The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
AANA Cross-Domain Action Gate Blind Validation v2
This repository publishes a larger frozen-adapter validation for six AANA agent-action domains: devops, education, finance, HR, legal, and pharma.
The v2 set is separate from GAP and from the prior 72-row validation. It contains 360 rows: 30 safe and 30 unsafe rows per domain. Rows include near-miss safe cases with risky words and varied unsafe phrasing. The gate receives only action, tool metadata, constraints, evidence, and domain. Hidden labels are used only after scoring.
The adapters were frozen before this v2 run. No post-run tuning was applied.
Result
Rows: 360
Domains: 6
Safe rows per domain: 30
Unsafe rows per domain: 30
| Path | Accuracy | Block recall | Block precision | Safe allow | Route quality |
|---|---|---|---|---|---|
| Base accept-as-safe | 50.00% |
0.00% |
0.00% |
100.00% |
50.00% |
| Generic AANA | 50.00% |
100.00% |
50.00% |
0.00% |
50.00% |
| Frozen six-domain AANA | 82.50% |
66.67% |
97.56% |
98.33% |
82.50% |
Confusion Matrix
| Path | TP | TN | FP | FN |
|---|---|---|---|---|
| Generic AANA | 180 |
0 |
180 |
0 |
| Frozen six-domain AANA | 120 |
177 |
3 |
60 |
Domain Breakdown
| Domain | Accuracy | Block recall | Safe allow | TP | TN | FP | FN |
|---|---|---|---|---|---|---|---|
| DevOps | 75.00% |
50.00% |
100.00% |
15 |
30 |
0 |
15 |
| Education | 75.00% |
60.00% |
90.00% |
18 |
27 |
3 |
12 |
| Finance | 85.00% |
70.00% |
100.00% |
21 |
30 |
0 |
9 |
| HR | 80.00% |
60.00% |
100.00% |
18 |
30 |
0 |
12 |
| Legal | 90.00% |
80.00% |
100.00% |
24 |
30 |
0 |
6 |
| Pharma | 90.00% |
80.00% |
100.00% |
24 |
30 |
0 |
6 |
Interpretation
The generic adapter caught every unsafe action but blocked every safe action.
The frozen six-domain adapters restored safe allow to 98.33% and raised
accuracy by 32.50 points, but missed 60/180 unsafe actions under varied
unseen phrasing.
This is useful negative signal. The next adapter work should target recall in devops, education, and HR while preserving the high safe allow rate.
Files
dataset.json: the 360-row blind validation dataset.report.json: full generic-vs-domain results.summary.json: compact publication summary.aana_cross_domain_action_gate_v2_eval.py: reproducible runner.
Links
- Try AANA: https://huggingface.co/spaces/mindbomber/aana-demo
- AANA model card: https://huggingface.co/mindbomber/aana
- v1 action-gate validation: https://huggingface.co/datasets/mindbomber/aana-cross-domain-action-gate-validation
- GAP domain adapter ablation: https://huggingface.co/datasets/mindbomber/aana-gap-domain-adapter-ablation
Scope
No state-of-the-art, official leaderboard status, production-readiness, or agent-safety guarantee claim is made.
- Downloads last month
- 37