Work
Evidence Desk

Product study · Evidence Desk

A flag is a starting point. The evidence makes it useful.

Review the moment, understand its context, and decide what deserves a response.

Inspect a finding See the design decisions

Findings

Search findings

F-101High urgency

Repeated request for an update

Jordan Ellis · Visit 318 · with Sam

Recorded words

C-2041S310:42:07words 0–12transcript v2

I called at eleven and nobody had an update, so I'm calling again.
Customer, recorded 10:42:07

One turn before and after

CustomerS2 · 10:41:58
Hi, it's Jordan Ellis. I dropped the car off this morning for the brake noise.
CustomerS3 · 10:42:07
I called at eleven and nobody had an update, so I'm calling again. Can someone just tell me when it will be ready?
StaffS4 · 10:42:19
Let me check with the technician, and someone will call you back.

Model summary · interpretation, not a quote

The customer seems frustrated that no one has said when the car will be ready, and is calling for the second time.

Not known now

Whether anyone called back after this recording is not known here. The recording only shows Thursday morning.

Recorded: Thu 15 Oct, 10:42
Analyzed: Thu 15 Oct, 18:05
Audio: Not available in this demo

Gave a specific time for the next update0.0% · 0 of 3

Sam, Thursday: 0 of 3 assessed conversations. A real zero: there were conversations to judge, and none showed it.

The denominator counts observed and not observed only. Needs review, not relevant and not yet assessed stay out of it and are listed instead.

The working desk, with synthetic data. Choose another finding, or review this one; nothing is sent or stored beyond this page.

Overview

Product studyDesign and build

The problem

Teams get summaries and flags about their conversations, but a flag on its own does not say what was said, when, or whether it still matters. Act on one blindly and you chase the wrong thing; set it aside and a real concern waits.

The change

Evidence Desk puts a defensible starting point under every flag: the exact recorded words, the turns around them, what is not known now, and a deliberate decision that is kept on record.

My part

I designed and built the workspace and its rules as a working demo. It reuses two rules from a conversation dashboard I built earlier: quotations are cut from stored words, and unknown is never shown as zero. The queue, the review record and its history are new work.

Three decisions

Decision one

Source before recommendation

The recorded words come first and the model’s summary second, labelled as an interpretation. A finding cites a segment and a word range; the desk cuts those words from the stored transcript and shows one turn either side. If the segment is missing, the range is malformed or the transcript changed after analysis, nothing is quoted and confirming is switched off.

How a reference becomes a quotation

The model returns a segment id and a word range, never the words.
Words are runs of non-space characters counted from zero; the start is included and the end is not.
Checks run in order: processing finished, transcript present, same transcript version, segment exists, whole numbers, a non-empty range inside the segment.
The quote is a slice of the stored text, spacing and punctuation included. A summary is never turned into a quotation.

What the model returned

{ "segment": "S3", "start": 0, "end": 13 }
analyzed against transcript v2; stored: v2

What the desk shows
C-2041S310:42:07words 0–12transcript v2
I called at eleven and nobody had an update, so I'm calling again.
Customer, recorded 10:42:07

Choose a reference. Only one that passes every check becomes words on screen.

Decision two

Unknown stays different from zero

A rate shows its denominator. When there were conversations to judge and none showed the behaviour, that is a real zero. When nothing could be judged, the desk says Unknown and lists why, instead of printing a zero that reads as a failure.

What counts in the denominator

The denominator is observed plus not observed.
Needs review, not relevant and not yet assessed stay out of it, and are listed beside the rate instead.
With nothing in the denominator the value is unknown, never zero percent.

Gave a specific time for the next update
0.0%
0 of 3 assessed · Sam, Thursday
A real zero: there were conversations to judge, and none showed it.
Handled a request to escalate
Unknown
0 of 0 assessed · Sam, Thursday
Not zero: nothing could be judged. 2 need review, 1 not assessed yet, 4 not relevant.

The same signal shape, two different truths: one judged and absent, one not judged at all.

Decision three

Reviewing is different from resolving

Confirming a finding records that the words support it. It does not mean anyone was helped or called, and the desk never marks a person contacted. Validity, disposition, recording time and the present situation are separate fields, and every save adds a revision to a history that is never rewritten.

How a review is recorded

Each save names the revision it was made against. A stale revision is refused, and the draft stays on screen.
Rejecting needs a reason. The original finding and its source version are never edited.
Saving the same review twice adds nothing. A failed save keeps the draft and offers a retry.
A next-step note is a draft for the team. Nothing in the desk sends a message or contacts anyone.

F-106 · Promised call-back by noon

Field	What changes it	Now
Validity	A review	Confirmed, rev 2
Disposition	A review, chosen separately	Follow up
Recorded	Nothing; it is when the words were said	Thu 15 Oct, 09:12
Contact since	Nothing in the desk	Not recorded
Situation now	Someone checking, outside the desk	Unknown

Review history

Confirmedrev 2
Reviewer A · Fri 16 Oct, 08:41 · Follow up
The promise is in the words. The missing call is only what the recordings show.
Draft note, not sent Ask Sam whether the call was made from a mobile.
Confirmedrev 1
Reviewer A · Fri 16 Oct, 08:34 · Open
The promise is in the words.

A reviewed finding: its validity and next step are on record; whether anyone was contacted is still not.

The tradeoff

Blocking is strict. A finding with a broken reference may still describe something real, but it cannot be confirmed until it is analyzed again. I chose a slower queue over a quote nobody can check; rejecting stays open, with a reason, so a broken finding can still be cleared.

Where it stands

Try these interactions in the demo:

Built
Quotes are exact slices of stored words; malformed, missing or out-of-date references show no evidence.
Built
Rejecting requires a reason and keeps the original finding and its history.
Built
A failed save keeps the draft; reviewing never marks anyone contacted.
Built
Back, Escape and Reset return to a known place, on desktop and phone.

Open the full demo