The system behind this note reads a property's official documents and drafts the appraisal report that a human appraiser then checks and signs. Hard rules govern what a machine is allowed to put on a page somebody signs. This note covers a narrower boundary: which parameters its Anthropic client puts into a request, and where the reason for leaving the others out is written down.

Which request fields does the client send?

The request body is limited to four parameters. Three common ones are deliberately left out:

Sent in the requestDeliberately not sent
modeltemperature
max_tokenstop_p
systemthinking
messages

This covers request parameters only. Transport details such as headers and authentication are outside the scope of this note.

Why are temperature, top_p and thinking left out?

A comment in the client code records that the model in use rejects sampling parameters. For that reason temperature, top_p and thinking are deliberately never sent. The evidence establishes the exclusion and the recorded reason. It does not describe how the exclusion is enforced, so this note makes no claim about that.

A common reason this kind of exclusion matters: engineers moving between providers bring habits with them. They set temperature to zero for repeatability, cap top_p, or turn on extended reasoning. Each habit can be valid for one model and rejected by another. Leaving these parameters out is a decision, and a decision is only durable if the reason survives along with it.

Why record a provider quirk in the code rather than only in a changelog?

A changelog entry describes what changed at one moment. A code comment stays with the code and is visible to anyone who opens the file. These do different jobs, and provider constraints benefit from the second.

Think about who touches an LLM client months later. It could be an engineer who wants to add a temperature setting for an experiment, a reviewer asking why the request is so sparse, or someone moving the client to a different model. Most of them will read the client before they read the project history. A constraint recorded in the client has a better chance of reaching them before they reintroduce a parameter the model rejects. Commit history is still a useful supporting record, but it has to be looked up on purpose.

The recorded constraint is about the model in use. As a practical reading, a model change is therefore a sensible moment to check whether the exclusion still applies, in either direction.

Why is max_tokens a 2048 constructor default instead of a per-call value?

The default output budget of 2048 tokens is set when the client is constructed, not supplied as a per-call value. That puts the baseline in one place, next to the rest of the client's configuration, rather than tying it to individual requests.

In general, token limits chosen at each call tend to drift. One call gets raised while debugging, another copies an old value, and eventually nobody can say what the effective ceiling is. A constructor default gives one answer to the question "how long can a response be by default?"

What should teams building LLM clients take from this?