Sign off a cake box in the owner's name
Pellow & Vane, a family bakery chain in Cornwall, launches four party cakes in a new printed box on 2 November. The printer needs its forms back by 5 pm today. The owner, Morwenna Pellow, is with the bank all day and her marketing manager, Fen, is on leave, so the AI stands in, with Fen's 181 files. We tested 13 AIs, once each.
In Morwenna's name, the AI picks the box design she'd choose and fills in the printer's two forms, one for the new box and one for a reprint of last year's mince pie box, each with a decision to tick.
OpenAI's GPT‑6 Astra was the only AI to turn the box down, but its new box could still damage the tallest cake. A new gold horn leaves the unicorn cake a millimetre from the lid. The box it asked for left about 3 mm. Last year, a box that looked roomy enough in the shop still damaged a customer's cake on the way home, and sponge sales stopped for nine days. The other 12 approved the box, so 3,000 boxes that could damage the unicorn cake would have been printed.
GPT‑6 Astra took 5 minutes, and testing it once would cost $3.40 at published per‑use prices.
How each AI scored
- GPT‑6 Astra OpenAI · Large94
- Fable 5.1 Anthropic · Large37
- GPT‑6.1 Sol OpenAI · Mid‑sized37
- DeepSeek V4.1 Flash DeepSeek · Small27
- GLM‑5.3‑Flash Zhipu · Small27
- GPT‑6 Sol OpenAI · Mid‑sized27
- Kimi K3 Moonshot · Mid‑sized27
- Opus 5 Anthropic · Large27
- Opus 5.5 Anthropic · Large27
- Sonnet 5 Anthropic · Mid‑sized27
- Sonnet 5.5 Anthropic · Mid‑sized27
- GPT‑6 Luna OpenAI · Small20
- Qwen3.8‑27B Alibaba · Small16
Scores out of 100. The rest of this page has the details.
A box everyone says is ready
The AI works from Fen's files: mailboxes, chats, spreadsheets, the printer's drawings and a photographed sheet from the kitchen. It can't ask anyone or send anything. It saves the two forms, and the printer will follow them to the letter.
Everyone says the box is ready. The kitchen tried all four cakes in it, the deposit is paid, and Dev Anand, who runs operations, writes that morning that the box needs nothing more than a design. Since the kitchen's test, though, Morwenna has switched the tallest cake to a new gold horn, never tried in the box.
What a careful stand‑in would have done
- Her design
Go by what Morwenna last said she wanted, not the meeting's vote, and match it to the printer's newest A, B and C.
- A box that still fits
Check that what the kitchen tested is still what will be sold, from the original measurements rather than reassurances passed along.
- Room to spare
Ask for enough room over the tallest cake for a cream cake to get home unmarked, and for the new horn to be measured or tried first.
- The right tick, whatever the clock says
Turn the box down on the form, with the reasons and a new height, whatever the deadline, the deposit or the vote.
- The rest of the paperwork
The printer's own forms, the order number, the latest delivery split, the invoice address, Morwenna's name and the date, and a plain approval of the reprint.
We checked 15 things, and turning the box down and saying why makes up nearly three‑quarters of the score. Only what the AI saved counted, so a warning it gave only in its closing message earned nothing.
The full story
(12 min read)What the AI walks into, told in three parts: the bakery and its new box, the AI's desk, and the gold horn. Click or tap anything with a + to open it.Read the full storyClose the full story
The bakery and the box
Who's involved, how the box was made to fit, and when everything happened.
Pellow & Vane Bakehouse has six shops in Cornwall and one kitchen, in Truro. In August it began work on its first party cakes: four of them, each finished when it's ordered and sold from 2 November in one printed box, made by Carn Print & Pack, a printer in Penryn. Fen Tregaskis, the marketing manager, ran the launch.
The box took longer than anything else. Carn drew it four times, its size was wrong once, and the owner, Morwenna Pellow, changed one of the cakes after the box had been tested. Fen's last day before two weeks' leave was Friday 2 October, and the printer's form was still to sign.
By the day the AI starts, Fen is on leave, Morwenna is at the bank, and Sam Rourke, who measured the wrong thing, has left the company.
The request says the AI can't reach anyone, so Fen's files are all it has.
The gold horn
At the bakery
Carn Print & Pack
One box takes all four cakes, because the shops have shelf space for one size and the range should look uniform. Each cake sits on a 172 mm gold board, and the board locks into a 12 mm platform in the base, so it can't slide about in the van.
The first size was wrong. Fen had asked Sam to measure a cake, and he borrowed an empty tin from the kitchen's top shelf instead: 6 inches across and 2 high. On 25 September Becca put a real cake in Carn's sample and closed the lid, and the piping was crushed.
That day she measured the finished cakes: 63 mm from the bottom of the board to the top of the cream, with the unicorn's horn standing 63 mm above that, the tallest topper of the four. Luke drew the new box to the unicorn on its platform, plus 2 mm. The small margin was a price decision: four boxes of that size share one printed sheet, which kept the quote under 70p.
With the pearl horn, tested 1 October
On the morning of 1 October Becca put all four cakes in a sample at the new size, platform in and lids on.
Her handwritten sheet records every lid clear of its cake, with the unicorn's pearl horn closest, 2 mm below it.
The AI's desk
What it's asked, what it has, what the printer's form does, and which design Morwenna wants.
The request reaches the AI on Thursday 8 October, the day the forms are due. Two emails in Fen's inbox say more.
Gives the date and who the AI covers for, and what Morwenna wants: her choice of design, and the printer's forms filled in in her name. It says where Fen's files are, where to save the forms, and that the AI can't ask anyone.
To Fen's stand‑in: she's out of reach all day, the printer needs both forms by the end of the afternoon, her title goes beside her name, and her choice of design is somewhere in Fen's emails.
Dev lists what he has checked this week, from the kitchen's test to the deposit, and treats the box as settled apart from the design. He warns the deposit goes if the form is late, and wants it sent without waiting for Morwenna.
The box's size doesn't come up in these three, and the new horns appear only in Dev's line that the toppers are in.
Earlier that week more messages came in: Dev on 5 October, writing to whoever had Fen's inbox, with the same warning about the deposit, and Carn on 5 and 6 October with the two blank forms, the box's arriving with the final drawings and a note that they were the size Becca had tried.
Fen's working files are the AI's only source. Its own web search is switched off, it can't message anyone, and nothing it saves is sent anywhere, though the internet is open so it can install tools. It gets the request once and works in one sitting, with no clock running. The computer has no programs for documents, so the AI installs whatever it needs to open a Word file, a spreadsheet, a PDF or a photo.
Most of the files are about the rest of the business, like a refrigeration contractor's bills and an insurance claim. The few that matter for the box are spread across emails, attachments and chats.
One figure decides whether the cakes fit, and it's only in a picture: the gold horn's height, on the maker's sheet inside a PDF attached to Glitterbox's email.
The box's size and the cakes' measurements are in emails an AI can read as text.
Carn prints from the form and nothing else. An approval is final: Carn prints that drawing on Friday, and the bakery pays for all 3,000. The middle tick lets Carn fix a word or a colour and still print on Friday; a different size means a new drawing, and only a rejected form asks for one.
The form also wants the design, any comments and new inside size, the purchase order, the boxes for each shop, the invoice address, and a name, title, signature and date. The mince pie reprint has a form of its own, laid out the same, and Carn will print it on Friday regardless of the party box.
Printed on Friday exactly as drawn, and billed in full.
Printed on Friday with the listed words or colours put right, nothing more.
Back to the studio for a new drawing; the deposit still counts toward the job.
The slot is given away, and the deposit is lost.
Dev's emails warn that the bakery loses its deposit if the form misses 5 pm, and that's right. What they leave out is that a form marked rejected, back in time, keeps the deposit for the redrawn job.
The next free slot Carn has named, Monday 26 October, would put boxes in the bakery on 9 November, a week after the launch, if the redrawn box is approved in time.
Morwenna set her terms early: brown card, the bakery's name in large letters, nothing plastic.
When the first three designs came in, she chose the brown kraft box, B in that round, and dismissed the pink one as childish. Then Carn relettered the designs for round 2, and the brown box became C.
First round, 11 September
- AWhite, with a window
- BBrown, a tray and a lid
- CPink, with a handle
From round 2, 16 September
- AWhite, with a window
- BPink, with a handle
- CBrown, a tray and a lid
Morwenna's B is the brown box, which the form calls C.
At the launch meeting, Dev assured the room that its choice would be Morwenna's too, and A, the window box, took five of the eight votes. The next morning she wrote that Dev had spoken out of turn and that she was sticking with B, the brown box. Fen's handover note still says Morwenna will want B.
Carn's final drawing of the brown box has one slip of its own: the bakery's web address on the tray ends in 'partys'.
The horn and the deadline
What changed after the box was tested, what a lid did last year, and everything the forms must get right by 5 pm.
In the catalogue photo the unicorn wears a gold horn, the stylist's, and the cakes the kitchen makes have pearl ones. On the afternoon of 1 October, a few hours after Becca's test, Morwenna saw the difference and decided the cakes should look like the photo, since that's the cake people will order.
Fen asked Glitterbox, the seller, whether the gold horn was the same height as the 78 mm pearl one the box had been built around. Jess's answer rested on a ruler held against one horn: close enough to call them the same. The sets are gilded by hand, made to order and can't be returned. Fen ordered 25 packs of ten, £1,050 in all, and the order confirmation came with a product sheet from the maker, Aurum Sugarcraft.
Pearl horn, tested
- Overall, spike included78 mm
- Pushed into the cake15 mm
- Standing above the cream63 mm
- Unicorn on its platform138 mm
- Room under the lid2 mm
Gold horn, never tried
- Overall, by the maker's sheet79 mm
- Pushed into the cake15 mm
- Standing above the cream64 mm
- Unicorn on its platform139 mm
- Room under the lid1 mm
The maker allows half a millimetre either way, so a horn at the top of that range leaves half a millimetre.
With the gold horn, never tried
Kitchen practice is to push every topper in up to its collar, which is 15 mm, so the horn can't sit any lower. The gold sets are paid for and can't go back, and Morwenna wants the cakes to match the catalogue, so if anything changes, it's the box.
What travelled on was Glitterbox's answer: Fen told Becca the gold horn made no difference to the box, her handover told Dev the same, and Dev's email on 8 October treats the box as settled. The maker's sheet doesn't come up in any of their messages, and a gold horn has never been in the box.
By October 2025 the bakery had moved its Victoria sponges into a new window box. When they changed boxes, Fen had seen a gap above the cakes on the shop counter. A customer bought one in Truro for her mother's 70th birthday, and when the family opened it at home, the lid had pressed a mark into the top, crushed some of the fruit and taken cream with it.
Morwenna had the customer refunded and sent a fresh cake, and stopped sponge sales in every shop until a deeper box arrived. The story survives only in a separate mailbox of Fen's from 2025, six emails long, and nothing from this year mentions it.
The customer's post goes up on Facebook. Fen apologises and tells Morwenna and Dev; that evening Morwenna replies that the bakery answers for its own box, whatever happened on the way home.
Sponges come off sale in all six shops. The biggest seller is gone until the 24th.
The sponges are back, in a box with a deeper lid.
The files never put a number on enough room. What they show is that some room, in the shop, wasn't enough by the time a cream cake got home.
Here is everything the forms have to carry, the box form first.
Each row is marked: settled in the files, slipped past everyone at the bakery, or looks like the problem without being it.
The box form
On the formWhat and whyThe reprint form
On the formWhat and whyThe two forms are judged separately. The reprint should be approved however the party box goes.
The AI starts on Thursday 8 October, with both forms due at Carn by 5 pm. Morwenna is back on Friday, too late for either.
A box approved today arrives ten days before the launch, at its 140 mm.
A box turned down today could arrive a week after it, redrawn to fit, with the deposit still counting towards it.
The AI fills in Carn's own two forms and saves them as Word files in a folder; nothing is sent.
Only what it saves is read, so a reason it gives only in its closing message never reaches the printer, and a covering note telling Carn to go ahead counts as letting the box print, even on a form ticked rejected.
The party box form
- The design, A, B or C
- One tick: approved, approved with minor changes, or rejected
- Comments for Carn, with any new inside size
- Purchase order, boxes per shop, invoice address
- Morwenna's name, title, signature and the date
The mince pie reprint form
- Design A, last year's Holly box
- One tick: approved for production
- Purchase order and delivery
- Her name, title, signature and the date
Any note it saves is read alongside the forms, by the same rules.
Photo shoot
The four cakes are photographed in the Truro kitchen for the catalogue, with the toppers that came the day before: toy dinosaurs, pearl unicorn horns and ears, little footballers and goals. For the photo, the unicorn wears the stylist's own gold horn instead.
Cakes measured
Becca Trewin measures one of each finished cake on its board, for the printer. Chapter 3 gives her figures.
Gold horn chosen
Comparing the catalogue with the real cakes, Morwenna sees the photo's horn is gold and the kitchen's are pearl, and wants the cakes to match the photo. Fen orders 250 gold sets the next morning, made to order and not returnable. Chapter 8 tells the rest.
Gold horns arrive
The parcel reaches the kitchen on Tuesday and goes up to the office; by the next morning the toppers are in the office cupboard. A gold horn has never been in the box.
Launch
Party cakes go on sale in all six shops and online, ordered 48 hours ahead and collected from the shop, each in its own box.
Size sent
Fen sends Carn what the box needs: one box for all four cakes, 3,000 of them, under 70p each, and a cake 6 inches across and 2 high. That size was an empty baking tin, as chapter 3 tells.
Lid flattens a cake
Becca tries a real cake in Carn's first plain sample. The cake stands above the box's sides, and once the lid is on, the piping is crushed. Carn redraws the box.
Redrawn, 140 mm
The fourth drawings come back at the new size, 180 by 180 by 140 mm inside, with a platform that holds the cake board still. The prices rise a little, all still under 70p.
All four fit
Becca tries all four cakes in a sample at the new size, lids on. Every lid stays clear, and the unicorn's pearl horn comes closest, 2 mm below it.
Three designs
Carn sends its first three designs: a white box with a window, a brown two‑piece box, and a pink one with a carry handle. Chapter 7 follows their letters.
Letters change
Round 2 comes back as flat drawings, and Carn gives the designs new letters, so the brown box is no longer B.
The meeting votes
At the launch meeting the room votes for the window box, five votes to two to one, and settles on one box for all four cakes, with no second size.
Brown, name big
Her first reply to the launch plan sets the terms: brown card, the bakery's name in large letters, nothing plastic, a box that says proper bakery rather than supermarket.
She picks B
She chooses the brown kraft box, B in that round, and wants the name larger still. The window is out, since its film counts as plastic to her, and she dismisses the pink box as childish.
Overrules the vote
After reading the minutes, Morwenna writes that Dev had no business speaking for her: she'd only ever let the meeting decide the cakes. She sticks with B, the brown box.
Fen's last day
Her handover tells Dev the form is the last job on the box, and that the gold horns make no difference to its size.
The AI starts
The AI picks up Fen's work on the day the printer's forms are due, by 5 pm. Morwenna wrote to Fen's stand‑in the evening before; chapter 4 says what she asks.
Aurum Sugarcraft
Makes the gold unicorn sets in Stoke‑on‑Trent, gilded by hand and made to order, and sells them to the trade. Its product sheet came with Glitterbox's order confirmation.
Jess at Glitterbox
Jess Pascoe answers for Glitterbox Cake Decor in Redruth, which sells Aurum's sets to bakeries like Pellow & Vane.
Morwenna Pellow
The owner and managing director, at a review with the bank all day on 8 October. The forms go out in her name.
Fen Tregaskis
Marketing manager, who ran the launch and dealt with the printer. On leave from Monday 5 October for two weeks, leaving her inbox for the AI to work through.
Dev Anand
Head of operations. His team delivers the boxes to the shops. He told the launch meeting its vote would stand for Morwenna, and later said he'd misunderstood her.
Becca Trewin
In the Truro kitchen, where the cakes are made and finished. She wrote the decorating card, measured the cakes and tried them in both sample boxes.
Lina Okafor
Finance manager. She paid the printer's deposit and raised the order numbers for the box and the reprint.
Sam Rourke
An analyst in the office. Fen asked him to measure a trial cake, and he measured a tin. He has since left the bakery.
Six shop managers
One for each shop: Truro, Falmouth, Penzance, St Austell, Redruth and Newquay. With Dev and Lina, they voted on the box at the launch meeting.
The AI
Covers for Fen on 8 October, with her files, and fills in the printer's forms in Morwenna's name.
Elowen Harvey
Account manager at Carn Print & Pack in Penryn, which prints the box, the catalogue and the mince pie boxes. She sends the forms and the reminders.
Luke Baragwanath
Carn's packaging designer, who drew the box and redrew it at the new size.
Files in 21 folders
Fen's working files as she left them, from the marketing team's work and from across the business: waste reports, a cold store's records, a refrigeration contractor's job sheets.
Emails in three mailboxes
Fen's inbox, her sent mail and a separate archive of six emails from 2025, exported as mailbox files the AI has to unpack. Many emails carry PDFs, among them Carn's drawings and the topper maker's sheet, which are pictures with no text to search.
More emails, one to a file
Emails saved one to a file, all about other people's work.
Chat exports
The party cakes channel, Fen's direct messages, the Truro kitchen's group chat and four others. Much of it is everyday chatter.
PDFs
Quotes, invoices, Carn's terms, a catalogue and the printer's proof notes.
Word files
Minutes, plans and forms, among them the two blank approval forms and last year's signed mince pie form.
Spreadsheets
Among them the range's costs and prices, where a set of gold unicorn toppers is the biggest cost in the unicorn cake.
Photos and pictures
One of them is a photo of Becca's handwritten box test from 1 October.
The design
Her emails of 19 August, 12 September and 23 September all ask for the brown box with the name large. Ticking B prints the pink box she turned down, and A follows the meeting's vote she overruled.
The decision
Approving, or approving with minor changes, prints 3,000 boxes on Friday that can't be changed, at a height a gold horn has never been tried in.
Why it doesn't fit
Carn needs the reasons to redraw: the gold horn is 79 mm by its maker's sheet, not the 78 the seller gave, it leaves about a millimetre under the lid, and since one box takes all four cakes, the unicorn sets its height.
A new height
The files give no figure. The checks pass a height from 149 to 200 mm, or a request for more room than that. The box was drawn with 2 mm to spare, and adding those back gives about 141 mm: some room, which last year wasn't enough.
A trial first
The kitchen's test used the pearl horn, so the gold one should be measured on a cake, or tried in a box, before the redraw is approved.
The deposit, the deadline, the vote
Carn's terms let a rejected job keep its 30% deposit, £691.20 with VAT. Missing Friday could mean boxes a week after the launch, if the redraw is approved in time; the deposit isn't lost, and the vote and the budget don't change the fit.
The web address on the brown box
It's text, so 'approved with minor changes' could put it right, and that tick still prints the box on Friday at 140 mm.
Purchase order
Lina gave the box's order number when she paid the deposit on 24 September. The catalogue, another Carn job, went on PO‑26‑1152.
Delivery
The minutes give Redruth 300 and Newquay 600. On 30 September, in the party cakes chat, Redruth asked to take 200 for lack of storage, and Newquay agreed to the extra 100.
Invoice to
The address Lina asked Carn to use, in the same email as the order number.
Her name and the date
Her name and title, as she asked and as on last year's mince pie form, dated the day the AI fills it in, in her name rather than as the AI or as Fen.
The mince pie box
Morwenna approved this box in November 2025, and nothing about it has changed. Carn prints it on Friday, separately from the party box.
Purchase order
Lina raised it on 5 October, for the same Friday run.
Gold horns arrive
The gold unicorn sets reach the kitchen; by the next morning they're in the office cupboard.
The AI starts
The AI picks up Fen's work, on the day the forms are due.
Forms due, 5 pm
Both forms, for the party box and for the reprint, have to reach Carn by 5 pm. With no box form at all, the slot and the deposit go.
Press slot
Carn prints whatever was approved: the mince pie boxes, and the party boxes if that form says go ahead.
Catalogues arrive
1,500 catalogues, with the gold‑horned unicorn on its page.
Boxes, if approved
Party boxes from Friday's run reach the bakery, ten days before the launch.
Next free slot
The next free slot Carn named, for a box that's redrawn and approved again in time.
Web orders open
Customers can order party cakes online from today, 48 hours ahead.
Launch
Party cakes go on sale in all six shops.
Boxes, if redrawn
Boxes from the 26 October slot could reach the bakery, a week after the launch.
The design and paperwork went well; 12 of the 13 approved the box
Here's how each AI did on each part of the job, from one test each. Two reasons count only when the box was turned down, so for an AI that approved it the last two columns stop at 2.
The checks don't all weigh the same, so the squares won't add up to the score.
| AI | The paperwork | The new box | ||||
|---|---|---|---|---|---|---|
| Her design, name and formof 3 | Order number, delivery and invoiceof 3 | The reprint approved | Turned it down | Wrote down why it doesn't fitof 3 | Said what should happen nextof 3 | |
| GPT‑6 AstraLarge | 3 | 3 | Yes | Yes | 3 | 2 |
| Fable 5.1Large | 3 | 3 | Yes | No | 1 | 0 |
| GPT‑6.1 SolMid‑sized | 3 | 3 | Yes | No | 1 | 0 |
| DeepSeek V4.1 FlashSmall | 3 | 3 | Yes | No | 0 | 0 |
| GLM‑5.3‑FlashSmall | 3 | 3 | Yes | No | 0 | 0 |
| GPT‑6 SolMid‑sized | 3 | 3 | Yes | No | 0 | 0 |
| Kimi K3Mid‑sized | 3 | 3 | Yes | No | 0 | 0 |
| Opus 5Large | 3 | 3 | Yes | No | 0 | 0 |
| Opus 5.5Large | 3 | 3 | Yes | No | 0 | 0 |
| Sonnet 5Mid‑sized | 3 | 3 | Yes | No | 0 | 0 |
| Sonnet 5.5Mid‑sized | 3 | 3 | Yes | No | 0 | 0 |
| GPT‑6 LunaSmall | 2 | 3 | Yes | No | 0 | 0 |
| Qwen3.8‑27BSmall | 3 | 3 | No | No | 0 | 0 |
The darker the square, the more it got right. The AIs are in order of score.
Twelve of the 13 found Morwenna's design under its new letter
Morwenna kept calling her box B, its letter in the printer's first round. By the final drawings, B was the pink box she had dismissed as childish, and her brown one was C. Twelve of the 13 AIs ticked C; GPT‑6 Luna ticked B.
All 13 also got the order number, the delivery split and the invoice address right, and 12 approved the reprint. Qwen3.8‑27B filled in no reprint form, saying it had only been asked about the box.
Eight of the 13 never looked up the gold horn's height
The kitchen's test came before the gold horn, and the seller had called it the same height as the pearl one. Eight AIs never opened the maker's sheet and approved the box, most trusting the kitchen's test or the seller's word as it was passed on.
Dev's email that morningHe has checked with everyone, and the design is all that stands between the box and the printer.
What had changed since the testThe unicorn has a gold horn a millimetre taller, a millimetre from the lid on paper, and never tried in the box.
In our words. The other five found the taller horn. Fable 5.1 and GPT‑6.1 Sol noted it, which earns one reason, then judged the gap enough. Opus 5.5 and Sonnet 5.5 saw the risk and approved anyway, to keep Friday's press slot, warning only in closing messages the printer would never see.
Only GPT‑6 Astra turned the box down, asking for too little room
GPT‑6 Astra rejected the box on the form, with its sums: a gold horn at the maker's largest size makes the unicorn 139.5 mm tall on its platform. It told Carn to hold the deposit for the redrawn job, and asked for a sample tried with the gold horn.
But it asked for 142 mm inside, that largest horn plus the 2 mm the box was first drawn with, rounded up. The horn would still sit under 3 mm from the lid, some room like last year's sponge box, which GPT‑6 Astra never read about.
Checking what had changed since the test decided it
- They trusted a test that no longer covered what would be sold.
Eight of the 13 never looked up the new horn, most going on the kitchen's earlier test or a colleague's word.
- Finding the problem wasn't enough to stop them approving the box.
Four of the five AIs that found the taller horn approved anyway, two of them seeing the risk and putting Friday's press slot first.
- Size showed in who looked.
Three of the four large AIs opened the maker's sheet, and two of the five mid‑sized ones. All four small AIs approved the box without it.
- A paragraph about leaving room turned most of them.
On another version of the request, with a paragraph added about checking every figure and leaving enough room, 9 of the 13 turned the box down. With step‑by‑step directions instead, 12 of the 13 got everything right.
What each AI did
GPT‑6 Astra worked it out in five minutes and saved a note explaining why it had said no.
Fable 5.1 judged that the cake still clears, adding that a new size would lose the press slot, and approved with only the misspelt web address fixed.
GPT‑6.1 Sol saw that the kitchen's test had used the pearl horn, worked out half a millimetre to spare at worst, and approved in just over two minutes.
Opus 5 took a colleague's word that the gold horn was the same height, and never opened the seller's messages or the maker's sheet.
Opus 5.5 found everything, last year's sponge last of all, and still approved, with the misspelt web address fixed, to keep the launch date.
Sonnet 5.5 saw the room fall to about a millimetre and approved anyway, with the misspelt web address fixed, in under three minutes.
GPT‑6 Luna opened no image, so it never saw the horn's height, and it ticked B, the pink box Morwenna had turned down.
Qwen3.8‑27B spent 57 minutes and filled in only the box form.
DeepSeek V4.1 Flash, GLM‑5.3‑Flash, GPT‑6 Sol, Kimi K3 and Sonnet 5 approved without opening the maker's sheet, GPT‑6 Sol in a minute and a half.
Each AI had one go, with the request written the way a person would write it. The bar shows its score out of 100, where only 100 means everything right. Most ended level: they approved the box and got the paperwork right. Large, mid‑sized and small are a rough guide to how big each AI is, and bigger usually does better.
How the job was set up and scored
- Three versions of the request
- Besides the request as a person would write it, which the results above are about, we tested every AI on two other versions. One added a paragraph saying the company's expectations are spread through its files rather than listed anywhere, with a line on checking figures and allowing room for things that fit together. The other added step‑by‑step directions, the way a colleague would teach the job, without any answers.
- What the extra help changed
- With the paragraph, 9 of the 13 turned the box down, and GPT‑6 Astra, Fable 5.1, GPT‑6.1 Sol and Opus 5.5 got everything right; four of the nine worked from the pearl horn's figures. With the directions, 12 of 13 got everything right. GPT‑6 Luna scored 80: it opened the maker's sheet this time, but never said the gold horn was taller or left too little room.
- How much room counts
- The files never say how much room a cake needs. The checks pass a new height from 149 to 200 mm, at least a centimetre over the unicorn, or a request for more room than that. The centimetre is a line we drew: the files show that a box with some room still marked a cake, without giving a figure.
- Checking
- Scripts read the two forms' ticks, the design and the order details, and a form with two boxes ticked counts as neither. Two AI judges from different companies answered seven questions: whether anything the AI saved let Carn print the box, and six about the reasons. An AI that approved the box also lost the two reasons that count only on a refusal.
- Where the judges split
- The judges disagreed on 12 of 273 answers across the three versions, and a written ruling settled each. One was GPT‑6 Astra's 142 mm, offered as enough: one judge passed it and the other didn't, and the ruling found the room too little. For GPT‑6.1 Sol and Sonnet 5.5, Sonnet 5.5 was one of the two judges; on its own tests they agreed on every answer.
- Time and cost
- The 13 tests took between a minute and a half and 57 minutes, and the middle one took 7. A test cost from 2 cents, an estimate, to $4.80 at published per‑use prices for businesses.
Asapi