# Lesson 02 — Assumptions of a model > Series: [index](README.md) · prev: [01](01-evidence-before-trust.md) · next: [03 — Counsel](03-counsel-choosing-an-approach.md) Every AI component rests on assumptions. The knowledge store records each as **held** (default) or `status="doubted"`. `trust = held / total`, so doubts discount the score until verified. | Component | Example assumption | Where to look | |---|---|---| | Neural nets | Inputs are scaled the way training expected | [neural-nets](../../artificial-intelligence/neural-nets/README.md) | | Character recognition | Glyphs are clean; scripts are European | [character-recognition](../../artificial-intelligence/character-recognition/README.md) | | Large language models | Training data and model licence permit your use | [large-language-models](../../artificial-intelligence/large-language-models/) | ```xml Input scaling matches training. Test data resembles production data. ``` The loader fails closed on any status other than `held` or `doubted`. ## Exercise Write four assumptions for an LLM connector, mark one doubted, and compute trust. *Answer:* 3 held of 4 gives 0.75.