ASSURE / EVIDENCE PACK
Proof you can inspect.
Exceptions you can follow.
Open each part of the example. Run context, source references and review history belong together—not in separate inboxes.
01 / SAMPLE RUN TV-DEMO-0042
A useful record includes the gap.
EXAMPLE / Return-leg audio missing · archive object available · exception open.
Run and timestamps
- Run ID
- TV-DEMO-0042
- Path
- LDN-004 / Cloud9 ringdown
- Start / UTC
- 2026-09-26T14:32:01Z
- Complete / UTC
- 2026-09-26T14:32:08Z
- Source linkage
- CALL-DEMO-42 → REC-DEMO-42 → ARCH-DEMO-42
Call legs and fingerprint
Outbound: audio present, known marker matched. Return: audio missing, marker not found.
The expected phrase and watermark are inspected per leg. Preserve a mismatch as a gap; do not let one matched channel conceal a missing speaker.
Archive and preservation
ARCH-DEMO-42 is retrievable in this example, but contains the same incomplete audio. Archive availability cannot repair capture quality. Retention policy needs review; legal hold is not tested in this run.
Exception and ownership
EX-DEMO-07 / OPEN. Missing return-leg audio. Owner: example recording operations. Next action: inspect capture mapping, record the change and authorize a new test. The original run remains in the history.
Reviewer sign-off
Example compliance reviewer · 2026-09-26 14:40:00 UTC. Decision: exception acknowledged, not closed. Acknowledgment records review; only a documented resolution and linked retest support closure. This sample contains no digital signature.
Artifact integrity
SHA-256 of the downloadable JSON bytes:
4fa6b40d6ab571630cd553a557edce2f4f41eec303fba6f64eae1664315afdf0This checksum detects changes to the sample file. It is separate from the known-audio fingerprint and does not authenticate the reviewer.
ENGINE EVIDENCE / FROM SCRIPT TO RESULT
Keep the words
behind the verdict.
A reviewer needs the expected conversation and the observed checks together.
Identity, answering party and both legs
Keep the unique call token, distinct per-leg script, expected answering party and observed answer classification. Record any human-answer courtesy message, call end, log entry and reschedule. Verification requires the expected answer as well as the configured content checks.
Recovered words, quality and timing
Link the source recording and recovered transcript to the expected tokens: correct leg, sequence and completeness. Retain the ground-truth script, archive or surveillance WER, STT confidence, MOS, jitter and loss. Preserve spoken/heard timestamps and mouth-to-ear latency per leg; for hoots, include each subscriber’s fan-out and delay result.
Surveillance and review routing
Pair each authorized trigger phrase with its expected rule and observed alert reference. Keep the token and classification tags that separate test traffic from human review. A missing alert remains an exception even when the audio was captured.
The existing sample above illustrates a missing return leg. The companion engine checklist below describes the additional evidence fields; it contains no measured scores or executed calls.
02 / REVIEW PRACTICE
A pass has a boundary.
Keep the tested path, configuration and time with every result. A passing synthetic call speaks for that test, not every production call between tests. An inconclusive access check is not a recording failure; a preserved file is not proof that both sides were captured.
Export the source identifiers with the findings. A reviewer should be able to trace a conclusion back to the recording, see excluded checks and follow an exception through a new run without losing the original evidence.
NEXT / DEFINE THE TEST
Start with the paths
you need to trust.
Bring your recording architecture. Let’s define the right test.