Part 1 of 5 · 2 min read
Overview
In this module you will
- Explain two proposed causes of hallucination and why retrieval reduces it without removing it.
- Name four places to act (the context, the instructions, the answer, the decision to answer) and the failure each one targets.
- Explain self-consistency and use disagreement between samples as a signal to refuse.
- Read a risk and coverage trade-off: fewer wrong answers means fewer answers.
- Design an evidence-based plan: change one thing at a time, and measure both errors and refusals.
Module 5 gave you a way to check an answer against its sources. This module is about producing fewer bad answers in the first place, and about knowing when a model is not on firm ground. Agents make this harder: they run for many steps, and an early wrong claim becomes the starting point for later ones. A system that is "mostly right" but never signals doubt is hard to trust or to automate. The controls below are cheap compared with the cost of a wrong action.
About the examples
The "model" in this module is a sampling imitation: it builds candidate answers from the source sentences and picks one at random, with odds that rise with word overlap with the question and fall as temperature rises. Two behaviours are assumed and built in: a sampling temperature and a penalty for sentences in the middle of a long context (the shape described in "lost in the middle" research). It is also noisier than a good real model, so its baseline error rate (about 40 percent for the sloppy setup) is far higher than you would see in practice. The questions are the 24 from module 5, run 40 times each (960 trials per setting). Read the numbers as showing which direction each control pushes, not how much it would help a real model. No real model or service was called.