Refusal
NULL-brake
When measured progress backslides more than one wheel radius — the body's own scale — the controller refuses: relax, let momentum bleed, re-center. The eye of the storm, made mechanical.
Open experiments
the ledgerEach experiment ships as its own frozen fork: the code, a one-command runner, the raw results, and a write-up that keeps the negative findings. If a number is wrong, the fastest way to show it is to run it.
Wheel-Amp v0.1 — soft-body locomotion, made robust
Demonstrated · in-simulationA soft-body wheel — a hub, a ring of rim points, and spoke muscles — learning to roll to a goal across differentiable physics (AlgoVivo). The muscles are driven by a persistent UDI controller that commits, refuses, or returns NULL. The question: can the adapter be made robust to the wheel's starting orientation, without tuned gains and without touching the protected internals?
16/16
start rotations reached
UDI, amplified (v0.4.0)
15/16
before amplification
one backward runaway at 90°
0
regressions
identical on all 15 working cases
| Controller | Rotations reached (of 16) | Note |
|---|---|---|
| Open-loop math gait | 11 / 16 | fixed sinusoidal wave; fails 5 starts |
| UDI controller (frozen) | 15 / 16 | fails 90° — rolls backward, never recovers |
| UDI controller (amplified) | 16 / 16 | 90° fixed; every other case unchanged |
What the amplification is. Two structural changes, no tuned gains added to the commit/refuse decision:
Measured, not learned
Re-seed to polarity
The failure's true cause was a spurious early confirm corrupting the learned direction. On refusal, trust the measured body polarity over the corrupted learning.
Falsification kept. An earlier version of the runner set stiffness after the settle step; under that ordering the 90° failure vanished and the fix showed no benefit. The failure is initial-condition sensitive. The reported result uses the host-faithful ordering (stiffness before settle), and it is recorded in the write-up so the sensitivity is not hidden.
Not a physical-robot result — in-simulation on one 10-spoke wheel, one course. Not a claim of general locomotion or intelligence. Wider forks — spoke count, friction, terrain — are the next screen.
# reproduce (pure Node, no browser)
cd experiment-wheel-amp-v0.1
node sweep.mjs # regenerates results.json
sha256sum -c FREEZE-SHA256SUMS.txt
cd experiment-wheel-amp-v0.1
node sweep.mjs # regenerates results.json
sha256sum -c FREEZE-SHA256SUMS.txt
Amplified controller · sha256
addf25d7f882eb7f163cb266a1076da5cf17ebf8fa855696ea7ad13df57e1e2e
Sweep runner · sha256
d3b0ef9d7cb0848cc8972e60dbbb15cd509d4c99ca7003a8728f2c0c303d6e10
Results · sha256
5a4690ac064119b4a120380ac10ec46acb4369e1317a9c362cebb8f5fd2363f0
Write-up · sha256
3318cd55a37ce890fd0bc8e831d176b18255cf1c966151a0ca7c296b878a6785
PreregisteredFrozen forkHost untouchedNegative result keptNo tuned gains
Also on the record
elsewhere in the fleetBenchmark
LongMemEval-V2
Self-reported result with verifiable hashes, published on our terms. results.ado.earth →
Evidence
Sealed releases 001–005
Outcome-only releases with sealed-source commitments. evidence.ado.earth →
More experiments in progress. Wider wheel forks (spoke count, friction, terrain), the memory-benchmark gauntlet, and the TinyHive substrate integration land here as they freeze — each with its runner, data, and hashes.
Come falsify us.
Every experiment here ships the code and the command that regenerates it. If a result is wrong, reproduce it and we will publish what you find — positive or negative. That is the point.