My Peer-Reviewed T. rex Looks Worse Than a 1996 Video Game

I wanted to know what a Tyrannosaurus rex actually looked like. Not the movie one. The real one, the one the papers describe. So I told AI agents to read the science and build it in 3D, model and animation, with no human touching the mesh.

Three days later I had a potato.

That is the apex predator. Beige, calm, faintly disappointed. It has the body language of a man waiting for a bus.

And here is the annoying part: it is scientifically correct. Every bone length traces to a published paper. The body mass sits inside the window from peer-reviewed volumetric studies. The walk cycle matches a Royal Society paper to three decimals and a script verified it. My dinosaur can cite its sources. It just cannot frighten a chicken.

I am oddly proud of it and I cannot show it to anyone without explaining it first, which is usually a sign you have built the wrong thing.

What the agents taught me about T. rex

I never opened Blender. My job was approving phases and stopping Windows from falling asleep mid-render. The agents did the rest: two Claude Code sessions, one per machine, plus OpenAI's Codex hired specifically to be unpleasant about the plan.

The research was the part I enjoyed, because the agents kept telling me I was wrong about a dinosaur I had been confident about since I was eight:

• It had lips. Teeth covered when the mouth is closed. More monitor lizard, less crocodile grin. Sorry, Jurassic Park.
• It was scaly, pebbly scales about a millimetre across. On a 12 metre animal that is a shader detail. Nobody sculpts a millimetre.
• It was slow. Preferred walking speed 4.6 km/h, set by the resonant frequency of its tail. Top speed maybe 25 km/h, always with one foot down, because an aerial phase would have snapped its legs. The jeep was never in danger.
• The tail was the engine. The biggest leg muscle lives in the tail, roughly 260 kg per side. Every skinny-tailed rex you have seen is wrong.
• The colour is unknowable. No pigment data exists for any tyrannosaurid. Ours is olive-brown because somebody had to choose, and the repo says so in writing.

Then Codex, the hired pessimist, earned its fee in an afternoon. My research agent had cheerfully declared the Smithsonian's T. rex scan public domain. Codex opened the actual page: only the metadata is CC0, the scan is non-commercial only. A hallucinated licence, caught by a rival model that bothered to click the link.

Codex also caught the mass formula counting the air cavities twice. My dinosaur was being deflated twice for the same lungs.

Building a dinosaur when nobody involved can sculpt

Neither I nor the agents can sculpt. So the pipeline became Python scripts driving headless Blender, where nothing is judged by eye and everything is judged by a number.

Bones first, from published measurements of "Sue". Then 17 muscle groups, each named after the real muscle, each traceable to an anatomy paper. The script prints every muscle's weight, which is how I learned my dinosaur's tail muscle had come out at 471 kg per side against a published estimate of 261. It had been lifting. Tuned down to 252 kg.

Then the skin, with one rule: mesh volume times published tissue density must land between 8.4 and 9.5 tonnes. First attempt, 11.4 to 13.1 tonnes. My scientifically rigorous T. rex was 25 % overweight and its belly hung 7 cm below where the ribs say a belly stops. Two rounds of dieting later it passed at 8.8 to 10.2 tonnes.

All of that worked, and I want to be fair to it. Build, measure, compare against a paper, fix. Dozens of times, overnight, with nobody watching and no human skill involved. Nothing about the science half embarrasses me. I would put those numbers in front of a paleontologist.

The moment I should have seen coming

The walk had a validator: no aerial phase, no foot sliding, no tail dragging, jaw inside its measured gape. It rejected the animation four times and every rejection was a real bug. Right-foot plants generated at the left foot's coordinates, so the poor thing dragged a foot two metres per step. A clamping bug at the clip boundary. Feet planted too close to the hip, shoving the femur past its allowed arc.

Then the validator said PASS. I opened the render, and the dinosaur had stick legs and floating feet. The rig was rotating each leg 90 degrees, so the wide thigh faced the camera edge-on. Every number was perfect. The picture looked like the animal had been assembled by somebody reading the instructions upside down.

Validation catches what you thought to measure. I had measured whether the foot reached the ground. I had not measured whether the leg was on straight.

Here is the finished walk, machine-validated, zero violations, at the exact speed the paper says the animal preferred:

Two full strides at 1.293 m/s. Every number in it is right and it walks like a metronome with a mortgage. MP4 in 1080p.

I showed it to my wife. She asked what it was supposed to be doing.

So why did I get a potato?

I assumed it was my fault: bad prompts, wrong tools, not enough iterations. So the agents ran one more research pass, this time on themselves. The potato was predictable, and the literature is almost rude about how predictable:

• The models cannot see properly. "Vision Language Models Are Blind" tested frontier models on things a child does instantly, like whether two circles overlap: 58.57 % average. Probes suggest the vision half knows and the language half cannot read it out loud.
• Their taste is not our taste. VisJudge-Bench puts model-to-human agreement on aesthetics at 0.18 to 0.41, and MLLM-as-a-Judge found they are fine at "which of these two is better" and unreliable at grading one thing alone, which is exactly what an agent does to its own render. Mine could tell broken legs from normal legs. It could not tell mediocre from good.
• There is a benchmark for agents driving Blender, and it is humbling. BlenderGym scored thirteen vision-language systems 2 to 10 times worse than human Blender users, on tasks the authors call easy for people.
• Humans are freakishly good at spotting bad motion. Twelve moving dots are enough for you to recognise a walking person. Worse, uncanny-valley research found degraded motion drops a character below its own still image. My walk cycle is not neutral. It is actively working against me.

Then the structural reason, which is the part I keep chewing on. What made these models good at maths and code is training against verifiable rewards. "Femur = 1.321 m" is verifiable. "Feels alive" is not. Correctness automates first by construction, not by accident.

I hired a correctness engine and handed it a beauty problem. It solved the half it was built for and gave me a vegetable.

The 1996 problem

The T. rex in Tomb Raider, a few hundred polygons and a hand-painted texture, is scarier than my 52,000-quad specimen. Artists in 1996 had no science and no polygons and made a monster. My agents had all the science and all the polygons and made an exhibit.

Which is, if I squint, the actual finding. An artist starting from this repo would skip weeks of paleontology and inherit guardrails against movie-monster anatomy. My agents would get the one judge they lack. Neither half is worthless. They are two different jobs, and I spent three days proving it with a dinosaur.

The repo stays private and non-commercial, because the museum scan I used as a shape reference allows personal use only. Nobody is fighting me for the rights to a beige potato anyway.

What I would tell you

Make the agent's claims falsifiable with numbers from the literature. Hire one hostile reviewer from a different vendor. Look at the renders, because four green validator runs hid a dinosaur with sideways legs. Write down what you do not know.

And expect the last 10 % to be 90 % craft. Correctness automated beautifully. Beauty did not, and now I have citations explaining why. My dinosaur is anatomically defensible, machine-validated, and about as frightening as livestock.