Specimens require JavaScript for editing
The entire exhibition framework, examples and field guide remain available below without scripts. See the field guide for manual judging worksheets.
THE CREATIVE VARIANCE EXHIBITION ● OFFLINE CAPABLE
THE BLACK
BOX.
One prompt. Four plausible outputs. Six lenses. No borrowed authority.
What happens when we stop asking which model is “best” and begin asking what the work actually does? This is a transparent editorial demonstration—not a live AI benchmark.
OPEN THE COMPARISON CHAMBER ↘A
B
C
D
DIFFERENCE ≠ ERROR
CONSENSUS ≠ TRUTH
SCROLL TO INVESTIGATE ↓ ILLUSTRATIVE ENTRIES · NO MODEL ATTRIBUTIONS
01 / THE CHAMBERHold the prompt constant.
Hold the prompt constant.
Change the answer.
Every shipped entry is fictional editorial sample work. Edit freely; all changes and votes stay in this browser unless you export them.
Illustrative dataset loaded.
PROMPT CONTROL · VERSION 1
A new prompt version does not regenerate existing outputs; the entries remain illustrative until manually edited or imported.
OUTPUT SPECIMENS
02 / YOUR RESEARCH RECORD
Make a comparison defensible.
Scores are local, subjective ordinal ratings (1–5), not measurements of actual model capability. A vote is a preference in this browser session, not public polling.
03 / FAILURE ARCHIVEKeep the wreckage.
Keep the wreckage.
It teaches.
Document what broke instead of quietly removing it from the gallery.
Enable JavaScript to manage the editable archive. Example cases and method remain in the field guide.
04 / THE METHODSix lenses.
Six lenses.
Zero magic scores.
Novelty
Does the response find a fresh mechanism, not merely fresh adjectives?
Coherence
Do its parts fit together without contradictions?
Usability
Could a visitor understand what to do and what changes?
Reproducibility
Are conditions, dependencies and steps clear enough to repeat?
Accessibility
Are keyboard, text, motion and contrast needs addressed?
Functionality
Does the described result truly operate? Text alone is not proof.
The trap behind a leaderboard
A prompt defines a task, not a universal intelligence test. A polished paragraph may win for narrative clarity but lose for execution. A prototype may be original and still inaccessible. When the evaluation criteria change, the apparent winner can change with them. That is a property of judging, not evidence of the superiority of a particular model.
Evidence boundary: The four named specimens in this edition are author-written demonstration artifacts; provider labels are explicitly fictional. Their rubric values illustrate how judging works, not measured model performance. Manual scores, votes and your edits are locally entered data. The interface makes no network requests.
Three guided experiments ↗
A · Blindfold the brand. Score each specimen without its label, then reveal the labels. Notice what you expected, but do not infer an experimental effect from one encounter.
B · Shift the brief. Add a stricter accessibility or reproducibility constraint. Commit a prompt version, annotate which samples would need revision, then rescore.
C · Preserve the failure. Deliberately test a claim that is not implemented. File it in the failure archive with a reproducible observation.
05 / TAKE THE FIELD GUIDELeave with a method,
↓ DOWNLOAD FIELD GUIDE (PDF)Leave with a method,
not a verdict.
A multi-page original workbook with rubric anchors, experiments, interpretation cautions, annotated examples and reading paths.