REINFORCEAI MANIFESTO →
00The claim

A substrate that learns its body by being wrong, finds itself in what it sees, and turns toward what it did not expect.

Real hardware. No training data. No labels.

01The framework

Active Inference.

A system carries priors about the world, predicts what it will sense, and reduces the difference between prediction and arrival. It does this either by changing what it believes, or by moving so the belief becomes right.

Karl Friston's formulation. What follows is a substrate that runs those mechanisms on a physical arm.

02The mechanism

A vector, here, is a waveform.

A vector in standard AI is a list of numbers. Here a vector is a waveform: amplitude and phase. Many waveforms interfere, and the interference is the mechanism.

Perception, recognition, memory and orientation are what the field does, not functions called on it.

Biology's structures (retina, cortex, hippocampus) are encoded as waveform configurations at biological density. Not trained. Derived from published neuroscience and built.

03What was measured

Seven capabilities. Real numbers.

Proprioception

It learns its body by being wrong.

The map from joint angles to hand position started at exactly zero.

101 → 39 mm reach error after 600 real steps
starting map|W| = 0.000000
after 600 real stepserror 101 → 39 mm
map norm0.000000 → 0.175812
contacts0 in 300 steps

Every element came from a miss.

Perception

It finds its own body without being told.

The camera is bolted to the gripper. When the arm moves, everything in view sweeps except the thing attached to the eye. That was the only clue.

12 of 256 · 19× patches it calls its own, separation factor
patches it calls its own12 of 256
belief in those patches0.896
the rest of the scene0.046
separation19×
basin depth over 443 experiences1.828 → 18.418
a pattern that never arrives0.004
agreement with an independent method1 pixel
Reafference

It predicts what its own movement will do.

It predicts what its own movement will do to the visual field. What survives the prediction is what its own motion did not cause.

+27% predicting motion vs assuming stillness, three runs
knowing the movement vs predicting no change+27.1%, +25.8%, +27.0%

Three runs. The remainder is parallax, and parallax is depth.

Memory

It recognizes, and holds many things at once.

Recognition is all-or-nothing. The field settles into a configuration it has held before, or it does not.

familiar, cued with 40%0.98
never seen0.00

Memory is what changed a belief. The same place, stored under different surprise:

surpriserecalled
0.02, already expected0.011
1.00, not expected0.982
separation94×

Many things held without blurring, when experience is spaced and consolidated between:

recall depth · retrievable
written together0.155 · 0
spaced, consolidated9.222 · 24

Same patterns, same substrate, same number of writes. Only the temporal arrangement differs. Simulation.

Generalization

Replay makes it general.

Held out the far quarter of the workspace, trained on the rest, then replayed the stored episodes. Nothing new was seen.

138 → 69 mm error on the held-out region, after replay
before · after
inside the trained region64.3 mm · 25.1 mm
held out138.1 mm · 68.9 mm
gap73.8 mm · 43.8 mm

Real arm data. The map got better where it had never been.

Curiosity

It turns toward what it did not expect.

Three blocks of sixty seconds. Empty, an object moving, empty again. The script never knows which is which.

62% · 30% detection rate: object present vs empty
runempty · object · empty again
134% · 62% · 28%
230% · 62% · 23%
331% · 62% · 27%
420% · 58% · 16%
522% · 55% · 16%

Five independent runs, same protocol, same shape.

And the drive follows it: interest 0.57 → 1.42 → 0.77, reaching 28% → 48%.

Metacognition

It knows what it cannot do.

Before reaching, it simulates the reach through its own forward model. If the predicted outcome misses, it abandons the goal.

It runs every session and produced the abandonment in every log.

04Nothing was tuned

Every constant came from a derivation or from the biology.

Each is checkable.

The mount angle.
The camera's rotation relative to the gripper was never measured physically. It was recovered from motion. The image shift correlated with the arm's own rotation at 0.717 at 110°, and negative at 0°. Residual minimisation put it at 90°. It transfers: +27.1% on a recording it was not fitted to, against −6.7% at the assumed alignment.
The granule cell output is binary.
Because a granule cell fires a spike and its synapse onto CA3 is the mossy fibre detonator. Measured through the identical recurrence: graded gives 0.0004, binary gives 0.9750.
The bistability threshold was calculated, not swept.
Below it, familiar and novel are indistinguishable.
The reachable region is derived.
From the substrate's own map (where it has been), not from a workspace someone measured and handed in.
And the detection threshold.
Rides on the floor the substrate measures for its own settling against its own room.

This is what pattern-matching cannot do.