The web chat could not finish a booking. Three faults were in the way. All three are fixed, and the phone got better too.

Ken's rule is no email, no booking. So the email is required. But the software only counted the caller's answer if Josh's question matched one of eight set phrasings. “What is the best email for the confirmation?” did not match. Neither did “Best email?” So the caller answered, nothing recorded it, and Josh asked for the email again and again instead of booking.

Fixed. An email address is now recognised by its shape, wherever it turns up. Ten cases were checked both ways, including five that must not count, like an agent's address the caller mentions in passing. This is the same family as the street address fault fixed on 10 September. A few other required fields have the same gap. Each needs its own check before it gets the same fix.

At the end of a booking Josh reads back the name, address, size, year, foundation, who is attending, the package, the price, the day, the time and the inspector. Then he has to save. His replies had a size limit of 400 tokens, and the read-back alone filled it. He was cut off before he could save. We measured it on all four pages: four booking conversations, four cut-off replies, and the save never called. One was cut off at the exact words “Perfect. Let me get you booked.”

The limit is now 900. And a reply that does get cut off carries on once, with one instruction: if you were about to save, save now, before any more words. The limit is a safety net. It is not what keeps Josh brief; his instructions do that.

Earlier today we added a check that asks Josh to put it right when he claims a booking that did not save. He said it again on the retry, and there was no check left. So asking is not enough. Now the claiming sentence is removed before the reply goes out. The read-back stays, because the read-back is how a caller catches a wrong detail.

It cannot damage a real booking. It only runs until the CRM has confirmed the save. After that, every confirmation Josh makes goes out untouched. This is the phone's rule, brought over to the web.

Running the same booking three times, it died once at an ordinary turn (“Yes, that works.”). Nothing was wrong with the conversation. The shape of it was an overloaded call, not a bad request. And from outside, nobody could tell: the error message gave no code.

Now a passing fault like that is retried once, with a short wait. The retry only happens if nothing has been saved in that turn. Booking twice is worse than failing once, so a turn that already saved is never replayed. The error now carries a status code, so the next one can be diagnosed from outside.

This is the honest part. All 473 chats were graded on what Josh said. Not one check asked what the system did. A booking conversation that quietly never saved came back clean, run after run. The booking test was also only seven turns long, not enough to reach a save even on a healthy day.

Fixed. A new check fires when a booking case ends with no contact saved, and says whether the save was even attempted. Another reads whether a reply was cut off. The booking cases are now 18 turns of a patient caller, on all four pages, plus one that hands over everything at once.

Seven errors in one test run, all in a row. It was our own update landing mid-run. Every update restarts the web app, and anyone typing at that moment sees “Connection problem — please try again.” It happens several times on a normal working day. We had written this down for the phone and never for the web, where it happens far more often.

Not fixed, on purpose. The obvious fix is a gentle restart, but that only works in a mode the app does not run in. The other fix, a quiet retry in the chat window, risks booking the same job twice. Both need a decision from Dil, not a quick edit. The tester now retries on its own, so an update is no longer reported as seven faults.

Ken, 7 September: “we can't take the inspection without your email — have a blessed day. Boom, hang up… that's not even a qualified lead.” Dil confirmed it on 10 September. The web and the test line already had it. The phone did not: it booked a caller who refused her email on five runs in a row.

Now it is live. A caller is asked at most twice. Then Josh says Ken's sentence, saves nothing, and the line drops. A caller who was asked and simply never answered still books, with no email promised. That is Ken's own carve-out: refusal, not absence.

Proved on runs 7 and 8. Nobody who refused an email was booked. One caller who was booked on a refusal on run 7 was correctly turned away on run 8.

Run 8 showed the next problem. Josh saw the refusal, and then ignored the instruction to say the sentence word for word. On three calls out of three he improvised instead:

“Here's what I can do… your inspector will reach out to confirm your email.”
“No problem at all — we'll confirm your email… let me get you booked.”

Nothing was booked, because the save gate held every time. But the caller heard three different things, none of them Ken's. So the sentence is no longer trusted to Josh. Once the refusal is counted, the software speaks Ken's sentence itself, once, and drops anything else Josh writes for the rest of the call. Eight lines that said “the team will…” now say “your inspector will…”, which is what Ken asked for on 6 September.

Two hundred generated calls with a caller who answers whatever Josh asks. Two faults fell out. First, seven yes-or-no questions (is there a pool, any outbuildings, any add-ons, and so on) treated a bare “No.” as a filler word, not an answer. So the caller answered, nothing recorded it, and Josh asked again. Second, the add-ons question could be asked again and again, and it blocked every question behind it. The email was one of them.

Both fixed. Failures on those 200 calls went from 650 to 60, and the 60 that remain are one deliberately unfair case. Live on the phone since the morning's deploys.

The software that says numbers out loud was pushed through 120 ZIP codes in six written forms, every number sentence Josh has ever produced, and every reply chopped up 28 different ways. Four faults:

The last one is the serious one. Nothing was lost; the figure was simply wrong, and only when the network split the reply in the wrong place. So it was right nine times in ten and every existing test passed. None of the four would have been found by a phone call. All four are live since the morning's deploys.

Six test calls were written to prove the morning's fixes, on the client lines, because Ken's standing rule sends behaviour tests there. On the closing read-back Josh said, in his own audio:

“Mike Chen, twelve Oak Street, Pearland, seventy seven thousand five hundred and eighty four.”

The morning fix caught the ZIP written in digits after a comma. Nobody checked the ZIP written in words after a city. They are two separate doors in the same file, and both have to be opened every time. Fixed the same day, with 26 checks.

And most of the “one pass out of six” was our test file, not Josh. It held five Texas cases and one Delaware case, and the test company dials one number per file. On Chad's line the Delaware caller was correctly refused as out of area. Josh was right every time. The file is now split into one per firm, and a new check refuses any test file that names two.

Ted's suite is now built from Ted's own facts, not Chad's with the town swapped: his towns, his Saturday-morning rule, his $650 package, and Kennett Square in Pennsylvania. His area crosses two state lines, and a Josh who has learned “we serve Delaware” would turn that caller away. Nothing tested that before today.

On run 7 a caller asked about a 3,400 sq ft house and then a 2,200 sq ft one. Josh quoted $500 and then $450. Our guard read the second quote as a correction, threw the $500 away, and then blocked Josh saying it eight times over two minutes. That was the dead air and the “still with you” on that call. Quotes are now kept per size. Run 8 proved it: “$450 for the 2,200 square foot home, and $500 for the 3,400.”

Also from run 7: when the guard blocks a price Josh read off the table by eye, it used to replace it in the audio and tell Josh nothing. He carried on as if the caller had heard it. Now he is told the price was not spoken and to call the price tool. Live since the 8:02 a.m. deploy.

The phone used to promise an email and an agreement at the end of every booking, whether or not we had an address or the agreement was switched on. Now the server builds the closing line from what is true for that booking, and Josh says it word for word. The web already worked this way; both brains now use the same sentence.

Proved on run 8. One caller heard “you'll get an email with a secure payment link” and nothing about an agreement, because the agreement switch is off on the server.

This has been the stubborn one. A guard that catches a spelled-out email shipped on 11 September, and Dil's searches of the server log kept showing it had never once fired on a live call. Three reasons were found, one per run, and each was fixed: the caller's turn arrives in pieces and the guard was shown one piece at a time; the phone's transcriber puts a full stop after every piece; and it mishears the letter but not the code word, so “november” now decides the letter. And the address the caller spelled is now the one we save, whatever Josh typed.

Then Dil pasted run 8's log, and the guard had fired. On the call where the caller spelled D-A-N-A-K-I-R-K-B-R-I-D-E, the log shows the guard caught the whole address and replaced three wrong read-backs (danikirkbright, danikirkbride) with the letters she gave. The same log showed three more faults, all fixed the same hour: a correct read-back behind a preamble was being “corrected”; two second refusals (“No, thanks, I don't give my email”, “I don't wanna give my email”) were shapes the rule did not read, so Ken's sentence came at the third refusal instead of the second; and “saving fifty-eight dollars” was blocked as an invented figure when it was the difference between two prices the tool had quoted.

What is left is the transcriber, not our code. On one call eleven code words (“N for November…”) reached us as NIABEHLY. On another, letters said one at a time arrived as ordinary words, so there was no spelling to catch. Only logging the caller's exact words for one run will show what the transcriber writes. That is Dil's call.

Run 8 had a recap that said “Wednesday, October 1” for a Thursday. Run 7 had a recap that said 1727 Willow Bend for a house number the caller had confirmed as 1827. Both are now caught before they are spoken. A weekday that contradicts its date is corrected. A confirmed house number with one digit changed is never spoken; Josh asks for it again instead.

Ken's own method says the spouse deferral is about 90% of lost bookings, and that it is a trust problem, not a price problem. Both brains had a list of phrases that count as that deferral, and every one needed the word with. So “I have to talk to my husband”, “I'll run it past my wife” and “I need to ask my wife” all fell straight through. Four phrasings became eight, with nine sentences pinned that must not match, like “can I ask you a question?”

Four log lines that exist to say which field was missing had been printing the word “%s” instead of the field name since they were written. Fixed, with a check that fails the build on any repeat. The tool that reads a test run now catches today's four faults by name, and says on screen which of its own answers to trust: a transcript is Josh's audio, transcribed, so a “none” from it is not proof on its own. And Dil's deploy presses were checked rather than reassured: every one is recorded, and the button refuses to ship old code.

A curly apostrophe is a different character from a straight one, and real transcripts are full of them. So “I don't want to give my email” did not read as a refusal, and “No, it's Peachtree” came back as the surname Speachtree. A full stop was read as a dot in an email. “I won't be there, but the site manager will let the inspector in” read as a refusal. “Yes, that's fine.” was recorded as the name Fine. And an email said as two words, “anna king at gmail dot com”, lost its first half, five separate times. All six fixed.

When Josh offers two openings and the caller picks one, the code only understood the literal word “first”. “9am”, “morning”, “the 9 AM one” and an empty answer all booked the 1 p.m. when she had picked the 9 a.m., and nothing on the call said so. It now reads her words, and falls back to the opening Josh offered first. Two more from the same method: a question whose answer never made sense was asked forever, up to 59 times, and an empty diary sent 41 calls out of 200 round in a loop until the ten-minute cap. Both now end the call honestly, with the inspector told.

Every one of these is a sentence Josh has really said on a test call. “I've got you down for 88 Cypress Lane” is how a person says “booked”, and the guard had never heard of it. “The office will follow up with the details” is the exact promise Ken banned on 6 September, because most inspectors have no office. Both are caught now. Nothing that was blocked before is unblocked.

Four hundred whole calls, with chatter, corrections, mid-call changes, a refused email and a diary that sometimes has nothing, in about four seconds. Every call must end properly, save at most once, never say “you're all set” without a save, and never promise an email we do not hold. It also proves Dil's checklist ask on every call: 36 questions on the list, none missed. Delete one from the walk and 187 of the 400 fail.

Dil: “lets do this now.” Until today the test line answered every hesitation at the price with “No problem at all. Thanks for calling, and have a great day”, all but word for word the sentence Ken names as the failure. Now the code spots the deferral, picks her own word for her partner, and speaks Ken's sentence from the shared script. One question, then stop. She can still book from there, which is the whole point. Nothing here is written by us.