Prism Arena
Each contender emits one SceneSpec JSON object validated before rendering and evaluation.
- My role
- Creator and lead author
- Status
- Research

11models
10tasks
2attempts each
220runs
203valid scenes
17failed validation
The same test for everyone
Every contender receives the byte-identical task prompt through the identical command, with tools, skills, and rules disabled as a fairness control.
A label-blinded four-judge panel scores the first valid recorded run for each model and task pair, and the results report per-run validity next to the more flattering best-of-two figure.
The task, Bloom Sequence
Open twelve petals as one ordered, keyframed wave.
Ten tasks, each one checkable by eye
- Observatory ArrayPlace a calibrated observatory.
- Signal CascadeLand a five-pillar score at exact times.
- Retrofit Under RulesAdd an instrument without disturbing a station.
- Kepler RigProve counter-rotating shells and plotted orbits.
- Crank and SliderKeep samples, positions and a trace consistent.
- Standing WaveEncode three wave bases.
- Phase ChangeCarry five specimens through a phase cycle.
- Jointed ArmResolve a five-link target.
- Bloom SequenceOpen twelve petals as one wave.
- Coupled PendulumsChoose pass counts that all close together.
Who did what
What I owned
Conception and lead authorship
Lalo conceived and built Prism Arena and owns essentially all of it: the tasks and tests, the SceneSpec contracts, and the harness (owner-attested).
What the team owned
Collaborators
The remaining work came from Agentic Engineering collaborators, credited collectively: no individual is named and no ownership percentage is published (owner decision, 2026-08-20).
Built with
TypeScript, React Three Fiber, Zod, Anime.js, SQLite
Receipts
| Source | Kind | Verified |
|---|---|---|
| Prism Arena benchmark README | benchmark output | 2026-08-13 |
| Owner attribution and publication-rights decision | Source not public | 2026-08-20 |
Can’t trust your agent yet?