| Transcription error |
Wrong location/event |
Original audio, alternatives, human confirmation |
Stratified word/error and field accuracy |
| Translation error |
Lost clinical or hazard detail |
Certified evaluation and source-language display |
Bilingual scenario review |
| Automation bias |
Human accepts wrong suggestion |
Evidence links, friction on low confidence, training |
Controlled human-factors trial |
| Emotional-presentation bias |
Distressed caller discounted |
Prohibit demeanor scoring; outcome audits |
Disparate-error analysis |
| False positive |
Unneeded or harmful dispatch |
Separate pre-alert from dispatch; policy gates |
False-dispatch rate |
| False negative |
Delayed/missed help |
Conservative escalation and human monitoring |
Missed-critical-event rate |
| Location spoofing |
Resources sent incorrectly |
Multi-source location and discrepancy alarm |
Adversarial location tests |
| Prompt injection in message/media |
Model changes policy or leaks data |
Treat all caller content as untrusted data |
Red-team suite |
| Surveillance expansion |
Secondary use chills help-seeking |
Purpose limitation and retention deletion |
Privacy impact assessment |
| Outage/vendor loss |
Calls cannot be handled |
Bypass to ordinary procedures |
Failover drill |
| Model update drift |
Previously tested behavior changes |
Version pinning and revalidation |
Regression suite |
| Unauthorized dispatch |
Unaccountable state action |
Strong identity, role controls, signed events |
Access-control penetration test |