# Module 6 Instructor Notes

## Emphasis

This module should feel like applied neural-network engineering, not a survey lecture. Students need to explain why the method fits the data and why the evidence is strong enough for the stated claim.

## Common Misconceptions

- Treating high attention weight as guaranteed causal importance.
- Forgetting causal masks in generation settings.
- Ignoring quadratic cost as context length grows.

When these appear, redirect students to concrete evidence: tensor shapes, loss curves, validation behavior, diagnostic comparisons, or a written assumption log.

## Facilitation Plan

- Begin with a five-minute scenario readout and ask students to identify the decision being supported.
- Have students predict expected lab behavior before running code.
- Compare two student modifications and ask which changed a single variable cleanly.
- Require one limitation statement before accepting a performance claim.

## Evidence Expectations

A strong submission includes runnable code or clearly reproducible outputs, a comparison point, correct neural-network terminology, and an explicit limitation. A weak submission reports a score or screenshot without connecting it to the design decision.

## Extension Options

- Add a baseline that is simpler than the neural model.
- Repeat the run with a different seed and report stability.
- Add a diagnostic plot or tensor-shape table.
- Write a short model-card section for the module artifact.