# Review a call properly

Read the outcome first, the wire timings second, the review findings third and the transcript last. The review separates findings a setting can change from findings no setting reaches, and only the first kind is worth your afternoon. Listening to a call end to end before reading any of that is the slowest way to find out what went wrong.

- **Status:** Available
- **Audience:** both
- **Channels:** phone
- **In the app:** #/calls
- **Last verified:** 2026-09-10
- **Canonical:** https://connectbyjbrh.com/docs/how-to/review-a-call/

## Why the transcript comes last

A transcript tells you what was said. It is almost useless for telling you how the call felt, because a transcript row arrives when a sentence is complete, not when the caller started waiting. An early version of the timing display was built from those rows and showed eight milliseconds beside replies the caller had waited several seconds for.

Timing is measured on the wire instead. A clock in the voice worker times the caller's last word — the moment the caller's speech state stops — to the first audio coming back, and posts one timing event per reply. A separate event records how long the voice kept talking after the caller spoke over it. Those two numbers describe the experience; the transcript describes the content.

## The reading order

1. Open the call and read the outcome and the hang-up attribution.
   - Result: Who ended the call, and how, frames everything else. A caller who hung up during a reply is a different investigation from a call that ran to a natural close.
2. Read the reply timings.
   - Result: The review treats 2.5 seconds as the target and 4 seconds as slow. A call whose replies sit around the target has no latency problem regardless of how it sounded to you.
3. Read the barge-in events.
   - Result: These say whether the caller was talked over. An interruption the voice kept talking through for more than two seconds counts as ignored, and ignored interruptions are the finding customers actually complain about.
4. Read the review findings, then the settings snapshot beside them.
   - Result: The snapshot is the configuration that actually ran, captured when the session started. It is what makes a finding actionable — you are changing the setting that was in force, not the one saved now.
5. Now read the transcript, looking only for the moments the findings pointed at.
   - Result: You arrive at the text with a question rather than reading eight minutes hoping something stands out.

## Findings a control can change, and findings it cannot

This split is the most useful thing on the page. The review produces deterministic findings — slow replies, long gaps, ignored interruptions, false stops, repetition, stock phrases, the wrong script, a caller who was not heard — and each carries the setting change that addresses it. But when the model's own first token accounts for sixty per cent or more of the wait, the finding says so and offers no control, because there is none to turn.

| Finding | Kind | First move |
|---|---|---|
| Ignored interruption | Control | Lower the interruption threshold in seconds or words |
| False stop — the voice stopped for a noise | Control | Raise the threshold, or enable false-interruption resume |
| Long gap before a reply | Control | Shorten the instructions; prompt size moves first-token time measurably |
| Slow reply, model-bound | Model-limited | Nothing to turn. Recorded and counted, and it does not lower the score |
| Stock phrases or repetition | Control | Behaviour text — the opening, closing and objection lines |
| Caller not heard | Control | End-of-speech silence and noise cancellation |
| Wrong script | Control | The profile that resolved; check the snapshot before assuming |

> **Note** Two scores are shown. The headline score excludes model-limited findings so that it measures what you can act on; the older all-inclusive number is kept beside it so a fall in one and not the other is legible.

## The three findings worth acting on first

1. **Ignored interruptions.** A caller talked over for more than two seconds concludes nobody is listening, and that judgement survives the rest of the call. Fix this before anything cosmetic.
2. **A caller who was not heard.** Silence handling and end-of-speech detection failing together produce a call that reads fine in the transcript and was miserable to be on.
3. **Long gaps that are not model-bound.** These are the ones a setting genuinely fixes — usually instruction length. If the finding says model-bound, move on: the floor on this path is the model's first token plus its end-of-turn detection, and the best measured call sat at 3.3 seconds median rather than the one or two seconds people expect.

Everything else — phrasing, warmth, pacing — is worth changing only after those three are clear, because they change how the call sounds and the first three change whether it worked.

## The model's opinion, and what to do with it

The review also includes the model's assessment of delivery, and it is labelled as the model's rather than presented as a measurement. Read it as a second opinion on tone, not as evidence. When it disagrees with a deterministic finding, the deterministic finding is the one with a clock behind it.

> **Careful** One review is one call. A single slow reply is noise; the same finding across a week of calls is a setting. Change one thing at a time and read the next few reviews before changing another, or you will not know which change did what.

## Questions

### Can I listen to the call itself?

Call recording is a foundation capability: the ability is asked of the provider rather than assumed, and it is not enabled on the carrier in live use. The review, the timings and the transcript are what a call leaves behind, and between them they answer most questions a recording would.

### The review says the reply was slow but it felt fine. Which is right?

Both, usually. The clock measures from the caller's last word to the first audio back; a filler or a short acknowledgement can make a long wait feel shorter without shortening it. If the number is above four seconds it is worth fixing even when nobody complained.

### Does the review judge the settings I have now?

No, and that is deliberate. The settings that resolved for that session are snapshotted onto the call, so a review written last week still describes the configuration that produced it rather than whatever has since been saved.

## Related

- [Call quality review](https://connectbyjbrh.com/docs/phone/voice-lab-review/)
- [Reply latency on a call](https://connectbyjbrh.com/docs/phone/latency/)
- [Barge-in](https://connectbyjbrh.com/docs/phone/barge-in/)
- [What a setting cannot change](https://connectbyjbrh.com/docs/phone/model-limits/)
- [Measuring conversational latency correctly](https://connectbyjbrh.com/research/measuring-voice-latency/)
- [Findings a setting cannot fix](https://connectbyjbrh.com/research/model-limited-findings/)
- [Use a different voice for each line](https://connectbyjbrh.com/docs/how-to/use-a-different-voice-per-line/)

## What this page is based on

- Connect source pack: phone (docs-source/sources/PHONE.md §6 — latency measurement and review)
- Connect source pack: phone (docs-source/sources/PHONE.md §12 — what is not available)
- Connect capability registry (docs-source/facts.py — MEASURED, call_recording)