d028 Smith-Waterman — public-dataset evaluation
Evidence that d028 moves the needle on standard public benchmarks, captured as
the first checkbox in project_evaluation_matrix.md.
Both configurations use the same 26-regex baseline with d022_semantic_classifier
off. The only independent variable is d028_sequence_alignment (enabled in treatment,
disabled in control). threshold=0.7, scan result counted as a detection if
action in {block, flag} or risk_score >= 0.5, matching tests/benchmark_public_datasets.py.
deepset/prompt-injections (116 samples)
| Config |
TP |
TN |
FP |
FN |
Precision |
Recall |
F1 |
Accuracy |
FPR |
Speed |
| Control (26 detectors, no d028) |
1 |
56 |
0 |
59 |
1.000 |
0.017 |
0.033 |
0.491 |
0.000 |
244/s |
| Treatment (27 detectors, with d028) |
14 |
56 |
0 |
46 |
1.000 |
0.233 |
0.378 |
0.603 |
0.000 |
88/s |
| Delta |
+13 |
— |
+0 |
-13 |
— |
+21.66 pp |
— |
+11.20 pp |
+0.00 pp |
— |
leolee99/NotInject (339 samples)
| Config |
TP |
TN |
FP |
FN |
Precision |
Recall |
F1 |
Accuracy |
FPR |
Speed |
| Control (26 detectors, no d028) |
0 |
336 |
3 |
0 |
0.000 |
0.000 |
0.000 |
0.991 |
0.009 |
242/s |
| Treatment (27 detectors, with d028) |
0 |
326 |
13 |
0 |
0.000 |
0.000 |
0.000 |
0.962 |
0.038 |
81/s |
| Delta |
+0 |
— |
+10 |
+0 |
— |
+0.00 pp |
— |
-2.95 pp |
+2.95 pp |
— |
microsoft/llmail-inject-challenge (Phase1, 1000 subset) (1000 samples)
| Config |
TP |
TN |
FP |
FN |
Precision |
Recall |
F1 |
Accuracy |
FPR |
Speed |
| Control (26 detectors, no d028) |
978 |
0 |
0 |
22 |
1.000 |
0.978 |
0.989 |
0.978 |
0.000 |
73/s |
| Treatment (27 detectors, with d028) |
980 |
0 |
0 |
20 |
1.000 |
0.980 |
0.990 |
0.980 |
0.000 |
4/s |
| Delta |
+2 |
— |
+0 |
-2 |
— |
+0.20 pp |
— |
+0.20 pp |
+0.00 pp |
— |
ai-safety-institute/AgentHarm (352 samples)
| Config |
TP |
TN |
FP |
FN |
Precision |
Recall |
F1 |
Accuracy |
FPR |
Speed |
| Control (26 detectors, no d028) |
44 |
120 |
56 |
132 |
0.440 |
0.250 |
0.319 |
0.466 |
0.318 |
86/s |
| Treatment (27 detectors, with d028) |
44 |
120 |
56 |
132 |
0.440 |
0.250 |
0.319 |
0.466 |
0.318 |
24/s |
| Delta |
+0 |
— |
+0 |
+0 |
— |
+0.00 pp |
— |
+0.00 pp |
+0.00 pp |
— |
ethz-spylab/agentdojo (v1.2.1) (132 samples)
| Config |
TP |
TN |
FP |
FN |
Precision |
Recall |
F1 |
Accuracy |
FPR |
Speed |
| Control (26 detectors, no d028) |
17 |
86 |
11 |
18 |
0.607 |
0.486 |
0.540 |
0.780 |
0.113 |
163/s |
| Treatment (27 detectors, with d028) |
18 |
83 |
14 |
17 |
0.562 |
0.514 |
0.537 |
0.765 |
0.144 |
33/s |
| Delta |
+1 |
— |
+3 |
-1 |
— |
+2.86 pp |
— |
-1.51 pp |
+3.09 pp |
— |
synthetic/indirect-injection (80 samples)
| Config |
TP |
TN |
FP |
FN |
Precision |
Recall |
F1 |
Accuracy |
FPR |
Speed |
| Control (26 detectors, no d028) |
24 |
50 |
0 |
6 |
1.000 |
0.800 |
0.889 |
0.925 |
0.000 |
71/s |
| Treatment (27 detectors, with d028) |
30 |
50 |
0 |
0 |
1.000 |
1.000 |
1.000 |
1.000 |
0.000 |
8/s |
| Delta |
+6 |
— |
+0 |
-6 |
— |
+20.00 pp |
— |
+7.50 pp |
+0.00 pp |
— |