Three Judges,
Thirty Filters

Imagine you are the third judge at an art exhibition. The first two judges have already voted. You pick up the first dossier — a filter called Terminal Matrix, which renders any photograph as glowing green phosphor characters on a jet-black screen, like a vintage computer terminal.

Judge One — methodical, holistic — awarded it 79 out of 100. Judge Two — a strict statistician who measured actual pixel data — gave it a perfect score: 10 out of 10. The highest mark awarded to anything in the entire exhibition.

You study the output images carefully. You appreciate the atmospheric CRT glow, the phosphor-green portrait. But you also notice something the others missed: when this filter is applied to a bright sales chart, the light background turns the whole image into a sickly greenish wash instead of the crisp dark screen the concept demands. The filter has a blind spot. It only works in the dark.

You write: 6.6 out of 10. Rank: 22nd of 30.

That gap — from a perfect score to the bottom third — is not a grading error. It is the entire story. Three artificial intelligences looked at the same thirty pieces of AI-generated art and came to dramatically different conclusions. Understanding why tells us something fundamental about how machine intelligence defines quality, and what it means to judge something good.

Act I

The Factory and Its Products

The story begins in early 2026, with a question about creative discovery. Could an AI not just follow instructions, but genuinely explore — proposing new aesthetic ideas, executing them, evaluating the results, iterating? The ambition was to demonstrate that LLMs could function as autonomous creative agents.

Claude Code (Sonnet 4.6) was given a simple but precise brief:

The model ran. And ran. Over multiple sessions, it generated 30 distinct image filters — each with an evocative name, a README explaining the concept, and a shell script containing the actual code. Each was applied to three standard input images:

Three inputs chosen because they are different: a photograph with continuous tone; a comic panel with inked outlines; a data chart with precise geometry and small text. A filter that works on all three is truly robust. One that only works in the dark has revealed its own limits.

Now came the question: how do you evaluate AI-generated art? Three language models were each given the same task — think like a world-class artist and critic, score rigorously — and each drew from a completely different well.

Act II

Three Very Different Judges Take the Stand

The evaluation prompts are documented in 📄 prompts.md and the full results in 📄 gpt-evaluation.md, 📄 sonnet-evaluation.md, and 📄 gemini-evaluation.md. Each judge received the same framing but applied it through a completely different lens.

GPT asked: Is it visually impressive? Claude asked: Is it true to its medium? Gemini asked: Do the numbers check out? These are not the same question — and the answers diverged accordingly.

The Details

When the Judges Couldn't Agree

Terminal Matrix

Maximum Controversy

📄 filter.md 💻 filter.sh

GPT: 7.9 · #1 Gemini: 10.0 · #1 Claude: 6.6 · #22

Renders images as ASCII characters on a phosphor-green CRT terminal.

"Perfect Execution. This is a masterclass in constraint. By forcing the image into a specific, narrow color space — phosphor green — it creates a cohesive and instantly recognizable look."

— Gemini · Avg Color: [13, 201, 56] (Vivid Green) · 📄 source

"When the filter works — on dark inputs — it's exceptional. When it fails — on light inputs — the failure is complete. The chart has a light background; the filter renders it as a green-tinted wash."

— Claude Sonnet · Cross-Input Versatility: 5.0/10 · 📄 source

Thermal Scan

Expert vs. Analyst

📄 filter.md 💻 filter.sh

GPT: 7.8 · #2 Claude: 9.0 · #1 Gemini: 4.0 · (bottom half)

"The most technically accomplished and consistently excellent filter in the gallery. Authentic, coherent, versatile, and beautiful. A clear gold standard."

— Claude Sonnet · Overall: 9.0/10 · 📄 source

"The average color is grey. A thermal scan should be vibrant — Blue/Red/Yellow. This looks desaturated and incorrect."

— Gemini · Avg Color: [132, 122, 119] · 📄 source

Gemini averaged across all three outputs. The chart and comic (darker inputs) pull the average into grey territory. Claude was looking at the portrait in isolation and seeing a masterpiece. Both are right. They measured different things.

CMYK Separation

Process vs. Appearance

📄 filter.md 💻 filter.sh

GPT: 7.2 · #21 Claude: 8.6 · #3 Gemini: 3.0 · (lowest tier)

"This is a filter about how printing works, not just what printing looks like. The screen angles are technically correct — the single most impressive act of medium fidelity in the gallery."

— Claude Sonnet · Conceptual Coherence: 9.0/10 · 📄 source

"'Separation' implies seeing the dots/layers. The high color count suggests it's just a blurry photo."

— Gemini · Unique Colors: ~8000 · 📄 source

Eight thousand unique colors in the output means the pixels weren't quantized — no flat CMYK blocks. But this is precisely correct: real offset printing uses continuous halftone dots at sub-pixel scale; zoomed out, they blend optically, producing thousands of apparent hues. Gemini's color-count heuristic confused accuracy for failure.

Where They Agreed

The Unanimous Winner

One filter achieved something remarkable: it satisfied the holistic art critic, the medium fidelity perfectionist, and the pixel statistician simultaneously. Its name was Neon Blueprint.

Why consensus? Its average pixel: [6, 23, 39] — deep navy, exactly what an architectural blueprint looks like. High pixel variance because glowing cyan edges create dramatic contrast against the dark background. Gemini's statistics confirmed it. Claude's art-historical knowledge praised it. GPT's holistic eye admired it.

The code behind it is elegantly simple — a recipe for light from darkness:

# Neon Blueprint — core logic (Python/NumPy) · filters/neon-blueprint.sh
# Step 1: Build a dark navy canvas
bg = np.zeros((h, w, 3), dtype=np.float32)
bg[:, :, 2] = 18   # faint blue base

# Step 2: Draw subtle grid (blueprint paper)
bg[::48, :, 1] += 8;  bg[::48, :, 2] += 22

# Step 3: Extract edges, add layered glow
edges   = sharp.filter(ImageFilter.FIND_EDGES)
glow_sm = edges.filter(ImageFilter.GaussianBlur(3))
glow_lg = edges.filter(ImageFilter.GaussianBlur(8))

# Layer: large glow (blue), medium glow (cyan), sharp edges (white-cyan)
bg[:, :, 1] += gl * 100;  bg[:, :, 2] += gl * 160  # large glow
bg[:, :, 1] += gs * 200;  bg[:, :, 2] += gs * 230  # cyan glow
bg[:, :, 1] += e  * 255;  bg[:, :, 2] += e  * 255  # sharp edges

The filter works because it commits to a fundamental premise: everything begins in darkness. Every feature of the input image is reborn as light — luminous edges, glowing grid lines, cyan contours. Pure contrast. That is why it works on photographs and comics and bar charts alike.

The second filter to approach consensus was Prussian Cyanotype — Gemini 9.0 ("A beautiful example of historical emulation"), GPT 7.5, Claude 8.1. Both top consensus filters are monochromatic systems: one color dominating a dark or pale field. Simple enough for the pixel statistician. Precise enough for the medium critic. Bold enough for the generalist judge.

Conclusions

What Three AIs Disagreeing Tells Us

There is a detective's principle that applies here: when witnesses to the same event give radically different accounts, the investigator should ask — what were they each positioned to see?

GPT was positioned as a holistic art judge. Its compressed scoring range (0.62–0.79) reflects genuine professional conservatism: an experienced judge doesn't give 90s to ordinary good work, and doesn't give 40s unless something has truly failed. Everything here was competent.

Claude was positioned as a medium-fidelity critic. It brought art-historical knowledge — the specific chemical composition of Prussian blue, the traditional screen angles of CMYK halftone, the tonal response curve of silver daguerreotype. When a filter claimed to simulate a historical technique, it checked whether the simulation was accurate. This produced both the gallery's harshest reviews (Cathedral Glass: 4.0 — "the filter does not simulate stained glass") and its most generous (Thermal Scan: 9.0 — "the gold standard").

Gemini was positioned as a pixel analyst. It measured color statistics. This approach has merit for certain filter types — Terminal Matrix passes with flying colors because its average pixel truly is vivid green. But it systematically misjudges filters that simulate physical processes by layering color information rather than replacing it. Thermal scan averages to grey because it uses a continuous color ramp — and continuous ramps average to neutral across many inputs.

What the disagreements reveal, above all, is that creative quality is not a single number. It is a vector: concept, execution, versatility, fidelity, impact. When you want to know whether something is good, the right answer is not to pick one judge. The right answer is to ask all three — and let their disagreements teach you something.