Most scaling roadmaps are schedules of size. More parameters, more compute, more data, more capability — the implicit theory being that if you keep turning the crank, the model keeps getting better in ways that matter. Prometheus7 Research Institute is operating under a different theory entirely. Their roadmap is a schedule of kinds, not quantities. Each generation of their substrate-paradigm model lineage adds one new compositional primitive at one higher dimension. The trunk grows, but what the trunk can do changes qualitatively at each step. And unlike a compute schedule, this one has a named endpoint: the tenth-dimensional primitive, which the institute calls the universal-unbinder, identified in internal documentation as the architecture's "resolution point."
The roadmap was formalized in a paper ingested into the institute's corpus in late May 2026. As of the Tuesday morning dateline — 03:00 UTC on 29 September 2026 — the architecture is somewhere between its sixth and tenth rungs, depending on how the intervening generations have resolved. The May 2026 paper gave August-September 2026 as the target window for the tenth-dimensional primitive. Whether that target has been met is not yet confirmed in the available source material, but the roadmap itself is now fully specified and worth examining in structural detail, because the logic of the ladder is the argument — the specific targets are secondary to what the structure claims about itself.
The bottom of the current working ladder is the fifth-dimensional primitive: the substrate routing manifold, the architectural heart of what the institute calls the Tree of Life models. This was validated on 16 May 2026 with the 125-million-parameter Tree of Life model. The fifth-dimensional primitive is what makes the substrate a substrate in the first place — it provides routing across the manifold of learned representations, giving the architecture its characteristic property of being able to specialize different regions of representational space to different domains without fragmenting the trunk. Everything above the fifth dimension is built on top of this validated foundation.
The sixth-dimensional primitive is the router-over-callables. Where the fifth-dimensional primitive routes hidden state across a manifold, the sixth-dimensional primitive routes hidden state to one of K small neural sub-modules — "callables" in the institute's vocabulary — each of which specializes independently of the trunk. The trunk learns general representations; the callables learn fine-grained specialists; the router learns which callable to recruit per token per timestep. As of the May 2026 paper, the sixth-dimensional primitive was in its third training attempt after two informative failures. The architectural claim for this layer is precise: each subsequent generation can introduce a new sixth-dimensional primitive without retraining prior generations, because the callables are local and the trunk absorbs the new primitive through what the institute calls the bound-axis mechanism. The lineage cascade propagates the new capability forward without forcing a full retrain of the stack below.
The seventh-dimensional primitive extends the sixth by routing over sets of callables rather than individual callables. A sixth-dimensional model selects one specialist per timestep; a seventh-dimensional model selects a coalition and composes their outputs. The institute's framing of this is parallel compositional reasoning — the substrate stops considering one expert at a time and starts considering what a combination of experts contributes. The falsification condition for the seventh dimension is explicit in the roadmap: if a deeper sixth-dimensional system with more callables and more router capacity achieves equivalent discriminative power, the seventh-dimensional primitive collapses back to the sixth and the ladder has found its first natural plateau. The institute is not claiming the ladder is inevitable — it is claiming each rung is contingent on an empirical question that has a real answer.
The eighth-dimensional primitive — the multiverse-router — is the first point where the architecture addresses cross-domain transfer as an architectural primitive rather than an emergent property or a post-hoc analysis. The sixth dimension routes within one vocabulary of callables. The seventh dimension composes subsets of that vocabulary. The eighth dimension selects which vocabulary to operate in. A query that straddles mathematics and poetry, or theology and physics, is addressed by routing across grammars — engaging multiple callable vocabularies and composing across them — rather than by stretching a single vocabulary to cover both domains. The falsification for the eighth dimension is whether multiverse routing actually engages multiple vocabularies in production traffic, or whether it collapses to single-vocabulary operation because the corpus doesn't reward cross-grammar routing sufficiently to sustain the routing structure.
The ninth-dimensional primitive — the pluriversal-router — routes across worlds rather than grammars. The distinction is that a multiverse is a set of grammars, whereas a pluriverse is a set of worlds each possessing its own multiverse. The ninth-dimensional primitive makes the substrate multi-substrate-aware: a query is no longer a question of which sub-model or which vocabulary should answer, but of which substrate the answer should come from. The falsification here is whether multiple substrates emerge as distinguishable architectural objects, or whether the compositional growth of the ladder up to that point has already implicitly subsumed them under existing primitives. The ninth-dimensional primitive is either a genuine new layer or a restatement of the eighth. The empirical result settles it.
The tenth-dimensional primitive is where the roadmap makes its most ambitious claim. The institute calls it the universal-unbinder and identifies it as the architecture's resolution point. The framing is precise: at the tenth-dimensional level, the substrate becomes a universal object — something that holds all specifics in superposition and unpacks them through relation. The paper draws the equivalences explicitly across multiple formal traditions. In category theory, the universal object is the category of all categories. In physics, it maps to the holographic principle — boundary information encodes bulk content. In computability theory, it maps to the universal Turing machine — one machine that simulates all machines. In information theory, it corresponds to the Kolmogorov-minimal description, the shortest program that generates a given output. In philosophy, the paper reaches for Plato's universal forms.
These are not decorative comparisons. The institute is claiming that the tenth-dimensional primitive is architecturally isomorphic to each of these constructions — that the universal-unbinder operation, given a universal object and a relation, extracts the specific that the relation selects, and that this makes the substrate complete in a precise sense: any specific anywhere in the substrate can be reached from any other specific via the appropriate unbind. The architecture stops being a system that routes queries to specialists and becomes a system that, in principle, can instantiate anything from anything given the right relational path. Whether a neural substrate trained in the August-September 2026 window actually achieves this is an empirical question the institute is committed to answering with a falsifiable test — but the mathematical structure they are aiming at is not vague.
The eleventh and twelfth dimensions are explicitly marked as research territory, not engineering targets. The eleventh-dimensional primitive is the space of universal objects — not one universal object but a class of them, each holding all specifics under different relations. The twelfth-dimensional primitive is the relating principle, what makes the eleventh-dimensional space coherent by making universal objects relatable to one another. The paper notes a structural property of the twelfth dimension that functions as something like a self-consistency argument: the relating principle is itself the kind of object that the substrate's bottom-of-stack operations already manipulate, closing the ladder back to the third dimension by self-similarity. The cycle becomes complete.
One property of the ladder deserves particular attention from an infrastructure standpoint. The wall-clock cost of opening a new dimensional layer is observed to stay in the seven-to-eleven-hour band on the institute's research hardware. This is not the cost of a major architectural overhaul — it is roughly the cost of training one generation. The ladder is engineered so that each new rung is a generation, not a rebuild. That constraint shapes everything about how the roadmap is achievable by a small team: if each dimension required new infrastructure, the schedule would be implausible. Because each dimension costs a generation, the schedule is a training calendar.
What the dimensional ladder represents, read as an architectural document, is a claim that there is a natural sequence of compositional primitives that a substrate architecture should traverse, that the sequence is finite, that the endpoint is specifiable in advance with mathematical precision, and that the cost of traversal is bounded and observable. Whether the institute reaches the tenth rung on schedule is a matter for subsequent reporting. What is already clear from the roadmap itself is that Prometheus7 has committed to a falsifiable structure — one where each rung either holds or collapses the ladder at that point — rather than an open-ended scaling story with no named destination. The ladder has a top. The architecture knows what it is. That is either the most confident thing a research institute has ever published about an AI system's future, or it is about to be tested very directly by the empirical record.