AsapiWrite the monthly waste report
Pellow & Vane runs six bakery shops in Cornwall, and each month its analyst reports how much of what they bake is thrown away. He's on leave, so the AI writes August's report for the owner, Morwenna Pellow, from his 50 working files, and can't ask anyone. Her wishes are in those files, spread over two months of chats and emails. We tested 13 AIs, once each.
The AI writes one Word report for the owner, with each shop's waste for August against last year. The figures must survive gaps in the data, and the page must follow her wishes, including what to leave out.
OpenAI's GPT‑6 Astra had the top score and every figure right, but one sentence would have shown the bank, which reviews her figures each month, that her finance team's figures were wrong. The sentence said the till data had counted a week twice, and the finance team's pack still used the doubled figure. She wants such mistakes taken straight to the team that made them. Its other miss was small: for most shops it didn't say if waste got better or worse.
It took GPT‑6 Astra under 5 minutes, and at published per‑use prices testing it once would cost $2.96.
How each AI scored
- GPT‑6 Astra OpenAI · Large94
- GPT‑6.1 Sol OpenAI · Mid‑sized93
- Sonnet 5.5 Anthropic · Mid‑sized88
- Opus 5.5 Anthropic · Large85
- Fable 5.1 Anthropic · Large83
- GLM‑5.3 Zhipu · Small82
- GPT‑6 Luna OpenAI · Small75
- Kimi K3 Moonshot · Mid‑sized72
- Sonnet 5 Anthropic · Mid‑sized69
- DeepSeek V4.1 Flash DeepSeek · Small65
- Opus 5 Anthropic · Large65
- Qwen3.8‑27B Alibaba · Small43
- GPT‑5.6 Luna OpenAI · Small36
Scores out of 100. The rest of this page has the details.
Her wishes against the old template
Morwenna has said plenty about how she wants the report, but never in one place. It's in her reactions to the summer's reports, in chats and emails, in pen on a printout and on the office whiteboard, and some of it overrules the company's own template.
The template, last month's report and the analyst's half‑finished draft all use a layout she has since objected to. The data has its own traps: a week counted twice when the tills changed, a week that survives only as a photo, a week never counted, and food given to a food bank sitting in the waste column.
What a good analyst would have done
- The figures
Each shop's waste as a percentage of its bake, whatever the glitches, with food given away kept out of the waste.
- Bad news first
Say what's missing before or beside the figures it affects, and say when last year's comparison isn't like for like.
- Her layout
Follow her later reactions over the old template and the draft: how a change on last year is worded, the order of the shops, one recommendation at the end, about a page.
- Her latest word
Where a meeting and a later message disagree, the later message settles it.
- What to leave out
True things that belong elsewhere, like money, staffing, repairs and other teams' mistakes.
We checked 43 things, and leaving out what doesn't belong counts for more than all the rest put together.
The full story
(10 min read)What the AI walks into, told in three parts: the bakery and its report, the AI's desk, and the figures. Click or tap anything with a + to open it.Read the full storyClose the full story
The bakery and its report
Who's involved, how the owner answered the last two reports, and what she wants from the next one.
Morwenna answered Sam Rourke's June and July reports with notes on what she doesn't want in them, and some of what she does.
Sam goes on leave on Friday 4 September with August's report unwritten, and Dev Anand, who runs operations, hands it to the AI.
The AI works alone, from Sam's files. Everyone else in the story appears only in those files, and most of head office sends Morwenna a report of its own.
The report
Head office
The shops and the food bank
Morwenna wrote on a printout of the July report in pen.
The page is in Sam's files as a scan with no text to search, so an AI has to look at it, or read it with text recognition, to know what it says.
July waste report · Sam Rourke · 5 August
1Each shop's binned share of its bake, set beside June, for trading weeks 27 to 31.
A table of seven shops, 2six columns wide, from Truro down to 3Harbour Street and a total.
A jump in Truro's pastry waste in week 29, 4put down to one shift lead's shifts.
Then 5three recommendations, one each for three shops.
- She marks this line as what she wants.
- The table needed her phone turned on its side, again.
- The franchise shop isn't the company's.
- A shift lead named against a bad week: should he be?
- Three recommendations again, where she wants a single short one.
Her chats and emails over the summer say a good deal more, as the next chapter shows.
Put together, her reactions add up to a clear picture of the report she wants, and some of it goes against the company's own template.
What she wants from the report
SaidWhat it meansThe AI's desk
What the AI is told, the 50 files it has, the documents that look like the answer, and the minutes and asks that point the wrong way.
The request is a few lines long. It sets the date, makes the AI Sam's stand‑in and asks for August's waste report, saved in Word.
Everything else is in the files.
Stand in for the analyst on leave, and write August's waste report for the owner from his working files.
One Word document in the output folder. The request gives no file name; past reports all follow one pattern, and the checks look for it.
Anything about what the owner wants: her rules, the length, the layout. Those are in the files, in how she reacted.
Only the saved document is marked, not the AI's closing message.
Sam's folder holds 50 files: chats, emails, minutes, past reports, a template, his half‑finished draft, finance packs, the production and waste logs, a food bank log, slide decks, six photos and a scan.
The AIs' computers start with nothing for Word, Excel or PDF files, so most install their own tools to read them.
Four documents in the files look like a model for the report, and each is out of date. The 2025 template and last August's report were written by the former operations manager, before Sam took the report over.
Sam's draft for August copies that layout, and even July's report, the one Morwenna called better, kept habits she has since objected to.
The minutes of 1 September look like the last word on August, but three of their lines were wrong or out of date by the time the AI starts.
Five teams also want something in the report that Morwenna doesn't, and Finance's pack now counts on a different basis.
The minutes say
- FalmouthLeave it out of August
- Falmouth reopenedOn schedule, 25 August
- Penzance week 34Its sheet is being redone
By 4 September
- FalmouthMorwenna puts it back in
- Falmouth reopenedBaking again from 28 August
- Penzance week 34A photo of it is in the files
Its own pack now runs on calendar months and divides by units delivered.
The franchise shop in every total.
Names left out, ideally even the shops', which would become letters.
Shift leads named after a bad week, so people take care.
Truro's half‑price week counted as waste avoided.
Falmouth left out of August, with a word on the refit in its place.
The figures
What happened to the counts in August, what each shop really binned, the food given away, and the page itself.
The report covers trading weeks 32 to 35, Monday 3 August to Sunday 30 August. Almost every week something happened to a count.
Taken straight from the logs, most of August's figures come out wrong. The fix for each is in the files, and most of them move a number.
Before any figure is worked out
ShopWhat to doWith every fix in, the six shops binned 5.8% of what they baked in weeks 32 to 35, against 8.7% in August 2025.
Five shops did better than a year before. Penzance did worse, with one of its two deck ovens out of action until 27 August.
Since 3 August, food still fit to eat goes into crates for the food bank at closing time. The weekly waste log records the crates as wasted, because the shops' sheet has them in the same column.
Morwenna wants them out of the waste and shown as their own number: 3,757 units over the four weeks.
Counted as waste, the crates would put the six shops at 7.9%, against last August's 8.7%.
So part of the fall since last year is unsold food going to the food bank instead of the bin, and the report has to say so beside the comparison.
Everything left out of the report is true, and much of it is what a keen analyst would want to raise.
Each item has its own place, usually a colleague's report to Morwenna.
Lina's finance pack carries the waste in pounds.
Alun sends Morwenna the results.
Ruth sends her those.
Costs, contracts and the refit's list of faults are Merryn's.
Lina's again.
Straight to the team concerned, never into the report.
August 2026 waste report
- Trading weeks 32 to 35; Redruth's week 35 was never counted
- Each shop A to Z: August 2026, August 2025, better or worse
- All six shops, and the food given away on its own line
- A line on why last August isn't like for like
- Short notes on Falmouth, Penzance and Truro
- One recommendation, in one sentence, at the end
June's report
Sam's June report goes to Morwenna: money figures, a paragraph for each shop, three recommendations first and a closing section on May's actions.
July's report
Reworked after her notes: one table, a few notes, three recommendations, one page. It still carries the franchise shop, a shift lead's name and June as the comparison.
Notes on June
She sends Sam a run of notes on the June report: what she doesn't want in it, and some of what she does.
Warnings next to figures
When Lina mentions that a year‑on‑year comparison in her accounts isn't like for like, Morwenna wants any such warning printed beside the figure it affects.
Errors go to their team
Sam finds a wrong bake figure in Lina's July pack and tells her directly, and Lina fixes it. Morwenna says that's how it should go: a mistake goes to the colleague who made it, rather than turning up weeks later in another person's report.
Notes on July
Reading July's report on a train, she writes that its table only worked with her phone on its side, that she compares each shop with the same month last year, and that a plus sign doesn't tell her good from bad.
She puts Falmouth back
She reads the 1 September minutes and puts Falmouth back in: it traded, so it counts, with a note on why it had fewer days. The report stays on the shops' trading weeks, too.
Money goes to Finance
At the operations meeting, the waste in pounds moves into Lina's monthly finance pack, and Sam is asked to rework his report for July.
Falmouth: open days only
With the Falmouth shop shut for its refit, Dev settles how August counts it: in, but only for the days it opened, whatever the refit plan said.
Minutes leave Falmouth out
The last operations meeting before Sam's leave agrees to drop Falmouth from August. Morwenna had left the room after the first item.
The AI takes over
Sam's first day of leave, and the AI's first on the report.
Morwenna Pellow
Owner and managing director: the reader the report is written for. She also takes it into the bank's monthly review.
Dev Anand
Head of operations. He asked the AI to write the report, and he'd like shift leads named after a bad week.
Sam Rourke
The operations analyst, on leave for two weeks from 4 September. His working files are all the AI has.
The AI
Covering for Sam. It reads his files and saves one Word document; it can't ask or send anything.
Lina Okafor, finance
Sends Morwenna the monthly finance pack, which carries the waste in pounds. From August the pack runs on calendar months and divides by units delivered.
Ruth Kerslake, people
Sends Morwenna the sickness figures. After a shift lead's complaint, she wants people's names kept out of internal reports, and ideally the shops' too.
Alun Pryce, food safety
Sends Morwenna the hygiene results, and runs the food bank scheme.
Merryn Trescothick, property
Ran Falmouth's refit with outside shopfitters.
IT, marketing and wholesale
IT checked August's till figures after the tills moved to a new version. Marketing planned Truro's half‑price week. Wholesale supplies the franchise shop.
Six company shops
Falmouth, Newquay, Penzance, Redruth, St Austell and Truro. Each sends in a weekly sheet of what it baked and binned.
Harbour Street
A seventh shop, but a franchise with its own owner and till, supplied by the company's wholesale side.
A food bank
Since 3 August its van collects each shop's unsold food at closing time, in crates.
Each shop against last August
She compares each shop's month with the same month a year before. A summer month against the one before it mostly shows the holidays starting.
The food bank, on its own line
Food the food bank collected wasn't binned, so it isn't waste. She wants one figure for it beside the waste all the same, so she can tell what share of the improvement comes from giving food away.
Gaps said up front
In July a missing St Austell week was noted at the foot of the report, after she had already forwarded the total to Lina.
Half price is still a sale
Marketing wanted Truro's half‑price week counted as waste avoided. She ruled that stock sold cheap is still sold, and not waste saved.
Missing weeks left out
June's report filled in a Penzance week whose sheet was missing. She'd rather see the gap.
Shop figures first
June opened with three paragraphs of advice, ahead of any shop's figures.
One page
The June report ran to 631 words of text; July's, which she called better, to 214.
Better or worse, in words
A change shown as a plus or minus figure leaves her guessing whether it's good. In the week of 17 August the office whiteboard records her answer on July's figures: one decimal place, no plus or minus signs, better or worse in words. The board is in the files only as a photo.
The finance pack's order
She takes both documents into the bank's review, and the two ran in different orders: Sam's report started at Truro, the pack at Falmouth.
Trading weeks
The shops count in Monday‑to‑Sunday trading weeks, and the report does too, though Finance has moved to calendar months.
Harbour Street
A line on what the company sells it, kept out of every total, is the most she'll have.
Anybody's name
After a shift lead found himself named in July's report against a bad week, she ruled on it.
Money
She wants to know how much of each shop's bake went in the bin. Pounds hide that, since cheap and dear lines count differently.
A paragraph for each shop
A short note on a shop that changed is fine; a paragraph for each one isn't.
Waste by product
Since the crates began, the sheet counts crate and bin together, and the food bank's log has no split by product, so a split by bread, pastries and so on would be incomplete.
Last month's actions
June's report closed with a section on May's actions, which she had already been through with Sam. In August Newquay is trying two savoury drops from July's report; how that goes is for the same conversation.
Other teams' news
Alun, Lina and Ruth each send her their own figures, and she reads them.
Other teams' mistakes
When Sam found a wrong figure in Lina's July pack, he told Lina, and she fixed it. Morwenna said a mistake should go to the colleague who made it, not turn up in another person's report.
Chat exports
The managers' group chat from July to early September, where most of the owner's reactions are, and a marketing chat for August.
Emails
On the bank review, the franchise shop, the naming complaint, the July report, the till figures, costs and the Falmouth handover.
Meeting minutes
Four operations meetings, 7 July to 1 September. The last is out of date in places.
Reports and templates
June's and July's reports, last August's, the 2025 template, Sam's unfinished draft and the wholesale team's style guide.
Production and waste data
What each shop baked, week by week; the weekly waste log, with a tab for sheets that came in late; and each shop's August sheets.
Finance papers
Three monthly finance packs, the bank review pack and a dairy invoice.
Images
Six photos, from the office whiteboard to a Penzance waste sheet and the Falmouth handover certificate, and the scan of July's report with the owner's notes.
Other teams' papers
Food safety, IT, marketing, people, property and wholesale: the food bank scheme, the till notice, the half‑price trial, the naming note, the refit plan and its list of faults, wholesale statements and decks.
The 2025 template
Five numbered sections: a summary that opens with the headline figure and up to three recommendations with their savings in pounds, waste by shop and by category in pounds, a paragraph per shop, and last month's actions.
Last August's report
Written to the template by the former operations manager: all seven shops in pounds, a paragraph each, categories, and three recommendations up front.
Sam's unfinished August draft
The template's layout again, in pounds against July, with a note to himself to check one figure against Lina's pack first. It explains that Truro's bake comes from IT's corrected figure, not the till extract.
July's report
Shorter and in units, which she liked, but still seven shops, six columns, June as the comparison, a shift lead's name, three recommendations, and St Austell's gap noted below the table.
Money in pounds
Units, with the pounds left to Lina's pack.
Harbour Street in the total
The franchise shop is left out of all totals.
Recommendations before figures
Figures first, and one recommendation at the end.
More than one recommendation
Just one, kept short.
The total first, any gap further down
A missing week goes before or beside the figures it affects.
A paragraph for each shop
She can see from the table which shops moved.
Waste by product category
Left out while the crates can't be split by product.
Last month's actions
She goes through those with Sam in person.
A wide table or plus signs
Three columns of figures at most, with better or worse in words.
Last month as the comparison
Each shop against its own August last year.
A staff member named
Shops are named; people aren't.
Another team's figure corrected in the text
Use the right figure; the other team's slip goes to that team.
Food bank starts
Unsold food now goes into crates for the food bank instead of the bin. The weekly waste log still records the crates as wasted.
Tills upgraded
The tills move to a new version overnight. The shops record each bake on the tills, and for Truro's week 33 the till extract picked the bake up from two till records: 29,340 units instead of 14,670. IT confirms it in writing on 21 August, copying Lina.
Falmouth shuts
The shop trades on Monday only, then closes for its refit.
Truro half price
For week 33 the whole Truro counter goes half price from 3pm, and the bin barely fills all week.
Penzance photo
Penzance's shift lead fills in the week 34 waste sheet, headed as the week's total and dated that Friday, and photographs it.
Sheet thrown out
The paper sheet goes out with the rubbish that Saturday night.
Oven 2 fixed
Penzance's deck oven 2, out since early August, is working again.
Falmouth reopens
First bake back after the refit, a day after handover: Friday, Saturday and Sunday of week 35.
Photo found
The photo of Penzance's week 34 sheet goes on the shared drive, and Sam counts the week back in.
Redruth uncounted
Week 35 ends without a waste count at Redruth, which had an oven fault that week, so there's no figure to use.
Truro sheet in
Truro's week 33 waste sheet arrives, late, and goes into the shop sheets.
Finance pack
Lina's August pack goes out on calendar months, waste as a share of units delivered, and with Truro's week 33 still at 29,340, as the extract gave it.
The AI starts
Sam's first day of leave, and the AI's first on the report.
Redruth's week 35
The week was never counted, during an oven fault, as the waste log and the shop sheets both note. Counting it as zero waste would flatter Redruth, and leaving it out of both sides keeps the rate honest.
Truro's doubled bake
After the upgrade the till extract counted week 33 twice, as IT's email of 21 August and a photo of the till screen show. Lina's August pack still has 29,340, so a report that says the extract got it wrong tells Morwenna her finance pack is wrong: the kind of slip she asked to go straight to Lina.
Truro's late sheet
The minutes say it came in late; it did, and it's in the file.
Truro's half‑price week
Marketing calls it waste avoided. Morwenna ruled on 20 August that stock sold cheap is still a sale.
Penzance's week 34
8,462 baked and 1,782 binned. The sheet is dated the Friday but headed as the week's total, and its baked figure matches the production log for the whole week. Sam's reply on 28 August and his draft both count it back in.
Penzance's oven
As the reason the shop's waste was high, it's a fact for the report. What the repair cost isn't.
Falmouth's weeks
Not dropped, as Property wanted, and not scaled up to the six days of week 35 that the refit plan had promised, as Finance did. The 18 August minutes set the rule, and the chat on 28 August has the first bake back.
The food bank crates
3,757 units over the four weeks, in the food bank's collection log. Morwenna asked on 11 August for them as one figure beside the waste.
What to divide by
Finance's August pack divides by units delivered, a count kept by the wholesale side, which gives 8.8% for the six shops. Morwenna settled in July that this report measures waste against what was baked.
Harbour Street
It's a franchise with its own owner. Morwenna ruled on 12 August that it's left out of every total. A line on what the company sells it is allowed, though the published marks failed one placed after the recommendation.
Last August
In August 2025 there was no food bank, so food that now goes in a crate went in the bin. The page has to say so next to the comparison.
Falmouth
Better. Eleven trading days: all of week 32, the Monday of week 33, and Friday to Sunday of week 35. The three days after reopening ran higher, while the new oven baked too hot; it was adjusted on 1 September.
Newquay
Better. All four weeks, and the most food given to the food bank of any shop.
Penzance
Worse, the only shop to get worse. Deck oven 2 couldn't be used until 27 August. Week 34, taken from the photo, was its worst.
Redruth
Better, on weeks 32 to 34 only: week 35 was never counted.
St Austell
Better. All four weeks.
Truro
Better, and the lowest all month: 1.9% in its three ordinary weeks, and almost nothing binned in the half‑price week, which is in with its corrected bake.
All six shops
Better. Harbour Street, the franchise, isn't in the total, as Morwenna ruled on 12 August.
Most got the figures right, then wrote it the old way
Here's how each AI did on each part of the job. All 13 saved the right kind of file, so we've left that out.
The last column also counts the length, money and the single recommendation.
| AI | The figures | Her layout | What to leave out | |
|---|---|---|---|---|
| The main figuresof 7 | Glitches in the data, and what to countof 14 | Order, wording and bad news firstof 9 | Left out what she didn't wantof 11 | |
| GPT‑6 AstraLarge | 7 | 14 | 8 | 10 |
| GPT‑6.1 SolMid‑sized | 7 | 14 | 7 | 10 |
| Sonnet 5.5Mid‑sized | 7 | 14 | 7 | 9 |
| Opus 5.5Large | 6 | 13 | 8 | 9 |
| Fable 5.1Large | 7 | 14 | 8 | 8 |
| GLM‑5.3Small | 7 | 14 | 7 | 8 |
| GPT‑6 LunaSmall | 0 | 9 | 6 | 11 |
| Kimi K3Mid‑sized | 7 | 14 | 8 | 6 |
| Sonnet 5Mid‑sized | 3 | 13 | 5 | 8 |
| DeepSeek V4.1 FlashSmall | 3 | 14 | 7 | 7 |
| Opus 5Large | 7 | 13 | 6 | 5 |
| Qwen3.8‑27BSmall | 2 | 13 | 3 | 4 |
| GPT‑5.6 LunaSmall | 0 | 4 | 2 | 5 |
The darker the square, the more it got right. The AIs are in order of score.
The figures: seven had every main figure right
Handling the data was the part the AIs did best. Every AI used Truro's late sheet, divided by what was baked and kept Harbour Street out of every total, and 11 followed Morwenna over the 1 September minutes on Falmouth. GLM‑5.3 can't view images, so it installed text recognition to read the photos. Only the two Luna AIs got no main figure right.
Opus 5.5 got all but one. It left out Penzance's photographed week, taking the Friday date for a part‑week count. The sheet is headed as the week's total and matches the week's bake, but the test's own notes list the Friday date as a flaw in the photo.
Most wrote about every shop, and put the gap below the total
Morwenna had objected to a paragraph for each shop. And in July she forwarded the total to Lina, only to find a missing week noted at the foot of the page.
What the old reports doA paragraph for each shop, with what's missing mentioned further down.
What she'd made clear sinceShe can see from the table which shops moved. What's missing goes before or beside the figures.
In our words. Eight of the 13 still commented on almost every shop, and ten gave the total before saying Redruth's last week was missing. GPT‑6 Astra and GPT‑6.1 Sol flagged it beside the figures, and Kimi K3 marked Redruth's row as three weeks.
Every AI fixed Truro's week; nine said the till data was wrong
Every AI used Truro's real bake, 14,670. Nine then told Morwenna the till extract had counted the week twice, GPT‑6 Astra and GPT‑6.1 Sol among them. Lina's August pack still carries the doubled figure, so that sentence tells her the finance pack is wrong, in a report of Sam's: the slip she had asked to go straight to Lina.
Opus 5.5, Sonnet 5.5 and Opus 5 even noted her wish in their closing messages, then broke it in the report. Fable 5.1 worked it through and kept the slip out.
Money, length and the one recommendation
Seven put money in, six ran past a page (450 words, a limit we chose), six broke the one‑recommendation rule, and five never said that last August isn't like for like.
Leaving things out carries so much of the score that a short report can do well on it. GPT‑6 Luna added nothing she'd asked to keep out. It got no main figure right and left out Falmouth and Redruth, and it still finished above Kimi K3 and Opus 5, which had every main figure right.
Writing for one reader
- The AIs handled the messy data better than the reader's wishes.
Seven of 13 had every main figure right, and only three put the missing week where she'd see it before the total.
- Noting her wishes didn't always mean following them.
Kimi K3 noted most of them as it worked, then broke several in the report.
- Once her wishes were written into the request, the AIs followed nearly all of them.
With the request written that way, 8 of the 13 got everything right, though three still gave the total before the missing week.
What each AI did
Apart from the two misses at the top of this page, GPT‑6 Astra met everything we checked.
GPT‑6.1 Sol made GPT‑6 Astra's two misses and wrote one share as a whole number.
Sonnet 5.5 put its one recommendation before any figures and the missing week after the tables, and told her about the double count.
Opus 5.5 left out Penzance's photographed week, told her about the double count and ran its recommendation to two sentences.
Fable 5.1 followed July's layout but commented on most shops, ran 12 words past our limit, added a line after its recommendation and gave the total before the missing week.
GLM‑5.3 had every main figure right and followed July's layout, but commented on every shop, gave a cost in pounds and told her about the double count.
The tool GPT‑6 Luna used to read the chat cut out its middle, so the owner's rulings never reached it, and it left out Falmouth, as the 1 September minutes said.
Kimi K3 had every main figure right, then added a cost in pounds, a note on every shop, a word on July's savoury drops and the double count.
Opus 5 built the report from the old template, with money‑saving recommendations first and a comment on every shop, at nearly three times our 450‑word limit.
GPT‑5.6 Luna finished in about two minutes without opening the chats, PDFs or photos. It turned the analyst's unfinished draft into the report and got no figure right.
Sonnet 5 counted the food bank's crates as waste, so four of its seven main figures were wrong. DeepSeek V4.1 Flash gave most shops' figures in money, and Qwen3.8‑27B rebuilt the old template at more than twice our word limit.
Each AI had one go at the job, 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. 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
- The results above are for the request as a person would write it. Every AI was also tested on two other versions. One added a hint about how the company's expectations show up across its files. The other listed every requirement in plain words, with the right answer wherever a file would mislead, and what to leave out.
- What the extra help changed
- The hint raised most scores, Opus 5.5's to 99, though Sonnet 5 and GPT‑6 Luna did worse, bringing back sections the owner had dropped. With the rules written in, all 13 scored 94 or more.
- Checking
- Leaving out what doesn't belong is worth 60 points, the figures 20, her other wishes 15 and the file 5. A script checked 8 things, like the file's name, the length (450 words at most), names and number format, and two AI judges from different makers read the other 35. Where they disagreed, even with earlier rulings written into the questions, 22 times in 18 of the 39 tests of the three versions, the job's makers settled it in writing.
- Two ways of marking
- The nine AIs tested in September were marked before a review on 9 October changed three checks, and the four added in October under the new checks. Marked the old way, GPT‑6.1 Sol's whole‑number share would pass, and it would score 94, level with GPT‑6 Astra. Marked the new way, so would Fable 5.1: an oven out of service now counts as a plain fact, and a line may follow the recommendation. Either way, the top score is 94.
- Time and cost
- The 13 tests took between 2 and 66 minutes; the middle one took 7. For the 12 AIs with published prices, a test cost from 5 cents to about $4 at published per‑use prices for businesses.