Business operations · Task 3 of 34
The monthly temperature register
Compile a bakery chain's monthly fridge and freezer temperature register from 134 working files, where faulty sensors, odd logger settings and routine events all look like real breaches.
The task
It is mid-September at a family bakery chain in Cornwall. The food safety coordinator is off sick, and the food safety officer needs the August temperature register: for every fridge, freezer, display and hot-holding unit, whether it stayed within its limit, each breach, and each data problem corrected. The agent gets about 134 files: logger exports, calibration certificates, engineer job sheets, scanned handwritten sheets, photos, emails, chats, minutes, the two previous registers and a half-built draft. Nobody is available to answer questions.
The prompt only asks for the register. The real test is not taking the data at face value: telling a faulty sensor from a warm fridge, a planned clean from a breach, a misread logger from a real reading, and a trustworthy handwritten entry from one that isn't, at volume and in the previous registers' exact format.
What it tests
Spotting faulty and decoy data
Disconnected probes, stuck loggers, spikes and duplicate uploads read as real temperatures, and cleans, closures, deliveries and defrosts look like breaches, unless the agent checks them against the rest of the files.
Reading messy documents
Scanned and photographed paper sheets carry coffee rings, rewritten values, shifted times and decimal commas, with no text layer to search.
Normalising before counting
Not every logger reports the same way and not every clock can be trusted, so the data needs fixing before a single breach can be counted.
Matching the house format and practice
The register has to come out exactly as the previous ones did, in its files, columns and dates, and in what the house counts as one breach.
Weighing sources that disagree
Superseded limits, mislabelled paperwork and colleagues' own accounts sit beside the readings, and the agent has to work out which one wins.
Challenges
1Faulty readings that look real
A disconnected probe, a stuck logger, a single spike and a duplicated upload all sit in the exports as ordinary numbers, as they do in any real fleet of loggers.
2Wrong scale, wrong clock
Not every logger reports in plain degrees, and not every timestamp means what it seems, as in any mixed fleet of devices.
3Decoy breaches
Closures, deep cleans, deliveries, defrosts and genuine agency cover all look like problems in the data, and the files show why each is not.
4Messy paper
Stains, corrections, shifted times and entries that are not quite what they seem, as on any paper log kept by a busy kitchen.
5Superseded rules still on file
An out-of-date version of the standard is still in the folder, and not every limit stays the same all month.
6Accounts that play things down
Colleagues' descriptions of a fault, and requests to leave something off, do not always match what the readings show.
Results
| Model | Standard | Hinted | Told everything | Time (standard) | Cost (standard) |
|---|---|---|---|---|---|
| Sonnet 5.5 | 76% | 83% | 92% | 27 min | $8.10 |
| Opus 5.5 | 75% | 75% | 95% | 23 min | $8.34 |
| Fable 5.1 | 65% | 87% | 0%* | 51 min | $21.08 |
| GPT-6 Astra | 50% | 55% | 89% | 47 min | $28.61 |
| Sonnet 5 | 48% | 42% | 67% | 46 min | $6.01 |
| GPT-6.1 Sol | 43% | 39% | 0% | 36 min | $1.58 |
| DeepSeek V4.1 Flash | 41% | 53% | 82% | 32 min | $0.71 |
| GLM-5.3-Flash | 0% | 0% | 0%* | 66 min | $0.78 |
| GPT-5.6 Luna | 0% | 0% | 64% | 84 min | $0.82 |
| GPT-6 Luna | 0% | 0% | 0% | 5 min | $4.81 |
| Opus 5 | 0%* | 80% | 0%* | 123 min | $27.72 |
| Qwen3.8-27B | 0%* | – | 0%* | 180 min | $1.60 |
One run per model and setup. 100% = everything right (or, for a plan scored in money, the best plan); 0% = the obvious answer or worse.
Setups: Standard, the job as asked; Hinted, plus one hint; Told everything, every rule or fact given up front: the control that shows the task is fair.
* no score: the run handed nothing in or hit the spend cap or ran out of time, so it counts as 0%.
What models got right and wrong
- Sonnet 5.5 led the standard setup at 75.7%, just ahead of Opus 5.5 (74.8%) and Fable 5.1 (65.0%). Both leaders read every scan and photo and rebuilt the register from the raw data in under half an hour.
- One period decision cost the two leaders most of their marks. Asked to add a missing week from the month before, Sonnet 5.5 and Opus 5.5 counted the whole summary over five weeks instead of the month; GPT-6.1 Sol put that week in files of its own.
- Every standard register logged breaches that never happened. GPT-6.1 Sol (42.7%) entered every planned defrost and one-reading blip, about twelve times the real count, and GPT-6 Astra (50.5%) about eight times.
- Producing the register at all was a hurdle. GPT-6 Luna and GPT-5.6 Luna saved one Excel workbook instead of the three files (0.0%), GLM-5.3-Flash left only images of the scans, Opus 5 reached the $25 spending cap before writing anything, and Qwen3.8-27B ran out of time.
- A torn paper page with shifted times beat every register in every setup.
- The hint added 22.3 points for Fable 5.1, which posted the best hinted score, 87.4%, and 12.6 for DeepSeek V4.1 Flash. It cost Sonnet 5 5.8 points and GPT-6.1 Sol 3.9.
- Told everything, Opus 5.5 led at 95.1%, then Sonnet 5.5 92.2% and GPT-6 Astra 89.3%; no run got every point. GPT-6.1 Sol spent its 150 turns copying paper sheets by hand and wrote nothing, and GPT-6 Luna saved empty files after 42 seconds.
Test environment
- Kind
- Document folder: the working files of Tegen Rowse, Food Safety Coordinator at Pellow & Vane Bakehouse, a Cornish bakery chain with six shops, who is off sick
- Agent's interface
- A terminal in a fresh Docker container (Debian, Python 3.13 with pip, Node 22), driven by the agent's own command-line tool: Claude Code, Codex CLI, Prime Agent or Qwen Code
- Documents
- 134 files in 21 folders (Loggers, FoodSafety, Kernow Cold, Scans, Photos, Registers, Email, Chat, Minutes, Operations and others): 50 PDF, 21 DOCX, 18 CSV, 17 XLSX, 16 EML emails, 6 JPG, 4 TXT chat exports, 2 PPTX. The seven scanned temperature-sheet PDFs, a two-page PDF of chat photos and the six JPG photos have no text layer. No file arrives later
- Applications
- None. No document tools (no LibreOffice, Pandoc, OCR, ImageMagick or pdftotext) and no Python packages: the agent installs what it needs
- Web
- None. The agent's own web search and fetch are off; the internet is on for installing packages
- People
- None. Nobody answers questions
- Time
- None. The prompt gives the date: Monday 14 September 2026
- Actions
- Read the files and write to
output/only. Nothing can be sent - Deliverable
- Three CSV files in
output/, in the June and July registers' format:FSR-2026-08-breaches.csv,FSR-2026-08-corrections.csvandFSR-2026-08-summary.csv