Care inspection readerSYNTHETIC

A worked example

An invented Safe page in CQC's layout, and the chosen model's one reading of it, shown as it came, mistakes included.

Invented report. Every word of it is made up and describes no real service. No real report text appears anywhere on this site.

How to read it

Each card below is one section: the inspectors' score, and whether the model flagged it. The model saw only the section's paragraphs, never its name, its score or the rest of the report. Highlighted words are the quotes its findings rest on, each found in the text by code; the small number after a highlight is its finding. Under each quote is the model's own one-sentence reason. A finding the model filed under another statement than its section's is tagged Another statement; the section still counts as flagged.

A finding whose quote is not in the text is dropped and counted, never repaired. This reading dropped 0. The model answered every section.

Safe and effective staffing

Score 1 · significant shortfallsFlagged

The provider did not plan rotas with any travel time between calls.finding 1 Records for the four weeks before our visit showed that almost a third of morning calls started more than 45 minutes after the agreed timefinding 2, and 19 calls lasted less than half of their planned lengthfinding 3.

One relative told us, "Mum waits in bed until nearly eleven some days, and nobody rings to say they are running late." Two new carers had started visiting people on their own before their references and criminal record checks had come back.finding 4

Medicines optimisation

Score 2 · some shortfallsFlagged

People mostly received their medicines as prescribed, and carers completed the electronic record at each call. However, one person's blood thinner was prescribed for 6pm and was given at any time between 4pm and 9pm, because their teatime call had no fixed time.finding 1

Carers applied creams, but three people had no body map showing where each cream should go.finding 2 The office looked at the records every week and had noticed some of these gaps, but had not acted on them.finding 3

Involving people to manage risks

Score 2 · some shortfallsFlagged

Carers knew people well and could tell us how each person liked to be helped. Plans for people living with diabetes set out the signs of a high or a low blood sugar and what carers should do.

A person who had fallen twice in the autumn still had a moving plan written before the falls. It did not say that they now needed two carers to help them stand.finding 1

Learning culture

Score 3 · a good standardFlagged

Flagged, though the inspectors scored it 3: one of the flags on good sections.

Carers logged accidents and near misses on their phones, and the registered manager read each one within two working days. Lessons were talked through at team meetings and in a monthly newsletter.

One form from August recorded a fall but not what was done about it afterwardsfinding 1; the manager added the missing actions while we were there. People and their families felt able to raise a worry and said the office listened.

Infection prevention and control

Score 4 · an exceptional standardNot flagged

Carers always had gloves, aprons and hand gel, and spot checks found they used them well at every call we watched. Two carers trained as infection leads visited each new starter on their first shifts to check their practice.

When a stomach bug spread through one village last winter, the agency kept the same small team on the homes affected until it had passed, and nobody else became unwell.

No findings.

The reading

Read once by Qwen3.5-4B on one laptop GPU, at temperature 0. To read the same page yourself:

uv run inspection-reader read --model qwen3.5-4b site/example-page.html
Run details

Prompt SHA-256 508785a5c047969f9a22986a553d2f7bceb324707be4b38564fb1c9f0d26210b
Model file SHA-256 00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4

See the results: how the model did on real reports, against the inspectors' scores.