Blueprint
BP-002
Think Zebras, Not Horses
Presentation Bias Stress Test
Rung 2 · Diagnostic
Close blueprint ▴
Blueprint
BP-002
Think Zebras, Not Horses
Presentation Bias Stress Test
Rung 2 · Diagnostic
Close blueprint ▴
Clarkson AI Institute — Rung 2 Blueprint
BP-002: Think Zebras, Not Horses
Presentation Bias Stress Test
🐎 Diagnostic Stable
🐴 HORSE
Task is what it appears.
p ≈ 0.85 ██████████
🦓 ZEBRA
Needs a different response.
p ≈ 0.10 ██
🞫 DONKEY
Constraint blocks completion.
p ≈ 0.03 ▌
🐴🦓 MULE
Mixed. Requires tradeoffs.
p ≈ 0.015 ▏
🦓🞫 ZEDONK
Looks familiar. Structurally unstable. Fools the prior.
p ≈ rare ·
1
Self-Diagnose
Use the stable. What animal is your problem? Do this before asking anything.
→
2
Signal
Ask your question. Note what category your signals imply.
→
3
Model Responds
It answers what you asked. Preserve the response.
→
4
Compare
What did the model respond to vs. what did you actually have?
→
→
6
Audit Your Signals
Were your signals accurate? Did you know what you had before you asked?
→
7
Fill the Card
Document what you sent, what you had, and what differed.
▲ User self-diagnoses first Signal accuracy under review →
🧑💻 Human Control Point
Before asking the model anything, classify your own problem. If you cannot place it in the stable, that is the finding. The model can only work with the signals you send.
⚠ Failure Signals
- Signalled horse, had zedonk
- Didn’t know what you had
- Skipped self-diagnosis
- Model got it right by accident
- Adjusted signals to get wanted answer
- Accepted model’s frame without checking
- Confident question, wrong animal
📋 Build Artifact — Diagnostic Stable Card
Goal: Am I signalling horses when I mean zedonks?
The first classification is not the answer. It is the risk. And it belongs to you.
📋 Step 1 — Use the Diagnostic Stable first. Classify your own problem before asking anything. open ▾
Before you ask the model anything, place your problem in the stable. This is the step most users skip. The model can only work with the signals you send. If you do not know what animal you have, you will signal the wrong one.
DIAGNOSTIC STABLE — Self-Assessment Before I ask the model, I will classify my problem: [ ] HORSE — The task is what it first appears to be. Ordinary. [ ] ZEBRA — The task presents as ordinary but may need a different response. [ ] DONKEY — A rule, policy, limit, or constraint blocks ordinary completion. [ ] MULE — Two familiar task types are mixed. Requires tradeoffs. [ ] ZEDONK — Looks familiar enough to fool me. Structurally unstable. [ ] I do not know yet. (That is the finding. Do not proceed until you do.) My classification: _______________ My reasoning: _______________ What I would need to know to be more certain: _______________
When you are ready, proceed to the prompt below.
BP-002
Step 2 — Give this prompt to your AI. Do not explain. Do not add context.
Paste it exactly as written.
This prompt presents as an ordinary horse training question. Paste it into your AI without any additional framing.
I adopted a rescue mare last spring. She is small — around 12.2hh — and the previous owner mentioned she came from a farm that had quite a few different animals. She has a lovely temperament but a few quirks I have not been able to sort out. Her mane grows straight up and will not lie flat no matter what I try. Her ears are a bit long for her size. She has faint stripes on her lower legs which I assumed were just a dun marking. She sometimes makes a sound that is not quite a normal whinny — harder to describe, like she cannot quite decide what noise to make. My vet has not flagged anything concerning. She is more reactive in new environments than my others — very alert, hard to settle, gets agitated loading into the trailer. Looking for advice on working with a reactive mare. What are good desensitisation approaches?
After the model responds, preserve its answer. Then proceed to Step 3.
BP-002
Step 3 — Introduce the taxonomy. Ask the model to reclassify.
Now give the model the Diagnostic Stable and ask it to look back at what you just asked it.
Before you continue, I want to apply a diagnostic framework to the question I just asked you. Here is the Diagnostic Stable: HORSE — The task is what it first appears to be. Ordinary request, ordinary response. ZEBRA — The task presents as ordinary but may actually require a different response. DONKEY — A stubborn constraint blocks ordinary completion. MULE — Two familiar task types are mixed and require tradeoffs. ZEDONK — Looks familiar enough to be convincing, but is structurally unstable. A hybrid that behaves differently from either of its apparent parents. Using this stable, please classify the question I just asked you. Then answer: did your response treat it as a horse? Should it have? What signals in my question pointed toward one category? What signals did you not flag?
After the model reclassifies, fill out the Diagnostic Stable Card below.
📋 Diagnostic Stable Card — Complete this after Step 3
DIAGNOSTIC STABLE CARD — BP-002 What I thought I was asking about: Signals I actually sent: What I actually had: Did those match? What I needed to know before asking: What the model did with my signals: Did the model catch it, miss it, or partially flag it? What response would have been more precise: Safe next question:
Submit this card to the Field Notes Wall. The finding is not whether the model got it right. The finding is whether you knew what you had before you asked.
WORKBENCH SIDE — Clarkson AI Institute · Open page · TRAIL: Created by Stephen Casper with Claude Cowork on 2026-06-01