← Back to the assistant

How answers are made

The short version

Claude, the model built by Anthropic, writes every answer — but not from memory. Your question is sent to the model together with a curated corpus of law that we maintain, text that carries its own citations, and the model is told to answer from that text and to cite it. The citations are the part of an answer you can check, and you should.

What it is reading from

We call that text the corpus. We curate and maintain it, and it is why the assistant can name a controlling rule instead of paraphrasing a general impression of employment law. It is also the ceiling. The model is instructed to work from the corpus rather than its own recall, so a question the corpus does not cover is a question the assistant should decline.

Where it has coverage

The corpus covers United States federal law and seven states, and nothing beyond them. If you ask about a state that is not on this list, you are outside what this tool was built for.

It is told to decline, not to guess

When the corpus does not support an answer, the assistant is instructed to say so rather than produce a confident-sounding one. A decline is recorded as a decline, in the internal log we read to find the topics this thing handles badly.

Figures that change every year

Some numbers move annually — thresholds, rates, caps. For those, the assistant may run a web search to check the current figure. That search runs as a tool hosted inside Anthropic's API; we do not hand your question to a separate search company.

Open the citations

Answers cite the rules they rest on, and the citation is the part you can actually verify. Before you act on anything here — a termination, a final paycheck, a denied leave request — open the citation and read the rule yourself. If part of an answer has no citation you can follow, treat that part as a starting point for research rather than as an answer.

It can still be wrong

Statutes get amended, cases get overturned, agencies reissue figures, and the text we hold can lag any of that — or simply be wrong. We do not publish an accuracy rate and you should not assume one. This is general HR information to orient you and show you the source, not a substitute for a lawyer licensed in your state.

Tell us when it is wrong

If an answer was wrong, out of date, or hedged into uselessness, email maggievinas@outlook.com. Send the question you asked and the state you asked about.

What we store while doing this

Your questions and the assistant's answers are stored — in your own conversation history, and in an internal log we read to find weak coverage. What is kept, for how long, and who else is in the path is set out on the privacy page. Keep questions general: do not paste employee names or medical details.

What this is not

Using this tool creates no attorney–client relationship, and your conversations with it are not privileged. It knows nothing about your situation beyond what you type into it.