Week 7: Evaluating GenAI Outputs

Module: Evaluate | Deliverable: Prompt and output evaluation memo (due before Week 8 begins) | See assignment →

Learning objectives

By the end of this week, you should be able to:

Distinguish five evaluation targets: model, prompt, output, workflow, safeguard
Create a task-specific rubric with scored dimensions and calibrated anchors
Score GenAI outputs independently and compute multi-rater agreement
Compute Cohen’s κ across multiple rater pairs (human-human and human-LLM)
Test prompt sensitivity: does the LLM-as-judge produce consistent scores?
Compute minimum detectable difference (MDD) and determine whether score differences are meaningful
Design test cases with typical, edge, and adversarial inputs linked to rubric criteria

Sessions

Session 1 · Class 13

What does "good" mean? Rubric design and calibrated scoring

Five evaluation targets, a 10-dimension evaluation template, and rubric design for a specific task. Then a calibrated bake-off: 3+ students score the same three AI outputs independently, producing the data you'll analyze in Session 2.

Five evaluation targets Rubric design Calibrated multi-rater bake-off

Session 2 · Class 14

Measuring reliability and designing test cases

Put numbers on Session 1's disagreements: multi-rater Cohen's κ (human-human and human-LLM), prompt sensitivity testing (is the LLM judge consistent?), and minimum detectable difference - the threshold below which score differences are just rater noise. Then design test cases for your final project.

Multi-rater kappa Prompt sensitivity (3 runs) Minimum detectable difference Test case design

Prepare before class

Before Session 1:

Review the evaluation frameworks guide - rubric building and kappa interpretation
Think about what task from your final project you’d want to evaluate

Before Session 2:

Bring your Session 1 rubric and scores - you’ll use them for the kappa and MDD exercises

Key concepts

Concept	What it means
Multi-rater kappa	Cohen’s κ computed across multiple rater pairs (human-human and human-LLM) to test whether agreement is consistent
Prompt sensitivity	Running LLM-as-judge 3x on the same output at temp=0. If scores vary, the evaluation itself is unreliable.
Minimum detectable difference (MDD)	The smallest score difference that exceeds inter-rater noise. Differences below MDD are not meaningful.
Calibrated bake-off	3+ raters score independently before comparing - produces data for quantitative agreement analysis

Readings and resources

LLMs Visual Cards: LLM-as-a-Judge - the position, length, and style biases this week’s prompt-sensitivity test is designed to catch

Deliverable

Prompt and Output Evaluation Memo

8 components | 30 points | Due before Week 8's first class | Submit via Canvas

Rubric, scored outputs with justifications, multi-rater agreement (human-human and human-LLM κ), prompt sensitivity analysis, minimum detectable difference, and evidence-based recommendation.

Full prompt and rubric →

Looking ahead

Next week: Failures, Incidents, and Risk - real AI incidents, subgroup error analysis with formal disparity metrics, and the sociotechnical failure chain.