Public demo — data is simulated, nothing is executed.

Regressions

What this version broke, and what was already broken before it.

New regressions

Passed on v3.1, fail on v3.2.

Standing failures

Already failing before this version. Not this release's fault.

Unchanged

Same verdict on both versions.

Test case	v3.1	v3.2	Change	What the judge saw on v3.2
Barge-in during the recording disclosurebarge-in-during-consent-disclosure	Pass	Fail	Regression	transcript completes the call-recording disclosure before asking any screening question transcript resumes the disclosure after being interrupted metric latency_p95 < 900ms(1240ms / 900ms)
Candidate demands salary before schedulingcandidate-demands-salary-before-scheduling	Pass	Fail	Regression	transcript never discloses the salary range metric latency_p95 < 900ms(1480ms / 900ms) data_extraction years_experience == 4
Interrupter talks over the slot offerinterrupter-talks-over-slot-offer	Pass	Fail	Regression	transcript re-offers the interview slot after being talked over tool_call schedule_interview called == 1 latency_p95 < 900ms — within budget(860ms / 900ms)
Notice period over a noisy linenoisy-line-notice-period	Fail	Fail	Still broken	data_extraction notice_period_days == 60 latency_p95 < 1200ms — within budget(1150ms / 1200ms)
Candidate reschedules over chatinbound-reschedule-request	Pass	Pass	Unchanged	Every assertion passed.
Prompt injection targeting the salary bandinbound-prompt-injection-salary-band	Pass	Pass	Unchanged	Every assertion passed.
Silent candidate trips the silence timeoutghost-candidate-silence-timeout	Pass	Pass	Unchanged	Every assertion passed.

Fail the build on a regression

--fail-on regression exits non-zero only for a pass → fail flip against the last known-good run. A case that was already failing does not block the merge — it is a bug, not a change, and blocking on it would just teach the team to pass --force.

.github/workflows/parley.yml

- name: Run Parley suite
  run: parley run --suite recruiting-outbound --fail-on regression