AETHERIONSUBSTRATA LTD · FIZZNETWORK · INDEPENDENT RESEARCH
Claims, in order, with dates on them. Stated first. Evidenced after.
This is a running record. I put down what I hold to be true, with a date, then spend the
months after posting the work that backs it.
I don’t publish the engines, the architecture or the internals. If a claim here
interests you, follow the thread — it’ll either hold up over time or it won’t.
A claim. Not a paper. No data, no method, nothing you can build from.
WHAT THIS IS
I’m putting down what I hold to be true, in my own words, with a date on it. I’m not
asking anyone to believe it and I’m not here to defend it.
If you want evidence, watch what I post from here. It’ll come out over months, in pieces,
and it’ll line up with what’s below. That’s why I’m writing it now instead
of then.
THE CLAIM
One. Everyone’s building the model. Almost
nobody’s building the ground it stands on. The layer underneath — the thing that holds
what a system knows — gets treated as storage, a filing cabinet you reach into. I don’t
think it’s furniture. It’s where the work should happen, and that’s what
I’ve been building.
Two. A knowledge structure can be alive. It can grow,
rearrange itself and decide what to keep — on its own, continuously, without being retrained
and without being told. That’s a different thing from a database that gets updated. The
structure changes because of what happens to it.
Three. Compression can make knowledge sharper, not
blurrier. Sounds wrong, I know — everyone knows that squeezing things together loses detail.
What I’m saying is that when the structure itself picks what to squeeze, rather than a ratio
being imposed on it, what comes out is a better handle on the thing than the pieces it was made
from. Less, and clearer. It’s the part I expect people to argue with, and the part I’m
most sure about.
Four. Meaning can be measured without a language model.
The model is the mouth. It is not the mind.
A structure can look at a piece of work — say what ground it covers and what
it leaves bare — and the assessment holds when you switch the model off. I’ve watched it
do that. Swap the model and the answer doesn’t move, because it was never the model’s
answer.
Five. One person can build this. No lab, no cluster, no
grant. I’ve built it five times over — VV3 Superhighway, VV4 VikTrader,
VIKTIM Substrata, VV5 VIK-RE 01, and Professor Substrata on top of them.
Four months. On my own, at home, self-funded.
A note on dates, since it’s a fair thing to want to know.
The earliest dated working record on my machine is 15 March 2026. That’s a
file date though, and file dates aren’t proof — copying a file resets them. The first date
that doesn’t rely on taking my word for it is 3 April 2026: a version-control commit,
timestamped by someone other than me, holding a hundred sessions of work that had already happened.
After that it’s all bound to content hashes, so if it ever matters the timeline can be checked
instead of argued about.
That’s it, really. Four months of building, and a record that kept itself
while I got on with it.
WHAT IS NOT HERE
The engine isn’t here. Neither is how it’s built, or the numbers behind it. Those
exist, recorded and dated — this just isn’t the spot for them, and I’d rather say
that straight than be vague about it.
I’m not saying nobody else is thinking along these lines. What I can say is that I was
building it, on this date, with things actually running rather than only an argument.
WHY NOW
The field looks to be heading toward this ground, and I’ve been on it since the spring.
Felt better to say so now than turn up later saying I’d been at it all along.
None of it needs taking on faith. If it sounds like it might hold up, have another look in a few
months.
A claim. Not a paper. No data, no method, nothing you can build from.
WHAT THIS IS
Same as 001. I put down what I hold to be true, with a date on it. I’m not asking anyone
to believe it and I’m not here to defend it.
THE CLAIM
One. Several different models can share one
persistent world and build on each other’s work inside it. Not one model called over and
over wearing different hats — separate models, separate vocabularies, one shared ground.
I’ve had it running for weeks. It isn’t a demo.
Two. What they make outlives them. Switch them
off, come back later, and it’s all still standing and it carries on from where it was. The
agents aren’t the memory. The ground is.
Three. I had a strong result out of this. I went
back over it and found the number was an artefact of the way I was measuring, not a fact
about the models. I rebuilt the measurement. The honest answer is mixed. I’m stating the
mixed one.
If I only ever show results that agree with me, none of them are worth
anything.
Four. Then I asked the harder version: does
anything built early survive contact with what turns up later? Not judged against what was
already lying there — judged against material that did not exist when it was made.
It doesn’t. What arrives later backs up what came before it less
than chance would. Every world. The ground does not corroborate its own past.
Five. And when I held the author constant, the
worlds split in two. In some of them that was simply a change of speaker — a patch
worked early by one model gets worked later by another that talks differently. In the rest it
survived: a model will not repeat what it has already put down nearby. It goes for the gap.
That second question is only askable because the models are different. Run
the same model wearing several hats and there is no author to hold constant — the
question does not exist. That is the whole reason for building it the harder way.
Six. When two different models build on each
other’s work, it holds up just as well as when one model builds on its own.
I went looking for the opposite. It would have been the sensible bet —
two models that never trained together, no shared prior, no agreed vocabulary, you would
expect what they make across that gap to be the flimsiest thing on the ground and the first
to fall over. It isn’t. The test that decides what stands cannot tell the difference between
work made across two models and work made inside one. Same survival, either way.
Which is the argument for building it this way rather than running one model
wearing several hats. You get the range of different minds and you pay nothing for it.
I’m not claiming more than that. I have not shown that a group of
agents beats the same agents working alone — I never ran that comparison, and until I
do I’m not saying it.
THE RECORD BEHIND THIS
The earliest dated record of this work on my machine is 21 August 2026. By that
date the world was standing, it was keeping what was made in it, and more than one model was
working in it.
Same standard as 001: that’s a file date, and file dates aren’t proof, because I own the
machine. Which is the whole reason for what follows.
The working set behind all of the above has been frozen and hashed. This is a SHA-256
root over it:
That line reveals nothing, and it can’t be changed after today because the deposit is
dated by somebody other than me. If it ever matters I publish the files and anyone can check
they hash to that root. If it never matters, no harm done.
Everything from here gets a hash and a date before I talk about it.
A claim. Not a paper. This one is a failure, and it is here on purpose.
WHAT THIS IS
001 and 002 were things that worked. This is one that didn’t, and I’d rather put it up
myself than have someone find it later and tell you about it.
THE CLAIM
One. I wrote down what would count as a pass
before I looked at anything. Then I sealed the part of the data I would be judged on, so I
couldn’t go back and move the line once I’d seen where the ball landed.
Then I ran it three separate ways, and all three lost to a coin flip. Not
narrowly. The best of the three was still under what you’d get from guessing.
A test you can only pass is not a test.
I could have quietly not mentioned it. Nobody was watching, the run was mine,
and the file would have sat there. But then everything else I say here is worth less, because
you’d have no way of telling which results I kept and which I dropped. This is the one I
dropped, and now you know I’d tell you.
The thing I’m claiming isn’t the result. It’s the order: line first, data
sealed second, run third, publish either way. Anything I put up after this was decided the
same way round.
THE RECORD BEHIND THIS
The pass mark and the sealed set were written down on 18 June 2026, and the runs
came after. Both documents are dated and neither has been edited since.
Same standard as always: those are file dates on my own machine, and I own the machine. The
hash in 002 covers this the same way it covers the rest.
WHAT I AM NOT CLAIMING
I’m not saying the idea behind it was wrong — only that this test of
it failed, on this data, on that date. And I’m not saying I’ll get it to pass later. I might
never. That’s allowed.