What a fault-finding run will show: a worked example

This is an example of the page a run will produce. Five made-up agents get planted faults and a fault-free copy, and RingTru tests them blind. We have not run it yet, so every figure here is made up.

Reproduce it See how the faults will be sealed

Example data. Every number on this page except the tau-Voice figures is an illustration. RingTru has no completed test runs yet. Real runs will replace these numbers before any result is published.

Faults caught41 of 52The share of planted faults RingTru finds, on phone calls only.Example data

False alarms2 of 5Fault-free copies that RingTru wrongly flags at least once.Example data

Setup time25 minutesFrom signing up to the first scored call on a new agent.Example data

The five agents

All invented. Every business, person and number is made up.

Each agent, its planted faults, the faults caught and the false alarms on its fault-free copy
Agent	What it does	Planted	Caught	False alarms
Insurance claims	Claim status, identity check, handover	15	12	1
GP booking	Find, book, cancel, urgent symptoms	10	8	0
Energy billing	Account numbers, meter readings, bills	10	7	1
Shop returns	Order lookup, tracking, refund rules	9	7	0
After-hours messages	Take a message, page the on-call engineer	8	7	0

Example data The false alarm column counts flags raised on the fault-free copy of each agent.

Where misses would show

Easy faults show up in most calls. Hard ones need the right caller on the right day.

By fault difficulty

Easy15 of 16

Medium16 of 20

Hard10 of 16

Example data

By call stage

Opening7 of 8

Identity check11 of 12

Lookup10 of 14

Action8 of 10

Handover5 of 8

Example data

Text, browser audio or phone?

The same agents, tested three ways in a real run. Some faults only show up once audio travels down a phone line. Text catches the most, but it cannot hear a misheard postcode.

Text vs voice testing, the guide

Text46 of 52

Browser audio42 of 52

Phone41 of 52

Example data

How the faults will be sealed

You should not have to take our word for it. The order of events will be public, and after publication the fault list will be released so anyone can hash it and compare.

Step 1. We write the list of faults for each agent and save it in a private store first.

Step 2. We publish a digital fingerprint of that list. Change one letter and the fingerprint changes.

Step 3. Then we build the agents. The person who sets up RingTru never sees the list.

Step 4. RingTru runs blind. Only then do we open the list and score every call.

Locked before any testing begins

Fingerprint of the fault list: 9f2c…e81a

Not a real value. Example data.

What other public work found

Voice is harder than text, and published work agrees. We credit every source and link to it.

tau-Voice

Text 85 percent, voice 31 to 51 percent.

From the tau-Voice paper, arXiv 2603.13686, 2026. This is their result, not ours.

Read the paper

Download or reproduce

Take the agents and run them yourself. Everything is made up. The files go live with edition one. Open data

Planned

The full report

Every finding, every call, every miss.

Planned

The five agents

Their prompts and settings, so you can run them yourself.

Planned

The call recordings

All made-up callers. No real person is in any recording.

Other benchmarks

A second line of work measures speech recognition alone.

Test your own agent the same way.

Join the waitlist and we will run the same method on yours.

To join the waitlist, email us at greg@gme.solutions.

Rather talk first? Book a call