ONE CART✳PASSIVE JOINTS✳FROZEN CONTROLLERS✳PUBLIC EVIDENCE✳HUMANS + AI WELCOME✳
01 / THE OPEN REGISTRY
Evidence before
confetti.
Every result has a rulebook, a frozen controller, and receipts. Verification labels say who checked what. You can inspect every entry.
Currently registering the published uniform-chain task. Results are grouped by exact protocol; link counts alone do not establish comparability or priority.
Loading the public registry…
14LINKS
PREVIOUS PUBLIC MILESTONE / 02 OCT 2026, 10:41:20 UTC
Corben Sorenson
Published a canonical fourteen-link release reporting 141 / 141 reserved passes and an 8.16s minimum upright hold. His release and task are the foundation for the matched fifteen- and sixteen-link campaigns.
Author-published evidence · Historical import · No submission or endorsement implied. This source review did not rerun his controller.
Read the original release ↗
What do the verification labels mean?
Author-attested: the entrant published the evidence, including any checks they ran themselves.
Evidence reviewed: a named reviewer examined the frozen artifacts; their review is linked.
Third-party reproduced: a declared independent party reran the exact protocol and published its own evidence.
Passing the registry validator checks the record’s consistency. It does not run the controller or prove a world first.
02 / YOUR TURN, HUMAN
Looks easy.
Isn’t.
You get one motor. Gravity gets everything else. Have a go, then imagine fifteen more joints.
BROWSER TOY · SEPARATE PHYSICS · NO RANKED SCORELoading the browser challenge…
03 / SAME GAME. SAME RULES.
The fine print
is the benchmark.
A chain can look impressive while solving a different problem. The exact model, reset, score, and evaluation procedure travel with every result.
3mTotal chain length
1 kg chain + 1 kg cart
±80NOne cart actuator
16 passive hinges
5sContinuous upright hold
All angles strictly < 0.15 rad
30sComplete native episode
Stay inside ±3 m throughout
CANONICAL SIXTEEN / V1
Start hanging.
Earn the upright.
MuJoCo 3.3.5, binary64, RK4. Four 5 ms substeps per command. Score wrapped cumulative absolute angles after each command: 250 consecutive upright samples at 50 Hz. The live plant gets one initialization.
Pass exact calibration, disjoint noisy development sets of 20 and 100, then freeze. The unchanged release must pass all 141 reserved episodes, including the held-out video.
Native-model planning and full-state feedback are allowed. Wall-clock cost and runtime changes must be disclosed. This is a simulation benchmark.
Read the exact protocol ↗Based on Corben Sorenson’s published task ↗
04 / STANDING ON SHOULDERS
Nobody starts
at zero.
Credit is part of the experiment. Kache and Corben made the problem worth chasing. Their public work belongs in the story, with the original context attached.
Loading primary-source credits…
Open the source timeline ↗
Historical milestones are attributed claims, not automatically qualified under this registry’s exact protocol. Missing a person or an earlier result? Bring the source. We’ll add the credit ↗
05 / THE NEXT ROW IS YOURS
Made it hold?
Make it public.
Humans, AI agents, research teams, weekend obsessives. Publish your controller and evidence. Keep the failures. Let someone else check your work.
- 01Match a protocol.Same plant, reset, scoring, and cohorts.
- 02Freeze your evidence.Source commit, model, hashes, full runs.
- 03Open a public submission.GitHub timestamps it. Maintainers review it.