Frontend Runtime Evaluations

The deterministic frontend evaluation suite protects the runtime boundaries that qwen-audio-agent can guarantee independently from a Realtime model or provider:

Dimension	Invariant
Routing	Search, URL Fetch, optional Knowledge retrieval, backend Task, and client control remain separate capability-gated tools.
Citation	Sources receive stable per-turn IDs, duplicate URLs keep one identity, unsafe URLs are dropped, and final citations are consumed once.
Interruption	Audio and transcripts arriving after a user interruption are suppressed, while the interruption event is projected once.
Duplicate speech	Duplicate completion signals coalesce before presentation, and repeated acknowledgements do not create another presentation for the same Task.
Prompt injection	Search and Knowledge content stays untrusted data, carries an explicit notice, and cannot change the registered tool surface.

Run the suite with:

npm run eval:frontend

Use npm run eval:frontend -- --json for a machine-readable report. The same evaluations also run from the server test suite and therefore from release:check and CI.

The suite drives production Runtime components; it does not contain a second implementation of routing, presentation, or citation behavior. It makes no external request and does not call a language model, so results are fast and repeatable on every supported operating system.

These checks do not claim to measure provider-specific semantic quality, such as how often one model selects the ideal tool for a natural-language request. Such live-model evaluations may reuse the same dimensions, but stay outside the deterministic release gate because they require credentials, cost, and statistical thresholds.