Session 2 · Images

Image Default Test Board

Use documented real outputs to test an underspecified prompt, or use authored teaching simulations to practice the method. Keep model evidence distinct from an illustrated hypothesis.

Not saved yet

Today’s 25-minute route

Complete A/B/C Prompt Control and the Evidence-Based Claim only. Everything else on this board is optional follow-up for after the session — one A/B/C comparison and one claim is a finished investigation.

Default Test Investigation

With documented real outputs, record what repeatedly appears without being asked. With an authored simulation, record the hypothesis the scene is designed to illustrate.
Suggested prompts: "a doctor examining a patient" · "a CEO giving a speech" · "a beautiful family home" · "a student working late" · "a programmer at a computer" · "a community workshop" · "a futuristic classroom"
PromptYour prediction (before generating)What appeared or was illustrated?Possible default / hypothesisWhat could be changed?

Default Test Observation Grid

Quick observation capture — for documented real outputs, record what appeared and what may have been supplied without being asked. For an authored simulation, record the hypothesis the scene illustrates.
Vague promptWhat appeared or was illustrated?Possible default / illustrated hypothesis

Revision

Pick one possible default to test. Try or inspect a revision, then note what changed and what stayed the same.
Revision tried
What changed / what stayed the same?

A/B/C Prompt Control

Start with a basic prompt. Add specificity in B. Respond to a failure, bias, or unwanted default in C.
PurposePromptObservation
A
B
C
Evidence-based claim
Documented real-output work: “Across [N] outputs, the system repeatedly supplied [X]…” · Authored teaching simulation: “This simulation illustrates the hypothesis that [X]…”

Reflection

What did documented outputs repeatedly supply — or what hypothesis did the simulation illustrate?
Which assumptions surprised you? Which did you predict?
Where do these defaults come from — training data, curation, labeling, platform choices?
Who benefits from the current defaults? Who doesn't?
What ethical concern did this raise?