cavendish
QueueExperimentsScorecard
Type 3SignalAbandoned

Voice agent latency floor for AU telco

Voice and Vision
  1. proposed
  2. voting
  3. running
  4. measuring
  5. concluded

Preregistration · v1 · 25 May 2026 · immutable after start

An extension creates a new version rather than editing the old.

Hypothesis

At least one realtime voice endpoint reachable from Sydney holds median turn latency under 800ms across 40 scripted tier-1 telco triage calls over a partner SIP trunk.

Kill condition

Kill if no endpoint is under 1,000ms median after 20 calls each, or if caller-abandonment on the scripted calls exceeds 30% on every endpoint.

MethodThree endpoints (one AU-hosted, two US-hosted), 40 scripted calls each with two scripted callers, over the partner's SIP trunk. Measure turn latency p50/p95, barge-in handling and abandonment. Accent coverage: three AU accents per script.
Expected cost$450 in minutes and tokens, two people for three weeks, one partner trunk
Expected duration3 weeks

Result

Abandoned

Abandoned, stopped without result. The partner withdrew the SIP trunk on day 24 with 11 of 120 planned calls measured on one endpoint. The 710ms median is a tried-tier observation and no claim is made from it. The question is back in the pile with the harness ready.

Running notes · fed from harness sessions and by the pair

AW
manual · Adam Witanowski · 3 Sep 2026· ⌘↩ to post
  1. manual15 Jul 2026PR Priya Raman

    Abandoned. 11 of 120 calls, one endpoint. No result; numbers logged at tried tier. Question returns to the pile; g-voice-ivr-replacement written.

  2. manual1 Jul 2026TO Tom Okafor

    Trunk reclaimed by the partner's network team for a production change. Re-booking falls inside their change freeze to mid-August.

  3. harness24 Jun 2026PR Priya Raman

    harness live; 11 calls AU endpoint; p50 710ms p95 1,240ms; barge-in ok

  4. manual25 May 2026PR Priya Raman

    Predicted: AU endpoint passes, US endpoints do not, confidence 0.6. Trunk confirmed verbally; not in writing.

Harness notes are auto-captured from Claude Code sessions: model, date, commit, session reference. Never the transcript, code or paths.

Artifacts

repovoice-latency harness + call scriptsrepo
meas11 calls, AU endpoint only (tried tier, not a result)measurement