Adaptive junctions.
Study a junction whose response depends on its input and a retained local state, with feedback that changes selected connections. The material mechanism and its advantage require physical measurements.
Research notebook / October 28, 2026
Make the idea testable.
Let the evidence lead.
Today's work brings the hardware architecture, owner-controlled intelligence, collective-learning network, and production research into one pathway.
Research directions
A continuous computing fabric.
Explore local state, communication, energy, adaptation, and maintenance throughout a three-dimensional structure. Begin with compatible, characterized elements before tissue or body integration.
Learning across machines.
Test whether validated sharing improves held-out performance without spreading errors, losing previous capability, or exposing records without permission.
Recorded reference result
Software model · Synthetic taskCorrection, persistence,
Correction, persistence,
and recovery.
The reference sequence connects supplied knowledge → prediction → observation → correction → unseen cases → restart → lost-reply recovery.
| Check | Recorded result |
|---|---|
| Training observations | 3,200 |
| Held-out evaluation | 4,000 cases across 10 seeds and four supplied contexts |
| Uncalibrated equation error | 1.198832 mean absolute error, synthetic watts |
| Cell bounded-update error | 0.042687 mean absolute error, synthetic watts |
| Ordinary per-context estimator error | 0.039326 mean absolute error, synthetic watts |
| Fresh-process restart | Predictions retained |
| Lost acknowledgement | Reconciled by reading the receipt, with one receiver execution |
| Existing regression checks | 15 passed in the recorded run |
The Cell correction reduced error by 96.44% against the uncalibrated starting equation. The ordinary estimator was slightly more accurate.
The useful evidence is the integrated correction, persistence, and recovery behavior on this bounded fixture. It does not establish a superior learning method, autonomous knowledge organization, a physical Cell, or a material advantage.
Source: Humanus reference demonstration, RESULTS.md and recorded results.json, reviewed October 28, 2026. This website update did not rerun the experiment. An independent reproduction remains a next milestone.
Evidence boundaries
Ambition and proof,
both in view.
Executable architecture
Software and RTL work provide bounded evidence about state, execution, and recovery. Simulation does not prove physical Wi-Fi, power, timing, or manufactured reliability.
Physical development
The complete checked circuit, board, measured adaptive devices, and full integrated machine remain development work. Concept drawings are not manufacturing releases.
Commercial development
Engineering teams are a proposed early audience. Customer demand, unit economics, production yield, and a scalable service model have not yet been validated.
Bring a method.
Challenge a result.
We welcome collaborators who can help define fair comparisons, characterize devices, reproduce the reference behavior, and turn the next claim into an experiment.