RESEARCHApache-2.0 · open sourceReports a negative result about itself

Can this gate fail?

A small research substrate for online local learning in spiking neural networks, with two measurement instruments that read its own activations. It publishes one positive engineering result about training memory and one negative science result about integration, with equal weight. Pure Python and NumPy. The test suite runs on a laptop in a couple of minutes.

Read the substrate Research conversations

16×

sequence length, memory unchanged

results published, one positive, one negative

GPUs, services or network calls

consciousness claims

What it is

A reference implementation, released with the results that came out of it.

Backpropagation through time stores activations for every timestep, so training memory scales with sequence length. An eligibility-trace method replaces the backward unroll with forward traces, so a training step holds only fixed-size state. That is a well-known claim, and it is usually asserted rather than measured.

This repository measures it, two independent ways that have to agree: a deterministic byte count of what the training step holds, and a real peak-allocation measurement of the same run. It is not a framework, not a benchmark suite, and not a competitor to any sequence-model library.

The second half of the repository is two instruments that read the substrate’s own activation timeseries. They were built to look for a signal, and they report that they did not find one.

Two results, equal weight

One worked. One did not. Both are published.

Training memory that does not grow with sequence length

A sixteen-fold increase in sequence length leaves the eligibility-trace method’s training memory unchanged and multiplies backpropagation-through-time’s by sixteen. Measured deterministically and confirmed against real peak allocations.

BoundaryThis is a memory claim only. It is not a claim that the method matches backpropagation through time on accuracy, wall-clock, or any downstream task. No accuracy comparison is made, and none should be inferred.

No non-confoundable information integration at this scale

The integration and global-workspace probes separate a genuinely integrated control from degenerate ones, and they score this substrate as showing no non-confoundable integration. Its measure is real but not distinguishable from a noise control. That result is reported as it stands.

BoundaryAn earlier framing of this work reported a weak positive signal. That reading did not survive a gate requiring substrate-derived evidence, and it is retracted in the repository’s own claims document.

The instruments report negatively on their own substrate

The probes were built by the same people who built the thing being measured, which is exactly the situation where a result gets quietly shaded. Publishing the null, and the retraction of the earlier positive reading, is the only way that instrument means anything.

BoundaryA null at this scale is a statement about this substrate at this size, not a general result about the architecture class.

You can check both on a laptop

Pure Python and NumPy, Apache-2.0, no GPU, no service, no network calls, nothing phones home. The test suite runs in a couple of minutes, so both results are reproducible by anyone who wants to disagree with them.

BoundaryThe memory figures come from a machine with no CUDA device, so peak process memory is the honest stand-in for peak device memory. That substitution is stated in the code that makes the measurement.

Claim discipline

What it refuses to claim.

No claim is made, in the repository or by its authors on the basis of it, that this system is conscious, sentient, or a step toward either. The measures used here correlate with consciousness in systems already known to be conscious; measuring them in a small artificial substrate says nothing about its inner life, and here they read negative regardless.

The instruments are released as measurement code. Any consciousness framing placed on them belongs to whoever adds it, not to this work.

It is not a product and has no roadmap attached.
It is a reference implementation released for inspection and reuse, together with the results that came out of it.
The memory result says nothing about accuracy.
No accuracy comparison is made anywhere in the repository, deliberately, so that the memory claim cannot be read as a quality claim.
Part of the work is deliberately withheld.
A behavioural tier was held back because its probes construct a fresh subject rather than reading the substrate under test, which would not have measured what it claimed to measure.

Published in full: CLAIMS.md

What it isOnline local learning in spiking neural networks, plus two activation-based instruments

ImplementationPure Python and NumPy. No GPU, no service, no network calls

Positive resultTraining memory unchanged as sequence length grows sixteen-fold

Negative resultNo non-confoundable information integration at this scale

RetractedAn earlier weak-positive integration reading, withdrawn in CLAIMS.md

LicenceApache-2.0

Repositorygithub.com/BlackUnicornSecurity/morpheus-substrate

Next step

Two ways in.

Reproduce both resultsApache-2.0, pure Python. The suite runs on a laptop in a couple of minutes.

Research conversationsIf you work on local-learning or spiking systems, we are interested.

Morpheus is a research program that publishes its negative results. The substrate is open source, if you work on local-learning or spiking systems, we are open to research conversations.