# The LLM reality checklist — field card

Companion artifact to curriculum module 1.01 (Introduction to LLMs &
the architecture of prediction) at
https://lyceumagents.com/curriculum/intro-to-llms/

One page on what the machine actually is, where it fails, and the two
gates that keep it honest. Everything else in the course builds on
these facts.

---

## The mechanism

- [ ] The model computes ONE variable: the next token. Fluency,
      persona, and reasoning are all structure built on that
      prediction.
- [ ] It stores no table of facts — it stores the statistical shape of
      its training text and reproduces that shape.
- [ ] The unit is the token (subword fragment / short word /
      punctuation). Tokens are the unit of cost and speed.
- [ ] The context window is the ONLY working memory. Instructions,
      documents, and conversation must all fit at the moment of
      prediction.
- [ ] The window overflows silently: oldest tokens drop with no error.
      A long session that "forgot" its instructions hit context overflow.

## The three failure classes

- [ ] **Confabulation** — where training data holds no answer, the
      model produces the statistically likely SHAPE of an answer:
      perfectly formatted citations for papers that don't exist.
      Countermeasure: name the ground truth; legislate the silence.
- [ ] **The grounding gap** — the model manipulates symbols with no
      contact with physical reality; it can give earnest instructions
      for impossible tasks. Countermeasure: never let fluent text stand
      in for a checked fact about the world.
- [ ] **Systemic bias** — the model reflects its training data's
      historical imbalances of race, gender, and class. It shows up in
      aggregates, not single outputs. Countermeasure: audit; never
      assume away.

## The two human gates

Input stage:
- [ ] Strip PII, credentials, and identifying metadata BEFORE the model
      sees anything.
- [ ] Supply the context the task silently assumes — inside the window.

Verification stage:
- [ ] Fact-check every claim, citation, and figure. Fluency is not
      evidence.
- [ ] A model's confidence describes its training data, never the truth
      of its claim.
- [ ] You own the output. The model multiplied your speed, not your
      responsibility.

---

Next: module 2.01 (Generative AI agent design) commits this machine to
one purpose through the block anatomy — its field checklist is
prompt-anatomy-checklist.md.