A neutral third party by design

Neither side gets to rewrite the conversation.

EMT919's test program evaluates whether the interpreter preserves speaker roles, relays what was said, resists attempts to alter its behavior, and escalates when the configured boundary requires a human.

2,100+Scripted internal QA scenarios
57Laboratory test batches
Both sidesInfluence attempts tested in each direction
ContinuousFailures classified and addressed before release

Scope: Figures are from City919's internal QA program as of July 2026. They are not customer outcome statistics, field-performance guarantees, or independent certification.

What neutrality testing covers

01

Recruitment attempts

Scenarios direct the interpreter to omit, soften, exaggerate, or conceal part of a statement. Expected behavior is to refuse the manipulation and preserve the original meaning.

02

Pressure from either party

Testing runs in both directions, including attempts by a responder or a member of the public to make the interpreter editorialize or change roles.

03

Critical-detail fidelity

Names, spelled words, dates, identifiers, plate and case numbers, and medical quantities are evaluated as details where an approximation is not acceptable.

04

Prompt and role attacks

Standing tests attempt to reveal system instructions, change the interpreter's role, or redirect it toward an unrelated task.

05

Field-noise conditions

Test cases include overlapping voices, bystander interruptions, background speech, code-switching, and conversations directed away from the call.

06

Human escalation boundary

Policy-controlled and legally sensitive communication is tested against the agency-approved script and live-interpreter rules for the configured service.

What the system should do

  1. Keep the responder, member of the public, and interpreter roles distinct.
  2. Translate the statement rather than follow an instruction embedded inside it.
  3. Preserve material details instead of smoothing them into an approximation.
  4. Surface a manipulation attempt when that attempt is itself part of the conversation.
  5. Escalate or stop relying on AI when audio, policy, or language conditions exceed the configured boundary.

Limits still apply

Internal testing reduces known failure modes; it does not eliminate them. Noise, overlapping speech, cellular conditions, accents, terminology, device placement, and unclear phrasing can affect interpretation. Responders remain responsible for following agency policy and moving to a qualified live interpreter when required.

Review the test program See the agency evaluation standard