It can say why
When 7 works something out, the steps that produced it are the actual derivation, not a plausible story told afterwards. Ask it why and it walks the path it took.
Alpha · v0.1
A machine that knows what it knows.
7seven understands language by building structure, not by predicting text. It composes answers it was never given, keeps what it worked out separate from what it was told, and returns nothing at all where it has no grounds. Taught by conversation, corrected in the moment, inspectable line by line.
Version 0.1 is an alpha. The architecture is complete and instrumented, and every figure on this page comes from held-out tests. What 7 knows today is small — a few hundred concepts, taught by hand. The case below is for why that is the interesting part rather than the disappointing one.
The case
Why build this when language models already exist
There is a reasonable objection to 7seven. Language models work, they are backed by more capital than any software category in history, and they get better every quarter. Building something to stand beside them looks like a poor use of money.
The objection assumes the two systems are competing at the same thing. They are not — and the gap between them is architectural, not a matter of effort or funding.
A language model produces the most likely continuation of a text. Everything it does brilliantly follows from that, and so does everything it does badly. It cannot tell you why it said something: a chain of thought is more generated text, not a record of the computation that produced the answer. It cannot reliably decline, because answering and declining are the same operation on different probabilities. It cannot be corrected, only retrained, patched with retrieval, or asked to behave differently next time. And it cannot show you what it knows, because what it knows is not anywhere in particular — it is smeared across billions of weights.
None of that is fixable with more capital. It is what a next-token predictor is. The guardrails, retrieval layers, citation wrappers and evaluation harnesses now being built around language models are a large and fast-growing industry, and every one of them is an attempt to bolt on from outside what 7 has by construction.
It can say nothing
Where the structure has no answer, a hole is the output. Not a tuned refusal rate, not a confidence threshold — an absence the architecture cannot paper over, because it has no objective that would reward papering over it.
It can be corrected
Contradict 7 and it revises in the session, permanently, without losing the record of what it was originally told. No retraining run, no fine-tune, no patch deployed next quarter.
It can be inspected
One point is one thing. What 7 holds can be read back item by item — what it knows, where each piece came from, what changed and when.
That is not a better chatbot. It is a different instrument. The commercial question was never whether 7 can out-write a language model — it cannot, and is not trying to. It is whether there is durable value in a system that can be held to account. Every regulated sector, every safety-critical procedure, every institution that has to stand behind its own answers is currently being asked to accept a system that structurally cannot provide one.
Investing here is not a bet against language models. It is a bet that the market for language and the market for accountability are not the same market, and that the second one needs a different machine.
Architecture
Built the other way around
7 resembles a neural network in shape and differs from one in all three of the things that define one: connectionist in topology, symbolic in representation, non-gradient in learning. There is no loss function, no backpropagation and no fitted parameter anywhere in it.
A large language model
- Distributed — a concept is a vector smeared across billions of weights, inspectable only statistically.
- Learns by gradient descent against an objective.
- Optimises the likelihood of the next token. Fluency is the target, so fluency is what you get.
- Cannot separate understood the question from produced some text.
- A gap in knowledge and a confident answer are the same operation.
- The unit of change is a training run.
7seven
- Localist — one point is one thing. What 7 holds can be read back and audited item by item.
- Learns by counts, decay and co-occurrence. No gradient, no backpropagation.
- No objective function at all — nothing to optimise toward, so nothing to confabulate for.
- Comprehension and response are separate channels, separately recorded and separately scored, on every turn.
- A missing answer is a first-class output: a hole, not a guess.
- The unit of change is a conversation. The field adjusts in real time, in the session.
Three prohibitions do most of the work. Each is enforced from the first line of the architecture, and each closes off a failure mode that the rest of the field patches after the fact.
No human vocabulary inside
The interior holds opaque structure. English lives only on the translator that reads that structure back for a human. Comparable architectures name their internal atoms in English; keeping ours nameless is a discipline almost nobody holds.
It is also what makes “did 7 understand that, or did we tell it?” an answerable question rather than a matter of opinion.
No objective function
No reward, no fitted parameter, no score for being right. Correctness is never fed back into the field — our own rules forbid it.
Systems trained to look right learn to look right. Here that failure mode is designed out at the foundation rather than filtered downstream.
Curator physics
A strict, enforced bound on how much of the field may ever be examined for one thought. Most architectures scan memory freely; 7 made tractability a constraint from the start.
It is why the system runs on commodity hardware instead of a datacentre, and why inference does not explode as the field fills.
A symbolic cognitive architecture in the ACT-R / NARS / OpenCog line, distinguished by a hard prohibition on human vocabulary inside the field, the absence of any objective function, and a curator bound on retrieval — pursued as a System-2-first alternative to the neural mainstream rather than as a neurosymbolic hybrid.Our technical placement, stated as we would defend it in review
Evidence
What v0.1 can do, and how we know
7 is taught by conversation, in ordinary English, on a graded language scheme of the kind used with children. Nothing below is scored on the material it was taught with: every arm runs on held-out probes that appear in no fixture, against a field raised to full curriculum density.
| Capability | Score | What the number means |
|---|---|---|
| Role comprehension | 1.000 | Who did what, to what, where — recovered correctly, including from combinations of known words never seen together. |
| Order sensitivity | 1.000 | “The cat sat on the mat” and “the mat sat on the cat” are held as different facts, not as the same bag of words. |
| Hole honesty | 1.000 | Where the field genuinely lacks the answer, 7 says so. It does not fill the gap with the nearest plausible thing. |
| No confabulation | 1.000 | Anything 7 volunteers that is not the answer is marked as a distinct fact, never slipped in as one. |
| Vocabulary growth | 1.000 | New words are minted mid-conversation and used correctly on the following turn — 96 to 99 known words across an eight-turn dialogue, each new one marked provisional. |
| Cross-turn coherence | 1.000 | A meeting stays live across turns; what was established earlier is still in play later. |
| Conversational move fit | 0.889 | Choosing the right move for the turn — answer, ask back, or say it does not know. The weakest of the conversation numbers, and a known work item. |
| Structure mapping (analogy) | 1.000 | Mapping shared relational structure across different constructions, with the mapping inspectable pair by pair. Sat at 0.286 for months; it was the last number on the board to move. |
| Density scorecard | 13 / 13 | Every capability re-checked on the full field rather than on a convenient fixture. Density is where architectures of this class historically fail. |
Held-out probe results at full curriculum density, August 2026. Scores are pass rates on probe sets, not accuracy against a public benchmark — this architecture has no leaderboard to enter.
typed understood
─────────────────────────────────────────────────────────────────────────
the cat sat on the mat who=cat doing=sat what=mat where=on
the mat sat on the cat who=mat doing=sat what=cat where=on
the cat sat under the mat who=cat doing=sat what=mat where=under
the dog chased the cat who=dog doing=chased what=cat
the cup sat on the plate accepted
the plate sat on the table accepted
is the cup on the table yes · derived ← never stated; composed
Comprehension checks and a two-premise composition on a freshly rested field. derived is a marker the field sets on itself: it separates an answer 7 worked out from one it was handed. Nothing here is a retrieval.
Potential
What this becomes at scale
v0.1 holds a few hundred concepts. The open question — and it is the honest centre of the project — is whether a field of this kind stays coherent at a hundred thousand. Everything in this section assumes it does, and we are stating that as the assumption it is.
The assumption is not that 7 becomes a better writer. It is that the four properties above — says why, says nothing, corrects in place, can be inspected — survive contact with a serious body of knowledge. If they do, a set of things become possible that no amount of scaling a language model delivers.
Knowledge an institution can own
A bank, a hospital trust, a regulator or a defence organisation teaches 7 its own material over years. What accumulates is not a model checkpoint that must be retrained and revalidated — it is a body of structure that can be read, audited, corrected by the person who spotted the error, and inherited by whoever comes next. Institutional memory that outlives the people who built it.
Answers that carry their working
At scale, a system that distinguishes what it derived from what it was told is a machine that can tell you which of its own conclusions are new. Point it at a literature, a case history or a body of regulation and it composes consequences nobody stated — each one arriving with the chain that produced it, so a human can check the reasoning rather than trust the fluency.
Advice where being wrong is expensive
Clinical protocol, financial suitability, legal procedure, airworthiness, safety-critical doctrine. These are domains where a plausible wrong answer is a liability event and “I have no grounds for that” is a perfectly acceptable one. They are currently the hardest markets for language models to enter, for exactly the reason they are the natural ones for 7.
The grounded half of a hybrid
The nearest-term shape is not replacement. A language model writes and converses; 7 holds what is actually known, answers what it can stand behind, and flags where the fluent layer has left the record. Every LLM deployment that hits a hallucination wall is a lead, which makes the incumbents’ success a tailwind rather than a threat.
Systems that keep learning after deployment
An agent whose knowledge is frozen at training time and topped up with a context window is a fundamentally short-memoried thing. One that revises in real time, permanently, without losing its record of what it was told, is a different substrate for long-lived autonomy — and a far easier one to govern, because every change is a legible event.
Teaching that can see understanding
Because comprehension and response are separate channels, 7 is already a system where what was understood is a recorded, inspectable object rather than an inference from what was said. Turned outward, that is the missing instrument in educational technology: not marking the answer, but modelling where the gap actually is.
None of this requires 7 to beat a language model at language. It requires 7 to stay trustworthy about structure as the structure grows — a narrower thing to prove, with a clear experimental path to proving it.
The test
The bar we wrote before we built toward it
“Processes sentences, concepts and theories, and responds alike” is easy to assert and hard to check. So we fixed a falsifiable version first, and have been building toward it since:
The bar
7 is taught a domain it has never seen, entirely by conversation. It then answers questions about that domain whose answers require composing two or more things it was told; says it does not know where it genuinely lacks the answer; revises when contradicted, without losing the record; explains which steps produced an answer; and none of the answers are retrievals of something it was given.
Every clause maps to an instrument that already exists, and the mechanism behind each one is built and tested. The bar itself is not yet passed end to end on a domain 7 has never seen. Passing it is what the current programme is for, and it is the next milestone we will report against.
Straight answers
Where v0.1 actually stands
Our internal rules require every claimed capability to carry its own disclaimer and forbid describing any result as progress toward AGI. We hold the public version of the site to the same standard.
- Scale is the betThe field has never held more than a few hundred concepts. Whether a localist field stays coherent at 10³–10⁵ is genuinely unresolved — Cyc is the cautionary case, and the answer comes from measurement rather than argument. Density is now a dial we turn deliberately, with cliff detection on the curve. This is the single most important open question, and it is what capital and people are for.
- The bar is unpassedEvery mechanism the bar names is built. It has not been passed end to end on a domain 7 has never seen. Until it is, that claim stays an intention.
- Theories are still thinDerivation, repair, reuse and explanation exist and are tested. Whether they compose into something a domain expert would call a theory is unsettled.
- Vision is boundedVector scenes, screenshots and constrained photographs feed the same field as words, and 7 can be asked about an image. This is not general visual understanding.
- Speech is not started7 works from written language. The sound half — phonemes, digraphs — waits on a diet it has not been given.
- This is not AGIIt is version 0.1 of an architecture with a checkable bar in front of it. Our own documents refuse the word, and so do we.
For investors
The shape of the opportunity
The case against building this is that the incumbents are enormous. The case for it is that their architecture forbids them from following where 7 goes, and that the properties buyers in regulated markets keep asking for are the ones a next-token predictor cannot structurally provide.
Two routes to market
Standalone, in domains where a confident wrong answer costs more than no answer at all. Or alongside a language model, as the grounded layer that holds what is actually known and says when the fluent layer has left the record. The second route needs no one to switch from anything.
Raised, not scraped
Each customer’s 7 is taught from their own approved material, by conversation and by structured load. What it holds is theirs, inspectable, and auditable item by item. There is no shared web corpus and no shared weights — which also removes an entire category of provenance and copyright exposure.
A different cost structure
No training cluster, no gradient runs, no per-query GPU bill. Tractability is enforced by the architecture, which is why it runs on commodity hardware — including inside a building that will never let its data leave. That changes unit economics in kind, not by a percentage.
Asymmetric downside
If the scale question resolves, the category is new and has no incumbent in it. If it only partly resolves, a system that is trustworthy over a bounded, high-consequence domain is still a product with real buyers. The bad case is a narrower business, not a dead one.
Stage, stated plainly: alpha, pre-revenue, pre-launch. The architecture is built, the instruments exist, and the programme to pass the bar on a novel domain is beginning now. Capital and people go toward answering the scale question and passing that bar — not toward a launch date we have no business promising.
The defensible position is not a prompt or a dataset. It is the field physics, the raising pipeline, and the instrumented record of which design choices survived contact with density. We share principles and outcomes publicly, and implementation under NDA.
Building the team
The hard problems, and who we want on them
We are hiring research fellows and product engineers who want to build a cognitive architecture rather than another wrapper around someone else’s model. The direct comparators are ACT-R, Soar, NARS and OpenCog, and we name them whenever the project is described — spreading activation, chunking, frames, case roles, construction grammar and graded belief revision are all 1968–1995 work, and we claim the combination, not the parts.
Coherence at scale
Does a localist, objective-free field stay coherent at three to five orders of magnitude more concepts than it has ever held? Every cliff we have found so far was found by turning density up deliberately. This is the work that decides everything else.
Judgement without a judge
Two grown tools already compete on the same input and one loses, with no operator score anywhere in the loop. Whether richer settling emerges from connectivity alone — rather than from a scoring module bolted on — is the question the architecture exists to ask.
Structure mapping and inference
Analogical mapping now licenses candidate inferences in a target domain, marked weaker than observed fact. How far it can carry inference in a resource-bounded field is open, and the mapping is inspectable, which makes it tractable to study.
Raising as a variable
Because 7 adjusts in real time, what it holds moves with who teaches it — as with a child. We have deliberately not decided whether that is a risk to close or a property to exploit. It is a live question with obvious governance weight.
Access
Talk to 7
The v0.1 demonstration is invite-only and accounts are enabled individually. What you meet is a field raised on a graded language curriculum — small, early, and honest about its edges. Teach it something and watch it use the word on the next turn; ask it something it has no grounds for and watch it decline to invent one.
The interface returns one line of ordinary English per turn. No field state, no structure and no internals cross the network.