# Module 2 practice · a generator's decisions

Attempt these before reading [the solutions](../solutions/02-generator.md). Use `methods.py decoding` to inspect a trace, but calculate before running.

**A. A changed training table.** Train conceptually on “desk opens weekdays” three times and “desk closes holidays” once. What are the next-word probabilities after `desk` at temperature 1? What does greedy decoding select? At temperature 0.5, compute the two normalized weights. Does any result establish the actual opening policy?

**B. Replay and stopping.** In a Python session, use `train_bigram` and `decode_bigram` on the four sentences above. Generate greedily from `desk`, then from the unseen history `laptop`. Construct `{"again": {"again": 1}}` as a continuation table and decode from `again` with `max_steps=4`. Explain each stopping reason. Do not remove the bound to see how long a loop runs.

**C. Attention arithmetic.** For query 1, keys `0,0` and values `4,12`, calculate the normalized weights and weighted value. What happens to the weights if both scores receive the same added constant? Explain why neither calculation establishes which value is factually correct.

**D. A plain-language distinction.** A colleague says putting a revised handbook in a prompt “retrained the model.” Write a correction that separates source retrieval, inference context and parameter updates. Explain why our word-token count cannot establish a future model's context usage.

Self-review: a strong response gives correct denominators, reports all three stopping cases, computes the average and names the limits of the toy model. This is practice feedback, not platform credit.