Maybe superintelligence isn't one giant model that knows everything. Maybe it's a structure that keeps getting broader and can show how it knows each thing. Here is the theory, and what exists today.
Film, 2 min 43 s. Every number in it comes from the tests described below.
The theory
Most AI today is one network with billions of weights. Everything it knows is spread across all of them, and teaching it something new means adjusting them all at once. That is why large models forget, blur facts together, and can't point to where an answer came from.
Oracle7 is built the other way round. Its knowledge lives in specialists: small, exact parts, each a formula, a table, or a list of checked facts. Each one sits in front of a frozen core that handles language and exact arithmetic. A specialist answers only inside its own lane and declines everything else.
One specialist alone is narrow. The idea is to connect them, and to make connecting them cheap.
Two hundred and fifty specialists in one byte
Every specialist gets a one-byte address. A byte has 256 values, so up to 250 specialists fit, with a few addresses kept for the core. A path through four specialists is four bytes. Routing a question costs a table lookup instead of asking every part.
256 addresses. 99 hold standing specialists. 0xFE is the core's exact arithmetic. Point at a cell to read its address.
standing specialist (99)
core math, 0xFE
open address (room to 250)
"A van gets 25 mpg. How many kg of CO2 does a 200-mile trip produce?"
0xFEcore math: 200 miles ÷ 25 mpg = 8 gallons
0x59carbon specialist: EPA figure, 8,887 g CO2 per gallon
0xFEcore math: 8 × 8,887 g = 71.1 kg
How a question gets answered
The network reads a question into what it gives (25 mpg, 200 miles) and what it asks (kilograms of CO2). Then it searches for a path: specialists for the domain facts, the core's exact arithmetic for the rest. Every hop is checked. The units have to line up, every number in the question has to be used, and the answer can't just repeat a given back.
The path above is fe 59 fe, three bytes. If no checked path exists, it declines. It doesn't guess. That one rule is where most of the reliability comes from.
A neural net, in a sense
Think of each specialist as a node, but an intelligent one, with a verified skill and a frozen core behind it. The paths between nodes play the role that weights play in a neural network: they are what the structure learns.
The difference is how learning happens. A path that works is stored as a single vector in the substrate's own algebra, HRR, where hops are bound to their positions. A new question recalls a stored path by resemblance, even from a partial cue. The nodes themselves never change. So the structure learns without backpropagation, and nothing it already knew can be overwritten.
Errors behave differently too. When an answer is wrong, the fault traces to a named node and gets fixed there. This week that caught an anatomy specialist giving human facts for a dog, and a finance specialist computing only half of a two-phase account. Both were narrowed, and both were re-tested on their own sealed questions to confirm nothing else moved.
What exists today
These are measured results, not projections.
Component
What it does
Measured
99 specialists
Each has its own manifest (what it accepts and declines) and passed its own sealed held-out test before it could stand.release index, 29 Sep
99 standing
Combination harness
Runs up to four specialists together, each isolated, and settles which one's answer depends on the question.isolation check across the whole family
705,375 calls 0 mismatches
Core calls specialists
When the core would otherwise have to decline, it asks the specialist family instead.fold w504, admitted by the gate, live
live
One-byte dispatch
A token-to-address table picks which specialists to ask, instead of asking all 99.7,125 census inputs
99.58% same result 3.7 ms vs 15.4 ms
Low interference
Specialists stay out of each other's lanes. Overlap is measured from their behaviour, target below 0.15.phase 0 baseline
0.135 all output 0.062 answers
Path memory
Paths stored as HRR vectors, recalled by partial cue.1,024-dim codebook
4-hop recovery 100%
The sealed tests
The questions for the first test were written and sealed before any network code existed. It ran once. There were 31 problems:
The network
28 right · 0 wrong
Best single specialist
6 right · 6 wrong
The core alone
5 right · 9 wrong
The network declined the other three. A second sealed test came out 27 out of 27. That included three questions outside what the network was built for, and it correctly declined all three.
Both sealed tests were written by the same AI system that built the network, before the code existed and disclosed as such. That is weaker than an outside test, so we ran one.
For the outside check we used real textbook physics: OpenStax College Physics and University Physics, exercises from held-out chapters, scored against the book's own answers. The network answered 2 of 564. Both were right, none were wrong, and it declined the rest.
What it isn't, yet
This is not superintelligence, and we won't call it that. Two of 564 is a sliver of coverage. The hard part is reading open questions, where the numbers and the goal are buried in ordinary language. Open conversation is further still.
The network also isn't inside the public core yet. The gate rejected the fold that puts it there for three extra false answers. Our own re-run finds the same count with and without the network, so we are tracking down which measurement is off. Until that is settled, the fold stays out. The combination harness and the core's ability to call specialists are already live.
Why the shape matters
The property that matters is here already. This is an intelligence that grows by adding verified parts and remembering verified paths, never by retraining and never at the cost of what it knew. Each new specialist widens what the whole structure can reach. Each new path makes the next similar question cheaper. Every answer can be traced back through the bytes that produced it.
The next steps follow directly. We fill the address space toward 250, teach the reader to pull givens and goals out of open questions, and let conversation run on the same paths. If superintelligence arrives, we think it will look less like one oracle and more like this: a map that keeps getting larger, where every road on it has been checked.
Numbers: sealed network tests (sealed 42fc50f7 and 3179c28c), 30 September 2026. Independent check: OpenStax College Physics 2e and University Physics Vol. 1, held-out modules. Film narration voiced with ElevenLabs. Prometheus7 Research Institute.