Evaluation and claim boundaries

Three levels of claims

Determinism	The same bound state produces the same gate result.
Audit	Input, gate trace, and result show whether the transition was admissible.
Performance	Task benefit must be measured against alternatives.

How to verify, not just believe

We do not ask you to take our word for it. verify.py runs 35 automated checks directly on the engine — not on promises, on code. You run them yourself, locally, in seconds:

python -m fpf_thinking_map.verify

And when you build your own semantic map on top of the library — your own contexts, guards, and rules — you get the same verification level for free: the same determinism and audit properties apply to your map, not just the shipped example. You are not buying a promise. You install an engine you can verify from the first line of code.

And because we ask for verification, we are equally direct about where it stops:

What the product does not claim

It does not prove semantic correctness, does not guarantee safety, and does not make the LLM deterministic. It makes one chosen part of control explicit and mechanically verifiable.

And just as importantly: it does not interfere with the model's own generation (LLM). It takes over bookkeeping — state, rules, history — so the model does not have to keep that in memory itself; that is the cognitive load we actually remove. But it does not rewrite, edit, filter, or semantically reason in place of the model. Its role ends at checking the discrete move, before and after: what the model proposed, as language and as meaning, stays the model's own.