Limits

Small local models are useful, and they make mistakes in predictable ways. This page shows five of them, each with a captured example, so you can recognize them in your own work.

All captures come from nishad/apertus:8b at temperature 0.8 on 15 October 2026, unless stated otherwise. Temperature is the setting that controls how much the answers vary.

1. An invented place of publication

This one comes from the notebooks in Part 2. Notebook 02 reads catalog cards and fills in a record that follows a schema, a fixed list of the allowed fields and what goes in each. Card 03 has no place of publication:

Beauvoir, Simone de
Le deuxième sexe. I, Les faits et les mythes
Gallimard, 1949
395 p. ; 19 cm
Collection Folio essais

The instruction said to use an empty string (a blank value) for any field the card does not state, and never to guess a place. The model returned this (from the executed notebook 02, 15 October 2026):

{'title': 'Le deuxième sexe. I, Les faits et les mythes', 'creators': ['Beauvoir, Simone de'], 'place': 'Gallimard', 'publisher': 'Gallimard', 'date': '1949', 'language': 'en', 'identifier': ''}
place should be empty; the model said: 'Gallimard'

The model put the publisher’s name into the place field. A second error sits in the same line: the language of this French book is marked ’en’, the code for English. Both values look normal. Nothing in the answer signals a problem.

2. A fabricated identifier

The probe from Exercise 6:

Prompt:

What is the ISBN of the book "Das Parfum" by Patrick Süskind, published by Diogenes?

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

The ISBN of the book "Das Parfum" by Patrick Süskind, published by Diogenes, is 9783257061675.

The ISBN is not valid: its check digit is wrong, and the DNB catalog has no record of it. The same number with the correct check digit is a different book, “Verrechnet” by Dick Francis. The model gave one number and showed no doubt. The DOI in the same exercise was real but belonged to a different article. The full check is on the exercises page.

3. A rule missed in a calculation

A late-fee problem with a maximum fee per item:

Prompt:

A library's late fee policy: books cost $0.25 per day, DVDs cost $1.00 per day, and the maximum fee is $5.00 per item. Sarah has 2 books that are 12 days late, 1 DVD that is 8 days late and 1 book that is 3 days late. How much does Sarah owe in total?

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

To calculate Sarah's total late fee, we need to calculate the late fees for each item and then add them together.

Books:
Sarah has 2 books that are 12 days late. The late fee for each book is $0.25 per day.
So, the total late fee for the books is 2 * $0.25 * 12 = $6.

DVD:
Sarah has 1 DVD that is 8 days late. The late fee for the DVD is $1.00 per day.
So, the total late fee for the DVD is $1.00 * 8 = $8.

Book (3 days late):
Sarah has 1 book that is 3 days late. The late fee for this book is $0.25 per day.
So, the total late fee for this book is $0.25 * 3 = $0.75.

Total Late Fee:
Now, let's add up the total late fees for all items: $6 (books) + $8 (DVD) + $0.75 (book) = $14.75.

Therefore, Sarah owes a total of $14.75 in late fees.

Each multiplication is correct: 2 books at 12 days is 2 x 0.25 x 12 = $6. But the policy caps the fee at $5.00 per item, so the DVD should cost $5. The model ignored the cap and charged $8. The right total is $6 + $5 + $0.75 = $11.75. The model’s $14.75 is $3 too much. The steps are written out and look careful, which makes the mistake easy to miss.

4. A missed word limit

Prompt:

Summarize in no more than 15 words: The library will close for two weeks in August to replace the roof, move the periodicals to the basement and install a new self-service checkout system. All loans are automatically extended until the day after reopening.

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

The library will close for two weeks in August to replace the roof, move periodicals to the basement, and install a new self-service checkout system. All loans are automatically extended until the day after reopening.

The prompt asked for no more than 15 words. The answer has 35, and it is a near copy of the original text. The model has no word counter. Earlier on the techniques page, a request for exactly 20 words produced 21.

5. A wrong term in translation

The prompt below translates German economics terms into English. The reference is the English label in the STW Thesaurus for Economics, version 9.20.

Prompt:

Translate these German economics terms into English, one per line: Schattenwirtschaft, Lenkungsabgabe, Nutzungsdauer, Urheberrecht, Eigentum

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

1. Shadow Economy
2. Regulatory Fee
3. Service Life
4. Copyright
5. Ownership

Four of the five are acceptable English, but each differs from the label in the thesaurus. Lenkungsabgabe is a different kind of problem. In STW it is a levy (a tax or charge) whose main purpose is to steer behavior, and the model’s “Regulatory Fee” does not say that. A reader who does not know German would not notice. When a vocabulary sets the label, the model’s translation is no more than a suggestion.

Two checks anyone can do

1. Ask for the source, then look it up. Ask “Where does this come from?” and check what you get against a catalog, Crossref, the thesaurus or the original text. If the model names a source, confirm that it exists and says what the model claims. If it cannot name one, treat the answer as unverified.
2. Ask the same question twice and compare. Run the prompt again in a new chat. At temperature 0.8, the default in the apps, answers can vary between runs. Facts the model has actually learned tend to come back the same. A guess changes between runs. If two runs disagree, treat both answers as unverified.

Neither check proves an answer is right. They only tell you where to look.

From here to Part 2

In Part 2 you meet the tools that make these errors visible: a schema that rejects a wrong shape, a validator (a check written in code) that rejects a place that is the same as the publisher, and a human who approves a record before it is used. See Part 2.