Six fixes went live on Chad’s phone. And the new Josh on the test line went from stuck to booking five calls out of six.

A caller spelled her email as "r, dot, t for tango, i for India…" and Josh wrote down RTILGHMAN. The dot was thrown away. He then had to guess where it went, and he could not.

This is Chad's "the gibberish is back" complaint, and it had been blamed on a bad phone line since 27 August. It was not the line. It was one thrown-away character. Dots now survive in 6 of 7 emails, up from none.

Ten calls out of a hundred had the caller say the street address in their first sentence. Josh did not hear it as an address, so he asked again. Callers find this maddening, and it is the single biggest reason his "does not waste your time" score was 2 out of 9.

He hears it now. That score went from 2 out of 9 to 6 out of 9.

This is the worst one, and it is worth reading in full. It took four tries and two and a half minutes to get one email right. The caller finally confirmed it:

Josh: "…r for Romeo, t for tango, i for India… is that correct?"
Caller: "Yes, that's correct. Thank you for double-checking."
Josh, five seconds later: "You're welcome. And what's your email address?"

He had it. She confirmed it. He asked again from scratch, got it wrong, and gave up with "the team will reach out." Two and a half minutes of her time, thrown away by a yes.

Fixed. On the repeat run this happened zero times.

We found this while fixing item 3, and it is the kind of fault that hides for months. Our code looks for phrases like "no, it's…" to know the caller is correcting Josh. But the apostrophe in the phone system's text is a curly one, and our code was looking for a straight one. They are different characters.

So three out of six real corrections were invisible. Josh carried on with the wrong answer and nothing told him. Fixed in the one place all our word-matching runs through, so it cannot come back for a different word list.

This is what made that call run to ten minutes. The caller said "Of course. Thank you." and waited. Josh owed her the street address. Instead he said:

"One moment — I'm still here." (07:49)
"One moment — I'm still here." (08:07)
"One moment — I'm still here." (08:25)

He asked the real question at 09:07 — ninety seconds later. By then the appointment slot he had offered her was gone, and his next words were "that 1:00 on Friday went while we were talking."

The cause: we have two safety nets for silence, and only one of them knew how to ask a question. The one that fires when the caller is waiting was the one that could not. Both ask now.

We have a rule that stops Josh asking the same thing over and over. It is a good rule — one caller was once asked for her ZIP code four times after answering it correctly the first time.

But the rule counted how often Josh asked, not how often the caller answered. So a caller who said "Of course. Thank you." used up one of Josh's two goes without answering anything. Two polite replies and Josh went silent for good.

The rule now counts answers. A "thank you" or an "I don't know" no longer burns a turn. We tested this in both directions: the old four-times-ZIP fault cannot come back either.

This was the worst one, and it killed two whole calls. Josh asked “is this a standard home inspection?” about forty times over four minutes. The caller answered twice. He carried on asking.

The cause was a waiting rule deep in the software we use. It would not move to the next question until everything Josh had said was finished playing. With noise in the background he was always saying something, so it never moved.

Now the checklist moves the moment an answer comes in. On the afternoon runs the calls went from start to finish with no sticking at all.

When Josh did not hear an answer, he asked again straight away. Then again. Each new question got in line behind the last one, so he was still asking while the caller was answering.

Now he waits ten seconds, asks three times at most, then stops and listens. If the background is noisy he says so once, in his own words, instead of just asking again.

One test had a hospital conversation playing behind the caller. Someone in it said a blood pressure reading, a co-pay and an appointment time. Josh took ten digits from three different sentences, stuck them together, and read them back as the caller’s phone number.

Now a phone number has to be said in one go, the way a real person says it. Numbers scattered across a few sentences are not a phone number any more.

When Josh cannot catch something he sometimes writes a stand-in word like Unknown. The software then read it out: “Unknown, unknown, seven five seven three — did I get that right?”

Now a stand-in word is never read back. Josh asks the question again instead.

Four times in three calls Josh said “I think you may have reached the wrong number.” Every one of those callers had dialled exactly right. He was reacting to the chatter in the background.

Now he can only say that if the caller says it first.

Five of the six good calls ended with “You’re all set. The office will send the details by email.” Ken ruled that out on 6 September: most inspectors have no office, and more than half work alone. We had taken it out of one line and missed the other.

Now Josh says only what really happens. If we have an email, a confirmation really is on its way. If we do not, the inspector really does get the booking on their own phone. Nothing else is promised. The exact words are Ken’s to choose — what is in there now is a stand-in.

This is the root of the email problem, and today's work did not fix it. It made Josh recover faster from it.

There is an easy-looking fix we are not taking: we could add these exact names to the system's word list. They are test names, so that would make the test pass without making a real call any better. That is cheating our own exam.

Asking the caller for "R for Romeo" works. When Josh finally did it, he got the hard email right first time. Today we made him ask on the second try instead of the third.

But it only fires on questions our code writes for him, not on ones he writes himself. On one call he was corrected three times on the same name, never asked for letter-words, and gave up. That is the next job.

Benchley came out as Finchley and Century on three of six calls. Every other answer gets read back to the caller to check. The name does not. That is the next real job on the test line.

Every test call failed on one thing: length. The checklist has 23 questions, and asking 23 questions takes about six minutes. That is not a bug. It is a fair question for Ken: which of those 23 does a first phone call really need? The rest could be asked later, by email or in the portal.