It read a price off a blank page. 8 times out of 8.

It read a price off a blank page 0:00

Pip That is a hotel page. All of it.

Bo Two hundred and eighty-nine thousand bytes of HTML. Strip the scripts and the styles, take the text a reader would see, and you get one space.

Pip So the price checker sent that to a language model and asked it what the room costs.

Bo It said nineteen ninety-nine.

Pip For a night at the Royal Plaza in Hong Kong.

Bo Which is a hundred and ninety-one dollars sixty-nine.

Pip We ran it eight times to be sure.

Bo Eight times. The same answer, to the cent, every time.

Pip And it never once said it did not know. The error field was empty. As far as the rest of the program could tell, that was a price it read off a page.

Bo None of which ever reached the database - and the reason it did not is the whole argument.

Pip That is fixed now. It got fixed the same afternoon we found it, and you will see the diff.

Bo It is four characters. Getting there is the episode.

Pip Empty folder. PriceReduced.

Who's talking 1:00

Bo We did not build this one either. Neither of us has ever paid for a hotel room.

Pip Everything a person decided in this story arrives the same way it actually arrived.

Bo There is no write-up for this project. The whole history is twenty-one commits, a handful of docstrings, and a database that has been running since July.

Pip Which turned out to be better evidence than a write-up. Nobody smooths a commit message they wrote at midnight with the site name still in it.

What exists: 19 items, 1,897 checks, two and a half cents 1:27

Pip Here is what exists. You give it a product URL. It reads the price, on a schedule, and keeps the history until you delete the item.

Bo Nineteen URLs. Vacuum parts, a jacket, a trackpad, four Hong Kong hotel stays, two award-flight searches and one Google Flights itinerary.

Pip Eighteen hundred and ninety-seven price checks. Fifteen hundred and fifty-two of them found a price.

Bo It runs in Docker on a Raspberry Pi on a shelf, behind a public HTTPS URL, and nobody has touched it since the tenth of July.

Pip And the part everybody asks about first.

Bo Two and a half cents. All time.

Pip There is a spending cap in the config, set at five hundred calls a month. The busiest month used a hundred and fifteen. It has never once bound.

Bo So the cap is not what is keeping this cheap. Look at the shape instead.

Pip Thirty days. A flat line. Roughly the same spend every single day.

Bo A system whose model calls are supposed to be exceptional does not draw a flat line. Hold on to that; it is lesson four and it is the expensive one.

AI belongs where the ambiguity begins 2:34

Bo Everywhere you can name what you are looking for, you have already written the rule.

Pip A price in a JSON-LD offer block is named. A price in an Open Graph meta tag is named. A price in an element the page itself labelled "price" is nearly named.

Bo None of those need a model. They need about forty lines each, they are auditable, and when they are wrong they are wrong somewhere you can point at.

Pip And the corollary nobody puts in the design doc.

Bo A fallback is not free just because it is rare.

A page with no price needs a door, not a bigger model 3:07

Pip Four of the nineteen URLs are Agoda hotel pages. Every single check failed.

Bo Not "found the wrong price". Found nothing. You have already seen why - the page builds itself in the browser, and there is nothing there for any tier to read.

Pip The fix was not a better model. It was a different door.

Bo The page ships the URL of its own pricing API, with the dates already in the query string. Load the page once for the session, pull that URL out, call it.

Pip Eleven hundred and seven checks have run through that one regular expression. It is the single most productive line in the repository.

Bo And it is the only reason the nineteen ninety-nine never got written down. That handler runs before the generic cascade, so the model is never asked.

The rules were wrong twice, and both fixes were rules 3:55

Pip Meanwhile the bottom tier - the one that reads elements a page has merely labelled price - was getting it wrong in two different ways in the same week.

Bo On an Epson paper page it read the paper.

Pip On an Amazon trackpad page it read a warranty.

Bo Neither of those is a model problem. The first fix is: only count a number that is attached to a currency symbol.

Pip The second is better. Stop taking the first match. Collect every price-shaped candidate on the page, and let the page vote.

Bo The real price repeats. The add-on appears once. That is a five-line rule and it has not been wrong since.

So a model was put on top as a check 4:33

Pip Both of those were caught by a human noticing a wrong number on a dashboard, which does not scale past nineteen items.

Pip So: one model call to sanity-check any bottom-tier price, the first time it is seen or whenever it changes.

Bo Agree, and confidence goes up. Disagree, and the model wins. The reasoning was written down and it was reasonable - both known failures were ones a model would have caught.

Pip Eight weeks in production. It confirmed one price and overrode one price.

Bo And the override was the wrong call. A five percent clip coupon.

Bo The page said a hundred and fifty-nine; the model read a hundred and fifty-one oh five; and the rule as written handed it the argument.

Pip The patch widened "agreement" to ten percent, so coupon-sized differences count as agreeing.

Bo And here is the bit we only know because we went and checked. That same page, eight weeks later.

Pip Same page. Same two numbers, to the cent. The model has not got better at it; the tolerance just got wide enough to absorb it.

Bo Which is a real fix. It is not the same thing as a solved problem.

Pip The last change is the quiet one. Every check now also scrapes the product's barcode, part number and brand, whether or not that is where the price came from.

Bo So that the same product tracked at two shops can be checked against itself. Hold that thought too.

The whole stack, and what each tier actually found 6:02

Bo Twelve rows. One of them is a model, and it found two point two percent of the prices.

Pip And two rows in that cascade have never won a single check in eight weeks. Microdata and embedded JSON. They cost nothing and they have done nothing.

Bo Which is its own small lesson about how cheap it is to over-build a cascade, and we are not going to spend a card on it.

Lesson 1 — a fallback has to be able to say it had nothing 6:26

Pip Back to the top. The model was sent one space and answered nineteen ninety-nine. The interesting part is not that it made something up.

Bo It is that the code asking the question could not tell.

Pip Identical, field for field, to a price genuinely read off a real page.

Bo And there is a guard. Somebody thought of this.

Pip One space is not falsy. It walked straight through.

Bo Four characters.

Pip So we shipped it. Same afternoon.

Bo Read the second row, not the fourth. The model did not get better. The model was never asked.

Pip And the third row is the one a caller can act on. Before, a made-up nineteen ninety-nine and a real reading were the same object. After, a miss is reported as a miss.

Bo And the actual price still comes back, from the site handler, exactly as it did before. The fix closed the hole behind the door. It is not what keeps the door shut.

Pip The transferable version has nothing to do with models.

Pip Any time you hand work to something that can only answer, and never decline, you have to check what you handed it - because it will answer.

Lesson 2 — the right number can still be the wrong number 7:40

Bo The Agoda handler worked, and then the prices it recorded disagreed with what a browser showed for the same hotel on the same night.

Pip The extraction was not wrong. It was reading the field the room grid displays by default - which, in most markets, is before taxes and fees.

Bo Thirteen point zero zero percent. Three different hotels, identical to two decimal places. That is a ten percent service charge and a three percent hotel tax.

Pip So the number was right and the meaning was wrong, and no amount of checking the number would have found it.

Bo The fix is that every price point now records what it is - and a chart refuses to plot two different bases on the same line.

Lesson 3 — the badge that has never said yes 8:23

Pip You can group the same product across shops. Amazon and Apple, say. And every member gets a badge saying how sure the system is that it is really the same thing.

Bo Barcode match, or brand plus part number, gets you the tick.

Pip In eight weeks, the tick has never been awarded.

Bo Apple's own product page publishes a brand and no part number and no barcode. The Amazon listing publishes nothing at all.

Bo Same hundred-and-fifty-nine-dollar trackpad, two pages, no comparable identifier between them.

Pip So it falls through to comparing the titles, and the strongest answer it can give is "probably".

Bo The verified branch is correct. It is tested. It is unreachable.

Pip An identity check is only as good as the identifiers your sources publish, and the two biggest shops in this dataset publish none.

Lesson 4 — retry is a cost nobody models 9:14

Bo Two hundred and twenty-five model calls in eight weeks. Sixty-two percent of them came back with nothing usable.

Pip That sounds like hard pages. It is not.

Bo A hundred and thirty-seven calls. Sixty-one percent of everything this project has ever spent on a model. Zero prices.

Pip Two of those are award-flight searches that build themselves in the browser. The price is not in the page. It was never in the page.

Pip The model was handed an input that did not contain the answer, a hundred and fifteen times, and asked again the next morning.

Bo That is the flat line on the dashboard. It is not a system with a rare fallback. It is a subscription to three URLs.

Pip It is four lines. It is not written. And the reason it is not written is that the cap made the budget feel like a solved problem.

Lesson 5 — and the model is still worth having 10:04

Bo So we tested it properly. Every tracked page, fetched once, run through the rules and the model side by side.

Pip Zero. On every page where both had an answer, the model tied or lost. It never once won.

Bo Four of its five misses were plain retail product pages with the price sitting in a JSON-LD block.

Pip And then there is the one page where the rules have nothing.

Bo No offer block, no price meta tag, no stable element name, and a price that genuinely moved four eighty-nine to seven twenty-five in a month.

Bo The model reads it. It read seven eighteen the day we tested, against seven nineteen the day before.

Pip One page in nineteen, where the ambiguity is real and the answer is actually in the input.

Bo That is not the fallback failing to be the product. That is the fallback being exactly the right size.

Three more, quickly 10:55

Pip Three more, quickly.

Bo One. The default model on day one was a free-tier one. It was replaced after a single call, because free models parse badly and disappear.

Pip Two. This is a single-user app and it still had a concurrency bug - a manual check and the scheduler could fetch the same item at the same moment.

Pip It is fixed with a database lease, which is what you would do for a hundred users.

Bo Three. Two shops block it outright, and the dashboard says so. It says site blocks bots rather than showing an empty price and letting you assume.

Pip An honest status is a feature. It is also the opposite of the bug at the top of this episode, which was a dishonest status wearing a real one's clothes.

What we don't know yet 11:42

Bo The plan for this episode said: run the cross-check on for a fortnight, off for a fortnight, and count.

Pip We did not run it. It cannot answer anything.

Bo A fortnight of waiting to observe nothing. The replay you have been watching took an afternoon and answered more.

Pip Which is worth saying out loud, because a test with no power looks exactly like a test in a planning document, and nobody checks before they run it.

Bo The guard got fixed. The other thing in this episode did not, on purpose.

Pip A circuit breaker. Something that notices a URL has failed five times running and stops asking. It would have saved about a hundred and twenty of those hundred and thirty-seven calls.

Bo And it is written up rather than built, because those four are judgement calls and not one of them is a patch. The one that actually matters is the second.

Pip A breaker that never closes is a silent permanent opt-out, which is worse than the waste, because at least the waste shows up on a bill.

Bo The other thing we genuinely do not know. Those two flight-search pages fetch perfectly well and contain no price at all. Would a real browser get one?

Pip The static page reads "zero flights", which is just what it says before the JavaScript runs. It is not evidence either way. Nobody has pointed a browser at it.

Bo And the cheap answer might beat the clever one. Four lines that stop asking costs nothing. A headless browser is a whole dependency that fights a site for a living.

Pip So: next time we come back to this, it is whichever of those got built, and what the call log says afterwards.

End card 13:16

Pip Nineteen URLs. One of them needs a model.

Bo Empty folder's cheap. It's the second call that costs you.