The twelve are done. Nine more came in the same day, and six of those are done too.

Backup taken, copied from the repo clone, pm2 restart austin-voice. Verified: 5 fix markers in the running file, process started at the copy minute, no traceback, Bot ready!

⚠️ A rollback command was pasted by accident straight afterwards and failed safely — the glob matched thirty backups so cp refused, and the && meant nothing restarted. The command was mine and it was wrong; a rollback must name one exact file.

A phone call was not available, so the web Josh was tested instead — same rules, different runtime.

The fourth opened with “we’ve got good availability Monday through Friday” — softer than “plenty”, same broadcast. The rule was written against the word instead of the behaviour. Both brains now name the quieter phrasings and keep the half he got right, which was naming a specific slot.

Asked how to stop marketing a client he said “scroll to the bottom of the work order”. Those buttons moved to the top that same morning, so his answer went stale the moment the page changed. He was not inventing a feature — he was describing a layout nobody had told him.

He now carries a map of where things actually are. And the ask underneath it — “is there a system to automatically check James?” — is now a test, and the direction it runs in is the point: it reads the portal’s own markup and fails when the page and James’s description disagree.

Seven cases across all three lines, written against what actually broke rather than as a general sweep — and filed as three separate files, one per phone number.

⚠️ Three files, not one, and that is functional rather than tidy. A Hamming suite is pointed at one number when you run it. Run Ted’s two cases against Chad’s line and they fail correctly but uselessly — GC genuinely is weekdays-only and genuinely is banded by square footage. The CSV format has no phone column, so every description now opens with [RUN AGAINST <number>]; that line is the only thing standing between a suite and the wrong line.

⚠️ The two Quality cases are brain-separation tests. Both clients share one Josh, so a banded price or a weekdays-only answer on Ted’s line means GC’s setup has leaked into Quality’s call — which happened in July, to a real Delaware caller.

Written and validated; not run. Running it needs the Hamming platform.

Two separate buttons, because they are two different promises. Stop marketing is a pause an inspector can lift again. Don’t market is permanent and the server refuses to undo it — a customer who asked never to be contacted again should not be un-suppressed by a mis-click.

They sit at the top of the work order. They were below the history and the timeline at first, which is the same as not being there.

Both tags are written to the CRM, so anything reading GoHighLevel sees the suppression too.

The import window is now split: agents go back six months, clients three. One clock is read for both, so a long import cannot straddle midnight and apply two different cut-offs.

⚠️ The agent test runs first. It used to run after the date check, so an agent older than three months was dropped before anyone asked whether they were an agent.

Dil: “Josh shouldn’t book if there’s a schedule already sitting in that date. That is the rule.” The first version only flagged a clash on the calendar. It now prevents the booking — a slot is not offered at all once that inspector’s day is full.

Per inspector, not per firm. Chad being busy on Thursday must not stop Greg being offered Thursday.

The number comes from the onboarding form — a new Jobs a day field per inspector.

Point a phone at an agent’s card and the fields fill themselves — name, company, mobile, email. The image is shrunk in the browser before it is sent, so it works on a phone at an inspection rather than only on office wi-fi.

⚠️ Anything the reader is not sure about comes back blank rather than guessed. A wrong mobile in the CRM is worse than an empty one, because nobody goes looking for it.

The bulk CSV upload is off both tabs, replaced by add one contact — with its own permission attestation, the date you met them, and a tag so hand-added contacts are visible as such.

The CSV parsing itself was not deleted; it moved out of the page into its own file, so the newsletter accelerant’s list import still works and still blocks without the lawful-basis tick.

It is in the portal now, under Get help, and marked done on the launch board.

⚠️ Worth repeating from the research: there is no such thing as a Port-Out Kit. Telnyx has no such term and no such page, and Parker was promising one in five separate answer cards. The promise underneath — you can leave and take your number — is true and worth keeping. The mechanism was invented.

Ken described two mechanisms in one breath and they have different dependencies, which is why one shipped today and one did not.

The clickwrap — live. “I have read and agree to the Master Service Agreement”, unticked, gating the button on the enrollment page, with a link to the agreement beside it. The agreement now exists at a real address; before today /msa.html returned the marketing homepage, which is exactly the case Ken warned about — “a checkbox pointing at a 404 is worse than nothing, it proves the client couldn’t have read it.”

The per-clause ticks and initials — waiting. Those need Ken’s wording, and it has not been invented. A four-clause draft is in place for him to correct.

The thirteen testers — live. A free path with no plan, no price and no payment page. The agreement is signed before the questionnaire, on both paths, which settles the ambiguity in Ken’s two statements in favour of the earlier one.

Both, and neither was complete. Josh files a transcript onto the CRM contact — but only on calls that booked. Roughly two-thirds of test calls therefore have no written record on our side at all, which is why the failures were invisible.

Telnyx holds the audio for twelve months. An export endpoint now pulls what we do hold in one request rather than downloading and transcribing by hand.

An audition page lists the account’s own voices and reads the same fixed script through each, so they are compared on identical words rather than on whatever each demo happens to say.

This matters more, not less, since Deepgram Flux was rejected on quality — Dil: “it sounds like a radio… like a very old record”; Ken: “since we’re not using Deepgram, we’re staying with 11 Labs.” ElevenLabs is where any improvement has to come from now.

A Reports tab on the Inspector Portal, shaped like the monthly report Beth already sends — the headline number, business performance, call tracking, then what those calls are worth.

⚠️ Beth’s categories came back out. Dil: “that is only a base FORMAT for the report, that doesn’t mean we should have the same categories. Just use what we got right now… we need to know what business/agent gave us good numbers.” The Google Business Profile, website, blog, YouTube and social rows were five lines of things we do not have, on the report an inspector opens to see how his month went.

That space went to Who sent you the work — ranked by what each office was worth, not by how many jobs it sent. Four $450 standards and three $772 five-stars are not the same month, and a job count says they are.

Nothing on the page is estimated. Every figure comes from that inspector’s own bookings and his own call log; where something was not recorded, the report says so.

Nothing is blocked and nothing is being built yet — and the next move is Chad’s, not Larry’s. We already hold nearly every field a report header needs.

The two things that decide the whole job are visible inside Chad’s own Whisper account and he can photograph them in five minutes: the new-job screen (those are the fields we would fill) and the settings page (if there is an API key or an import button, the answer is already yes).

A copy-and-paste message to Chad is attached. Ask for the screenshots before the introduction — the call with Larry is worth far more when we already know what his software looks like from the inside.

Beth took this in the meeting, in her own words: “Let me handle it. I’ll go tell Ken I’ll do it too.” Out of scope here and left alone.

No successful lookup on this call → every figure is invented by definition, so the sentence is blocked. That is exactly the $622 call.

A lookup succeeded and this figure is not one it returned → logged, loudly. That is exactly the $452 call, and next time it is in the log the same morning instead of on a recording three days later.

⚠️ The square-footage tables are deliberately not a source of sayable figures. Every banded price must come from the tool, because picking the band is what he gets wrong. Two tests assert that from opposite directions, so a later edit cannot quietly re-open the $622 call.

If Josh reads back ten digits and then tells the caller it was nine, the claim is stripped and the rest of the sentence goes out unchanged.

⚠️ This does not fix the miscount. It removes the part that makes a caller feel accused of getting their own phone number wrong. Handing the model the count instead of leaving it to count is a bigger change with its own test call — and it is the top open item on the voice side.

Never ask the caller to confirm a price. From a real recap: “the 5-star package is 572 dollars. Is that right?” — “Well, you tell me. You’re the one coming up with the price.”

Say a price once and do not re-derive it. The $452.

Never tell a caller how empty the diary is. “We’ve got plenty of availability, so you’ll have no trouble finding a time” — said to a caller who was leaving.

The checker built this morning caught James saying something wrong. Dil asked for the other half: “James must be familiar with all the pages that he is on.” That is the absent answer, and it is harder to notice — nobody screenshots a tour that was merely improvised.

He is docked on three pages and his map covered two. On the Business Center he had nothing at all, so he invented the whole tour, confidently, on the public sales page. His portal tab list also named ten of the thirteen tabs, so Reviews, My Account and INSP Reports were invisible to him.

The checker now runs in both directions: every page he is docked on must have an entry, and every tab the portal has must get named. Parker’s tune-up has not been started.

The five reports under the monthly one are now folds. The monthly report deliberately does not fold — it is the one Ken copies and sends, so the tab shows it without a click.

A closed report carries its own count: 12 offices, nobody has gone quiet, $23,013. A list of bare headings just moves the work from scrolling to guessing, which is Ken’s actual complaint about the ISN’s forty reports — “I bet there’s nobody that ever pulls a quarter of these.”

“Who has gone quiet” now carries the phone number and the email. Same reason: the answer to “how can I remarket those people?” was, from that page, “you can’t.”

Dil: “we need to make this look like Chad’s. All features right now is so different.” Measurable rather than a matter of taste — the page’s own description promises “the same screens a subscriber gets”. A subscriber had thirteen tabs. The demo showed nine.

Added: Reports (the tab asked for this morning, folds and all), Reviews, My Account, and the Get help menu with the port-out kit visible in it before anyone signs. The library is now labelled INSP Reports, as it is in the portal, so it stops colliding with Reports.

The test reads the portal’s own sidebar and fails in either direction. Underselling loses a sale; overselling means he buys a screen that is not there.

The page ended with “tap the button to talk to Josh now”, pointing at a bubble in the corner. A demo chat converts badly not because people dislike the product but because a blinking cursor asks the visitor to invent a test, and most of them close the tab.

Six scenarios now start the conversation for him: book one start to finish · push back on the price · start in Spanish and switch to English · call as an agent · try to catch him out · talk, don’t type.

And the services are on the page, which Ken asked for so callers know what can be booked. The square-footage bands, both surcharges, the package with its arithmetic shown, all seven add-ons — and the limits, because what Josh says no to is part of the answer.

⚠️ Every number is lifted from the record Josh prices from, and the test reads that file rather than the page. A visitor who reads $475 and is quoted $545 has caught us contradicting ourselves on the one page whose entire job is to be believed.

It read: “He’s wired to a real inspection company right now. Chat with him, ask a question, book a test inspection. See exactly what your callers experience from the first ring.” Three of those clauses are checkable and none of them held.

So a prospect pressed Talk With Josh Now and landed on a page whose first line corrected the one he came from — on the exact screen where we are asking him to believe a demo. The honest version is the stronger one and it happens to be true: the same Josh, running the same brain, not a script written for the page.

This morning’s note put the missing piece on Whisper’s side and prepared four questions for Larry White. Reading our own ISN integration points somewhere different.

api/isn/availability.js is a complete ISN integration against Chad’s live account. It reads availability — live, and how Josh offers a real slot. And it writes a full order — twenty-two property fields plus the client record, which is the same thing as the report header we would be pre-filling. That write is finished, tested, and called by nothing.

⚠️ Two reasons it is dark and only one is a switch. ISN_WRITE_ENABLED is unset, deliberately, so test calls cannot create real orders in a live account. But isnCreateOrder is imported by nothing — flipping the env var alone would change nothing. That is the difference between “turn it on” and half a day, and it is worth saying plainly.

What I do not know is the whole thing: whether Whisper reads from ISN. Nobody here has seen Chad’s Whisper account. It is one question, thirty seconds, and it is for Chad rather than Larry — “when you sit down to write a report, is the job already in there, or do you type it again?”

Written against the demo firm’s real configuration, so every answer can be marked right or wrong rather than judged. Nine groups — price, the package, booking, getting the details right, agents, awkward and rude, honesty and limits, Spanish, voice.

The ones that matter most are the ones that already cost money on Chad’s calls: ask the price twice and see if it moves · give a valid ten-digit number and see if he calls it nine · read back a lockbox code and a gate code, which are different fields · 2,600 sq ft, built 1955, on a crawlspace should be $630, and he should say why.

Parker’s tune-up — not started. The ISN and Spectora contact APIs — can we pull the clients and agents list from Chad’s ISN? Not researched. That is a different question from the Whisper one above and has not been answered. An in-house CRM instead of GoHighLevel — not started.

Named here rather than left off, because a list that only shows what was finished is not a status report.

Dil: “edit the overview as right now we can understand… we need to have a sentence explainer 6th grader reading level.” The card was not badly written, it was dense — two sentences carrying six facts, 23.5 words a sentence, Flesch-Kincaid grade 8.6. It also opened with “two of them are now impossible”, a phrase he had already had to stop and ask about.

Then all twenty-one section headings, which are the list you read on the launch board before deciding whether to open this. Several were internal shorthand: “Don’t market, and stop marketing” names two buttons without saying what they do; “Six months for agents, three months for clients” gives two numbers and no subject. Everything is now measured at grade 6.0 or below.

⚠️ And a bug I caused while writing about it. I added a note explaining that these headings become the launch board’s list — and spelled the tag out in angle brackets to say so. The scraper is a regex, not an HTML parser: it found the tag inside my own comment and turned that sentence into section 1, swallowing the report’s real first heading. The board still said 21 sections. One was wrong and one was gone. Fixed at the source — comments are stripped before anything is scraped — and both cases are now tests.