table of the preprint and its supplement, per benchmark. AUROC, AUPR and FPR@95 in %, online detectors as the mean over five stream orders. Generated by scripts/build_site.py from the outputs of proton run. Back to the project page.
Ophthalmology
Zero-shot OOD detection (static scores, PROTON, PROTON+)
| Method | Sem AUC | Sem APR | Sem FPR | Cov AUC | Cov APR | Cov FPR | Far AUC | Far APR | Far FPR |
|---|
| Max-Logits | 54.0 | 54.1 | 79.3 | 27.9 | 23.3 | 88.0 | 83.9 | 76.4 | 63.0 |
| Energy | 51.8 | 51.5 | 79.7 | 27.2 | 23.2 | 88.0 | 83.9 | 76.5 | 62.7 |
| GL-MCM | 75.8 | 75.6 | 86.0 | 68.9 | 45.3 | 69.3 | 78.5 | 62.4 | 75.0 |
| MCM | 75.8 | 75.6 | 86.0 | 68.9 | 45.2 | 69.7 | 78.5 | 62.4 | 75.0 |
| TAG-MCM | 68.4 | 70.5 | 88.7 | 30.1 | 23.8 | 93.7 | 72.2 | 47.2 | 60.3 |
| MCM (L=1) | 66.5 | 65.0 | 84.7 | 84.2 | 66.3 | 47.3 | 51.8 | 38.4 | 98.3 |
| MCM (L=5) | 58.3 | 56.5 | 87.0 | 66.5 | 42.5 | 65.7 | 54.4 | 33.5 | 74.3 |
| PROTON (ours) | 68.0 | 72.5 | 92.7 | 46.7 | 29.4 | 83.6 | 86.8 | 76.2 | 55.0 |
| Δ vs MCM | -7.7 | -3.1 | 6.7 | -22.1 | -15.8 | 13.9 | 8.3 | 13.7 | -20.0 |
PROTON mean ± std over 5 stream orders: semantic: 68.0±0.9, covariate: 46.7±1.8, far: 86.8±3.6
PROTON+ against PROTON (components, ID fraction, sensitivity, base scores)
| Variant | Sem AUC | Cov AUC | Far AUC | Sem FPR | Cov FPR | Far FPR | t_calib |
|---|
| MCM | 75.8 | 68.9 | 78.5 | 86.0 | 69.7 | 75.0 | nan |
| PROTON (paper) | 68.0 | 46.7 | 86.8 | 92.7 | 83.6 | 55.0 | 14 |
| + z-scored fusion | 69.4 | 50.7 | 89.0 | 91.3 | 77.7 | 32.3 | 14 |
| + consistency admission | 70.6 | 40.1 | 92.6 | 83.0 | 89.5 | 18.6 | 14 |
| + domain-anchored prototypes (PROTON+) | 67.3 | 85.3 | 93.4 | 75.8 | 31.5 | 15.4 | 14 |
| domain-anchored only | 64.0 | 86.3 | 94.3 | 84.9 | 33.2 | 32.9 | 14 |
| consistency only | 68.9 | 30.5 | 91.5 | 90.1 | 95.1 | 35.6 | 14 |
| PROTON+ max-fusion | 65.7 | 86.6 | 93.2 | 85.6 | 28.0 | 15.6 | 14 |
| PROTON+ rejection-driven alpha | 64.4 | 86.9 | 93.6 | 75.9 | 29.9 | 15.5 | 14 |
| S_proto only (PROTON bank) | 62.7 | 43.7 | 90.5 | 91.6 | 80.8 | 47.0 | 14 |
| S_proto only (PROTON+ bank) | 57.0 | 90.4 | 97.9 | 78.7 | 26.0 | 8.9 | 14 |
| PROTON, oracle ID-only bank | 76.8 | 93.1 | 94.5 | 91.1 | 24.1 | 26.1 | 21 |
| PROTON+, oracle ID-only bank | 73.6 | 90.4 | 93.5 | 71.3 | 12.9 | 15.3 | 21 |
| rho | PROTON+ Sem/Cov/Far | PROTON Sem/Cov/Far | MCM Sem/Cov/Far |
|---|
| 90% | 70.6 / 88.5 / 90.9 | 73.1 / 85.5 / 90.4 | 76.2 / 67.7 / 75.5 |
| 70% | 68.7 / 74.7 / 92.9 | 71.2 / 47.7 / 87.4 | 74.8 / 69.0 / 79.0 |
| 50% | 67.4 / 22.3 / 36.7 | 68.2 / 34.0 / 55.3 | 75.8 / 69.1 / 78.4 |
| 30% | 64.5 / 28.8 / 27.8 | 60.3 / 28.0 / 26.4 | 76.2 / 68.3 / 78.0 |
| 10% | 56.8 / 41.4 / 20.5 | 50.0 / 41.3 / 16.7 | 75.9 / 69.2 / 78.1 |
| gamma (PROTON+) | Sem | Cov | Far |
|---|
| 0.5 | 65.8 | 86.3 | 93.2 |
| 0.6 | 66.4 | 85.7 | 93.3 |
| 0.7 | 67.3 | 85.3 | 93.4 |
| 0.8 | 67.8 | 87.8 | 93.5 |
| 0.9 | 68.0 | 88.2 | 93.6 |
| kappa (PROTON+) | Sem | Cov | Far |
|---|
| 1.0 | 67.7 | 87.7 | 93.4 |
| 1.5 | 67.4 | 87.5 | 93.4 |
| 2.0 | 67.3 | 85.3 | 93.4 |
| 3.0 | 67.2 | 85.0 | 93.4 |
| 4.0 | 67.2 | 85.0 | 93.1 |
| Base score | Base Sem/Cov/Far | +PROTON+ Sem/Cov/Far |
|---|
| Max-Logits | 54.0 / 27.9 / 83.9 | 57.1 / 87.1 / 99.1 |
| Energy | 51.8 / 27.2 / 83.9 | 56.1 / 87.2 / 99.1 |
| GL-MCM | 75.8 / 68.9 / 78.5 | 67.3 / 85.3 / 93.4 |
| MCM | 75.8 / 68.9 / 78.5 | 67.3 / 85.3 / 93.4 |
| Entropy | 75.8 / 68.9 / 78.5 | 63.3 / 86.5 / 96.6 |
| TAG-MCM | 68.4 / 30.1 / 72.2 | 65.2 / 58.7 / 91.5 |
PROTON component ablations
| Variant | Cov | Sem | Far |
|---|
| MCM baseline | 68.9 | 75.8 | 78.5 |
| + Proto. (alpha=.5) | 46.6 | 65.5 | 86.9 |
| + Adapt. fusion | 46.9 | 65.3 | 85.3 |
| + Conf. gate (Full) | 46.7 | 68.0 | 86.8 |
| gamma | AUC | APR | F95 |
|---|
| 0.5 | 46.9 | 29.4 | 82.5 |
| 0.6 | 46.7 | 29.4 | 83.8 |
| 0.7 | 46.7 | 29.4 | 83.6 |
| 0.8 | 47.8 | 30.1 | 81.3 |
| 0.9 | 49.4 | 30.7 | 79.9 |
| M | AUC | APR | F95 |
|---|
| 10 | 45.3 | 29.3 | 92.0 |
| 25 | 44.6 | 28.6 | 89.9 |
| 50 | 45.4 | 28.9 | 87.2 |
| 100 | 46.7 | 29.4 | 83.6 |
| 200 | 47.9 | 29.9 | 79.5 |
Stream ID-fraction sweep
| rho | Sem | Cov | Far | (MCM: Sem | Cov | Far) |
|---|
| 90% | 73.1 | 85.5 | 90.4 | 76.2 | 67.7 | 75.5 |
| 70% | 71.2 | 47.7 | 87.4 | 74.8 | 69.0 | 79.0 |
| 30% | 60.3 | 28.0 | 26.4 | 76.2 | 68.3 | 78.0 |
| 10% | 50.0 | 41.3 | 16.7 | 75.9 | 69.2 | 78.1 |
PROTON on six base scores
| Base score | Base (Cov) | +PROTON (Cov) | Δ | Base (Sem) | +PROTON (Sem) | Base (Far) | +PROTON (Far) |
|---|
| Max-Logits | 27.9 | 23.6 | -4.4 | 54.0 | 57.3 | 83.9 | 87.0 |
| Energy | 27.2 | 22.6 | -4.5 | 51.8 | 55.3 | 83.9 | 87.0 |
| GL-MCM | 68.9 | 45.7 | -23.2 | 75.8 | 67.6 | 78.5 | 85.0 |
| MCM | 68.9 | 46.7 | -22.1 | 75.8 | 68.0 | 78.5 | 86.8 |
| Entropy | 68.9 | 49.7 | -19.2 | 75.8 | 70.0 | 78.5 | 89.8 |
| TAG-MCM | 30.1 | 22.6 | -7.5 | 68.4 | 67.9 | 72.2 | 89.4 |
Broader baselines, including privileged ones
| Method | Cov AUC | Sem AUC | Far AUC |
|---|
| TAG-MCM | 30.1 | 68.4 | 72.2 |
| MCM (L=5) | 66.5 | 58.3 | 54.4 |
| kNN-ID | 100.0 | 76.8 | 99.8 |
| Mahalanobis-ID | 100.0 | 82.4 | 100.0 |
| PROTON | 46.7 | 68.0 | 86.8 |
| hier_l5+proton | 24.0 | 41.2 | 25.3 |
Prompt sensitivity
| Prompt set | ID acc | MCM Sem/Cov/Far | PROTON Sem/Cov/Far | PROTON+ Sem/Cov/Far |
|---|
| flair_default (default) | 84.0 | 75.8 / 68.9 / 78.5 | 68.0 / 46.7 / 86.8 | 67.3 / 85.3 / 93.4 |
| vanilla | 86.0 | 65.6 / 64.9 / 83.5 | 58.5 / 46.7 / 94.4 | 58.2 / 87.9 / 97.0 |
| l1 | 83.7 | 66.5 / 84.2 / 51.8 | 65.0 / 66.6 / 78.2 | 64.6 / 94.0 / 88.1 |
| l5 | 57.3 | 58.3 / 66.5 / 54.4 | 41.2 / 24.0 / 25.3 | 56.2 / 50.8 / 58.8 |
Robustness extras
sigma2_0 sweep mean AUROC: 0.005: 66.7, 0.01: 66.7, 0.02: 67.2, 0.04: 68.4, 0.08: 68.6
misrouting (cov AUROC): 0.0: 46.7, 0.1: 50.0, 0.2: 51.2, 0.3: 54.5
overhead: PROTON 0.020 ms/sample, PROTON+ 0.282 ms/sample, bank 0.41 MB
wins vs MCM over stream orders: {'semantic': {'proton': '0/5', 'proton_plus': '0/5'}, 'covariate': {'proton': '0/5', 'proton_plus': '5/5'}, 'far': {'proton': '5/5', 'proton_plus': '5/5'}}
triage macro-F1: 0.5968720012468626
PROTON gate x bank-size grid (change in covariate AUROC over MCM)
| M=10 | M=25 | M=50 | M=75 | M=100 | M=150 | M=200 | M=500 |
|---|
| g=0.3 | -22.5 | -23.3 | -22.6 | -22 | -21.5 | -20.9 | -20.5 | -20.3 |
| g=0.4 | -22.5 | -23.3 | -22.6 | -22 | -21.5 | -20.9 | -20.5 | -20.3 |
| g=0.5 | -22.5 | -23.3 | -22.6 | -22 | -21.5 | -20.9 | -20.5 | -20.3 |
| g=0.6 | -23 | -23.8 | -23 | -22.2 | -21.7 | -21.1 | -20.6 | -20.5 |
| g=0.7 | -22.9 | -23.7 | -23 | -22.1 | -21.6 | -21 | -20.4 | -20.3 |
| g=0.8 | -21.9 | -22.8 | -22.1 | -21.2 | -20.8 | -20.2 | -19.7 | -19.6 |
| g=0.9 | -20.2 | -21.1 | -20.6 | -19.9 | -19.5 | -19.1 | -18.8 | -18.7 |
Radiology
Zero-shot OOD detection (static scores, PROTON, PROTON+)
| Method | Sem AUC | Sem APR | Sem FPR | Cov AUC | Cov APR | Cov FPR | Far AUC | Far APR | Far FPR |
|---|
| Max-Logits | 69.4 | 56.1 | 82.3 | 0.4 | 10.8 | 100.0 | 75.7 | 64.6 | 90.2 |
| Energy | 70.2 | 57.4 | 84.8 | 0.4 | 10.8 | 100.0 | 77.0 | 67.4 | 90.5 |
| GL-MCM | 54.3 | 35.7 | 89.7 | 46.7 | 18.1 | 87.3 | 29.2 | 14.2 | 98.0 |
| MCM | 54.2 | 35.2 | 89.7 | 46.6 | 18.1 | 87.3 | 27.6 | 14.0 | 99.7 |
| TAG-MCM | 57.4 | 36.8 | 87.8 | 48.1 | 19.3 | 85.5 | 41.3 | 16.0 | 87.5 |
| MCM (L=1) | 54.2 | 35.2 | 89.7 | 46.6 | 18.1 | 87.3 | 27.6 | 14.0 | 99.7 |
| MCM (L=5) | 54.2 | 32.0 | 77.5 | 78.4 | 38.2 | 46.3 | 84.6 | 52.3 | 39.5 |
| PROTON (ours) | 55.3 | 38.3 | 95.2 | 61.2 | 31.0 | 67.6 | 93.9 | 92.5 | 33.5 |
| Δ vs MCM | 1.1 | 3.1 | 5.5 | 14.6 | 12.9 | -19.8 | 66.3 | 78.5 | -66.2 |
PROTON mean ± std over 5 stream orders: semantic: 55.3±0.9, covariate: 61.2±1.7, far: 93.9±2.3
PROTON+ against PROTON (components, ID fraction, sensitivity, base scores)
| Variant | Sem AUC | Cov AUC | Far AUC | Sem FPR | Cov FPR | Far FPR | t_calib |
|---|
| MCM | 54.2 | 46.6 | 27.6 | 89.7 | 87.3 | 99.7 | nan |
| PROTON (paper) | 55.3 | 61.2 | 93.9 | 95.2 | 67.6 | 33.5 | 30 |
| + z-scored fusion | 61.6 | 78.1 | 90.5 | 87.6 | 56.1 | 40.1 | 30 |
| + consistency admission | 61.9 | 78.4 | 92.8 | 87.3 | 60.5 | 36.1 | 30 |
| + domain-anchored prototypes (PROTON+) | 61.4 | 93.9 | 91.5 | 88.1 | 26.7 | 35.7 | 30 |
| domain-anchored only | 55.2 | 86.8 | 95.0 | 95.2 | 38.7 | 19.9 | 30 |
| consistency only | 55.4 | 63.2 | 94.9 | 95.2 | 68.0 | 22.0 | 30 |
| PROTON+ max-fusion | 56.0 | 92.2 | 90.8 | 89.6 | 27.4 | 34.7 | 30 |
| PROTON+ rejection-driven alpha | 60.8 | 95.8 | 93.8 | 88.3 | 17.4 | 35.8 | 30 |
| S_proto only (PROTON bank) | 61.8 | 92.3 | 95.0 | 95.2 | 41.5 | 19.8 | 30 |
| S_proto only (PROTON+ bank) | 62.6 | 98.8 | 96.2 | 83.1 | 1.3 | 20.5 | 30 |
| PROTON, oracle ID-only bank | 55.4 | 86.5 | 94.9 | 95.1 | 50.6 | 19.8 | 49 |
| PROTON+, oracle ID-only bank | 64.9 | 92.4 | 91.3 | 87.2 | 33.8 | 35.8 | 49 |
| rho | PROTON+ Sem/Cov/Far | PROTON Sem/Cov/Far | MCM Sem/Cov/Far |
|---|
| 90% | 64.8 / 96.4 / 94.1 | 57.8 / 71.0 / 94.1 | 53.7 / 46.8 / 27.5 |
| 70% | 62.0 / 89.5 / 81.0 | 54.9 / 54.3 / 81.8 | 54.4 / 47.2 / 27.9 |
| 50% | 58.8 / 69.0 / 73.0 | 54.5 / 49.9 / 58.8 | 54.5 / 47.9 / 28.5 |
| 30% | 55.6 / 12.2 / 14.2 | 54.0 / 38.8 / 37.1 | 54.5 / 46.7 / 27.7 |
| 10% | 52.5 / 12.7 / 13.1 | 53.8 / 23.7 / 21.4 | 53.5 / 47.5 / 27.5 |
| gamma (PROTON+) | Sem | Cov | Far |
|---|
| 0.5 | 60.1 | 95.2 | 94.8 |
| 0.6 | 60.8 | 94.8 | 93.7 |
| 0.7 | 61.4 | 93.9 | 91.5 |
| 0.8 | 62.4 | 92.8 | 87.0 |
| 0.9 | 59.6 | 77.5 | 71.8 |
| kappa (PROTON+) | Sem | Cov | Far |
|---|
| 1.0 | 61.6 | 94.1 | 91.5 |
| 1.5 | 61.6 | 94.0 | 91.5 |
| 2.0 | 61.4 | 93.9 | 91.5 |
| 3.0 | 61.2 | 92.8 | 91.5 |
| 4.0 | 61.2 | 92.8 | 91.5 |
| Base score | Base Sem/Cov/Far | +PROTON+ Sem/Cov/Far |
|---|
| Max-Logits | 69.4 / 0.4 / 75.7 | 69.1 / 92.9 / 98.8 |
| Energy | 70.2 / 0.4 / 77.0 | 69.2 / 92.9 / 98.9 |
| GL-MCM | 54.3 / 46.7 / 29.2 | 61.4 / 93.9 / 91.6 |
| MCM | 54.2 / 46.6 / 27.6 | 61.4 / 93.9 / 91.5 |
| Entropy | 54.2 / 46.6 / 27.6 | 64.3 / 98.1 / 95.5 |
| TAG-MCM | 57.4 / 48.1 / 41.3 | 63.6 / 86.8 / 83.9 |
PROTON component ablations
| Variant | Cov | Sem | Far |
|---|
| MCM baseline | 46.6 | 54.2 | 27.6 |
| + Proto. (alpha=.5) | 66.0 | 55.2 | 96.0 |
| + Adapt. fusion | 67.7 | 55.3 | 97.2 |
| + Conf. gate (Full) | 61.2 | 55.3 | 93.9 |
| gamma | AUC | APR | F95 |
|---|
| 0.5 | 67.7 | 37.8 | 58.7 |
| 0.6 | 65.0 | 35.1 | 62.3 |
| 0.7 | 61.2 | 31.0 | 67.6 |
| 0.8 | 57.0 | 28.1 | 79.5 |
| 0.9 | 52.0 | 25.2 | 92.8 |
| M | AUC | APR | F95 |
|---|
| 10 | 61.0 | 32.1 | 76.6 |
| 25 | 61.0 | 31.2 | 69.9 |
| 50 | 61.1 | 31.1 | 69.4 |
| 100 | 61.2 | 31.0 | 67.6 |
| 200 | 61.2 | 31.0 | 66.7 |
Stream ID-fraction sweep
| rho | Sem | Cov | Far | (MCM: Sem | Cov | Far) |
|---|
| 90% | 57.8 | 71.0 | 94.1 | 53.7 | 46.8 | 27.5 |
| 70% | 54.9 | 54.3 | 81.8 | 54.4 | 47.2 | 27.9 |
| 30% | 54.0 | 38.8 | 37.1 | 54.5 | 46.7 | 27.7 |
| 10% | 53.8 | 23.7 | 21.4 | 53.5 | 47.5 | 27.5 |
PROTON on six base scores
| Base score | Base (Cov) | +PROTON (Cov) | Δ | Base (Sem) | +PROTON (Sem) | Base (Far) | +PROTON (Far) |
|---|
| Max-Logits | 0.4 | 4.5 | +4.1 | 69.4 | 67.1 | 75.7 | 80.6 |
| Energy | 0.4 | 4.5 | +4.1 | 70.2 | 67.8 | 77.0 | 81.8 |
| GL-MCM | 46.7 | 60.6 | +14.0 | 54.3 | 55.1 | 29.2 | 93.5 |
| MCM | 46.6 | 61.2 | +14.6 | 54.2 | 55.3 | 27.6 | 93.9 |
| Entropy | 46.6 | 63.1 | +16.5 | 54.2 | 57.2 | 27.6 | 94.2 |
| TAG-MCM | 48.1 | 60.6 | +12.4 | 57.4 | 58.2 | 41.3 | 94.2 |
Broader baselines, including privileged ones
| Method | Cov AUC | Sem AUC | Far AUC |
|---|
| TAG-MCM | 48.1 | 57.4 | 41.3 |
| MCM (L=5) | 78.4 | 54.2 | 84.6 |
| kNN-ID | 99.9 | 67.8 | 100.0 |
| Mahalanobis-ID | 100.0 | 58.9 | 100.0 |
| PROTON | 61.2 | 55.3 | 93.9 |
| hier_l5+proton | 79.1 | 50.3 | 88.3 |
Prompt sensitivity
| Prompt set | ID acc | MCM Sem/Cov/Far | PROTON Sem/Cov/Far | PROTON+ Sem/Cov/Far |
|---|
| vanilla | 73.2 | 41.8 / 57.5 / 55.5 | 44.9 / 78.9 / 94.7 | 53.2 / 95.5 / 96.2 |
| l1 (default) | 69.2 | 54.2 / 46.6 / 27.6 | 55.3 / 61.2 / 93.9 | 61.4 / 93.9 / 91.5 |
| l5 | 55.3 | 54.2 / 78.4 / 84.6 | 50.3 / 79.1 / 88.3 | 49.9 / 88.5 / 96.8 |
Robustness extras
sigma2_0 sweep mean AUROC: 0.005: 72.7, 0.01: 72.0, 0.02: 70.1, 0.04: 62.0, 0.08: 58.8
misrouting (cov AUROC): 0.0: 61.2, 0.1: 65.7, 0.2: 69.3, 0.3: 72.8
overhead: PROTON 0.019 ms/sample, PROTON+ 0.153 ms/sample, bank 0.41 MB
wins vs MCM over stream orders: {'semantic': {'proton': '5/5', 'proton_plus': '5/5'}, 'covariate': {'proton': '5/5', 'proton_plus': '5/5'}, 'far': {'proton': '5/5', 'proton_plus': '5/5'}}
triage macro-F1: 0.8475022798662478
PROTON gate x bank-size grid (change in covariate AUROC over MCM)
| M=10 | M=25 | M=50 | M=75 | M=100 | M=150 | M=200 | M=500 |
|---|
| g=0.3 | 20.9 | 21 | 21.3 | 21.4 | 21.6 | 22.2 | 22.5 | 22.6 |
| g=0.4 | 20.9 | 21 | 21.3 | 21.4 | 21.6 | 22.2 | 22.5 | 22.6 |
| g=0.5 | 20.9 | 21 | 21.3 | 21.4 | 21.6 | 22.2 | 22.5 | 22.6 |
| g=0.6 | 18.3 | 18.5 | 18.9 | 19.2 | 19.6 | 20.1 | 20.4 | 20.4 |
| g=0.7 | 15 | 14.9 | 15.3 | 15.5 | 15.8 | 16 | 16.1 | 16.1 |
| g=0.8 | 10.8 | 10.8 | 11 | 11.3 | 11.5 | 11.6 | 11.7 | 11.7 |
| g=0.9 | 5.3 | 5.5 | 5.8 | 5.9 | 6 | 6 | 6 | 6 |
Pathology
Zero-shot OOD detection (static scores, PROTON, PROTON+)
| Method | Sem AUC | Sem APR | Sem FPR | Cov AUC | Cov APR | Cov FPR | Far AUC | Far APR | Far FPR |
|---|
| Max-Logits | 2.1 | 19.1 | 100.0 | 21.3 | 12.4 | 97.7 | 25.0 | 15.3 | 99.3 |
| Energy | 1.7 | 19.1 | 100.0 | 19.4 | 12.2 | 99.0 | 23.3 | 15.0 | 99.7 |
| GL-MCM | 56.7 | 40.5 | 94.0 | 67.2 | 30.9 | 77.7 | 67.2 | 29.1 | 75.2 |
| MCM | 59.8 | 42.4 | 90.3 | 66.1 | 29.8 | 85.2 | 67.1 | 31.5 | 77.2 |
| TAG-MCM | 48.7 | 33.1 | 97.2 | 48.0 | 19.4 | 92.0 | 45.1 | 18.7 | 95.0 |
| MCM (L=1) | 59.8 | 42.4 | 90.3 | 66.1 | 29.8 | 85.2 | 67.1 | 31.5 | 77.2 |
| MCM (L=5) | 15.3 | 20.4 | 99.8 | 64.6 | 30.7 | 77.3 | 61.1 | 26.3 | 87.3 |
| PROTON (ours) | 64.9 | 54.9 | 92.8 | 73.0 | 47.3 | 87.2 | 75.2 | 50.4 | 80.4 |
| Δ vs MCM | 5.0 | 12.6 | 2.5 | 7.0 | 17.5 | 2.1 | 8.1 | 18.9 | 3.2 |
PROTON mean ± std over 5 stream orders: semantic: 64.9±1.3, covariate: 73.0±3.7, far: 75.2±3.7
PROTON+ against PROTON (components, ID fraction, sensitivity, base scores)
| Variant | Sem AUC | Cov AUC | Far AUC | Sem FPR | Cov FPR | Far FPR | t_calib |
|---|
| MCM | 59.8 | 66.1 | 67.1 | 90.3 | 85.2 | 77.2 | nan |
| PROTON (paper) | 64.9 | 73.0 | 75.2 | 92.8 | 87.2 | 80.4 | 82 |
| + z-scored fusion | 77.7 | 84.7 | 82.8 | 78.6 | 57.0 | 55.1 | 82 |
| + consistency admission | 82.1 | 85.6 | 88.2 | 69.4 | 51.6 | 40.1 | 82 |
| + domain-anchored prototypes (PROTON+) | 82.4 | 85.6 | 92.1 | 67.7 | 50.7 | 35.0 | 82 |
| domain-anchored only | 68.7 | 74.0 | 81.1 | 92.8 | 86.9 | 73.7 | 82 |
| consistency only | 68.7 | 74.2 | 81.5 | 92.8 | 86.5 | 72.1 | 82 |
| PROTON+ max-fusion | 84.8 | 87.9 | 95.2 | 51.3 | 40.4 | 18.4 | 82 |
| PROTON+ rejection-driven alpha | 88.3 | 88.0 | 96.6 | 51.6 | 45.0 | 14.6 | 82 |
| S_proto only (PROTON bank) | 83.5 | 87.5 | 87.9 | 92.8 | 77.4 | 57.7 | 82 |
| S_proto only (PROTON+ bank) | 94.9 | 93.7 | 98.7 | 20.2 | 26.8 | 1.5 | 82 |
| PROTON, oracle ID-only bank | 70.6 | 74.4 | 84.4 | 92.6 | 86.6 | 79.6 | 82 |
| PROTON+, oracle ID-only bank | 82.1 | 85.8 | 91.4 | 67.1 | 50.2 | 36.3 | 82 |
| rho | PROTON+ Sem/Cov/Far | PROTON Sem/Cov/Far | MCM Sem/Cov/Far |
|---|
| 90% | 87.0 / 87.5 / 93.8 | 73.0 / 74.6 / 81.9 | 60.0 / 66.6 / 67.1 |
| 70% | 81.6 / 79.8 / 81.7 | 64.1 / 65.4 / 68.4 | 59.9 / 65.8 / 66.8 |
| 50% | 73.9 / 71.5 / 73.5 | 60.1 / 61.7 / 64.3 | 60.2 / 65.4 / 66.5 |
| 30% | 45.2 / 67.5 / 68.8 | 55.6 / 67.6 / 63.9 | 59.6 / 67.6 / 68.7 |
| 10% | 44.7 / 61.6 / 50.3 | 51.8 / 61.2 / 54.4 | 63.0 / 61.2 / 61.9 |
| gamma (PROTON+) | Sem | Cov | Far |
|---|
| 0.5 | 83.6 | 87.5 | 93.4 |
| 0.6 | 84.1 | 87.0 | 93.4 |
| 0.7 | 82.4 | 85.6 | 92.1 |
| 0.8 | 62.8 | 66.2 | 74.1 |
| 0.9 | 59.9 | 66.2 | 67.2 |
| kappa (PROTON+) | Sem | Cov | Far |
|---|
| 1.0 | 82.4 | 85.7 | 91.8 |
| 1.5 | 82.5 | 85.8 | 92.0 |
| 2.0 | 82.4 | 85.6 | 92.1 |
| 3.0 | 81.2 | 85.5 | 91.2 |
| 4.0 | 81.0 | 85.5 | 90.3 |
| Base score | Base Sem/Cov/Far | +PROTON+ Sem/Cov/Far |
|---|
| Max-Logits | 2.1 / 21.3 / 25.0 | 75.5 / 79.4 / 92.5 |
| Energy | 1.7 / 19.4 / 23.3 | 75.5 / 78.8 / 92.2 |
| GL-MCM | 56.7 / 67.2 / 67.2 | 83.5 / 88.0 / 94.8 |
| MCM | 59.8 / 66.1 / 67.1 | 82.4 / 85.6 / 92.1 |
| Entropy | 59.8 / 66.1 / 67.1 | 84.1 / 87.9 / 93.9 |
| TAG-MCM | 48.7 / 48.0 / 45.1 | 76.6 / 74.3 / 81.4 |
PROTON component ablations
| Variant | Cov | Sem | Far |
|---|
| MCM baseline | 66.1 | 59.8 | 67.1 |
| + Proto. (alpha=.5) | 75.6 | 66.5 | 78.8 |
| + Adapt. fusion | 73.6 | 65.0 | 76.7 |
| + Conf. gate (Full) | 73.0 | 64.9 | 75.2 |
| gamma | AUC | APR | F95 |
|---|
| 0.5 | 73.6 | 43.8 | 84.3 |
| 0.6 | 73.3 | 45.0 | 87.1 |
| 0.7 | 73.0 | 47.3 | 87.2 |
| 0.8 | 66.1 | 29.8 | 85.2 |
| 0.9 | 66.1 | 29.8 | 85.2 |
| M | AUC | APR | F95 |
|---|
| 10 | 72.8 | 47.3 | 86.6 |
| 25 | 72.9 | 46.8 | 87.5 |
| 50 | 73.0 | 47.1 | 87.3 |
| 100 | 73.0 | 47.3 | 87.2 |
| 200 | 73.1 | 47.5 | 87.1 |
Stream ID-fraction sweep
| rho | Sem | Cov | Far | (MCM: Sem | Cov | Far) |
|---|
| 90% | 73.0 | 74.6 | 81.9 | 60.0 | 66.6 | 67.1 |
| 70% | 64.1 | 65.4 | 68.4 | 59.9 | 65.8 | 66.8 |
| 30% | 55.6 | 67.6 | 63.9 | 59.6 | 67.6 | 68.7 |
| 10% | 51.8 | 61.2 | 54.4 | 63.0 | 61.2 | 61.9 |
PROTON on six base scores
| Base score | Base (Cov) | +PROTON (Cov) | Δ | Base (Sem) | +PROTON (Sem) | Base (Far) | +PROTON (Far) |
|---|
| Max-Logits | 21.3 | 29.7 | +8.4 | 2.1 | 12.7 | 25.0 | 31.4 |
| Energy | 19.4 | 28.1 | +8.7 | 1.7 | 12.4 | 23.3 | 29.8 |
| GL-MCM | 67.2 | 73.8 | +6.6 | 56.7 | 62.9 | 67.2 | 75.4 |
| MCM | 66.1 | 73.0 | +7.0 | 59.8 | 64.9 | 67.1 | 75.2 |
| Entropy | 66.1 | 76.4 | +10.3 | 59.8 | 69.8 | 67.1 | 78.2 |
| TAG-MCM | 48.0 | 61.0 | +13.0 | 48.7 | 58.4 | 45.1 | 59.5 |
Broader baselines, including privileged ones
| Method | Cov AUC | Sem AUC | Far AUC |
|---|
| TAG-MCM | 48.0 | 48.7 | 45.1 |
| MCM (L=5) | 64.6 | 15.3 | 61.1 |
| kNN-ID | 99.4 | 100.0 | 100.0 |
| Mahalanobis-ID | 99.8 | 100.0 | 100.0 |
| PROTON | 73.0 | 64.9 | 75.2 |
| hier_l5+proton | 61.5 | 32.6 | 65.7 |
Prompt sensitivity
| Prompt set | ID acc | MCM Sem/Cov/Far | PROTON Sem/Cov/Far | PROTON+ Sem/Cov/Far |
|---|
| l1 (default) | 59.3 | 59.8 / 66.1 / 67.1 | 64.9 / 73.0 / 75.2 | 82.4 / 85.6 / 92.1 |
| id_acc | 72.7 | 48.8 / 48.0 / 45.1 | 52.9 / 55.6 / 58.2 | 72.5 / 70.1 / 81.1 |
| vanilla | 64.5 | 64.3 / 53.4 / 72.5 | 66.7 / 58.9 / 82.2 | 75.8 / 72.9 / 92.5 |
| l5 | 65.5 | 15.3 / 64.6 / 61.1 | 32.6 / 61.5 / 65.7 | 15.4 / 73.4 / 81.5 |
Robustness extras
sigma2_0 sweep mean AUROC: 0.005: 75.4, 0.01: 73.9, 0.02: 71.0, 0.04: 68.6, 0.08: 68.1
misrouting (cov AUROC): 0.0: 73.0, 0.1: 74.2, 0.2: 75.7, 0.3: 75.2
overhead: PROTON 0.017 ms/sample, PROTON+ 0.096 ms/sample, bank 0.41 MB
wins vs MCM over stream orders: {'semantic': {'proton': '5/5', 'proton_plus': '5/5'}, 'covariate': {'proton': '5/5', 'proton_plus': '5/5'}, 'far': {'proton': '5/5', 'proton_plus': '5/5'}}
triage macro-F1: 0.3187348586668186
PROTON gate x bank-size grid (change in covariate AUROC over MCM)
| M=10 | M=25 | M=50 | M=75 | M=100 | M=150 | M=200 | M=500 |
|---|
| g=0.3 | 8 | 8.3 | 8.4 | 8.5 | 8.5 | 8.5 | 8.6 | 8.8 |
| g=0.4 | 8 | 8.3 | 8.4 | 8.5 | 8.5 | 8.5 | 8.6 | 8.8 |
| g=0.5 | 8 | 8.3 | 8.4 | 8.5 | 8.5 | 8.5 | 8.6 | 8.8 |
| g=0.6 | 7.7 | 7.8 | 7.8 | 7.8 | 7.9 | 8 | 8 | 8.2 |
| g=0.7 | 8 | 8 | 8.1 | 8.2 | 8.2 | 8.3 | 8.4 | 8.4 |
| g=0.8 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| g=0.9 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 |