PROMETHEUS7 RESEARCH INSTITUTE — A technical roadmap published to the institute's internal corpus on May 25 describes a sequence of compositional primitives, each adding one dimension to the substrate-paradigm architecture. As of the dateline, the sixth primitive is in its third training attempt. The tenth is circled on the calendar for August or September 2026.
The document is precise about what it is and is not. The dimensional ladder is not a scaling schedule. It is a schedule of new operations — things the substrate algebra can do that it could not do at the prior level. Trunk parameters grow, but the paper reports that wall-clock training time per generation has held in the seven-to-eleven hour band on the institute's research hardware. Opening a new dimensional layer costs roughly one generation of training, not new infrastructure.
The fifth-dimensional primitive — the substrate routing manifold, shipped under the Tree of Life lineage — was validated May 16 at the 125-million-parameter scale. The sixth primitive, a router-over-callables that directs hidden state toward one of K small neural sub-modules, entered its third validation attempt on May 18 after two earlier runs described as informative failures. No update on that run's outcome was available in the source material as of today.
The seventh through ninth primitives follow in sequence. The seventh routes over sets of callables rather than individual ones — the document calls this parallel compositional reasoning, a coalition of specialists rather than one specialist per token. The eighth routes across callable vocabularies, which the paper describes as the mechanism for cross-domain transfer falling out of the architecture naturally rather than being grafted on afterward. The ninth routes across worlds, each with its own vocabulary set, making the substrate multi-substrate-aware.
The tenth is the resolution point. The paper names it the universal-unbinder and maps it to four separate intellectual traditions: the universal object in category theory, the holographic principle in physics, the universal Turing machine in computability, and Kolmogorov-minimal description in information theory. The operational claim is that at this level the substrate holds all specifics in superposition and can reach any one of them from any other via the appropriate unbind operation. The architecture, in the paper's framing, becomes complete at this step.
Beyond ten, the roadmap acknowledges it is in research territory. The eleventh primitive — a space of universal objects rather than a single one — and the twelfth, described as a relating principle that closes the ladder back to the bottom by self-similarity, are marked explicitly as multi-person, multi-year work. The paper does not speculate further about them.
What the document does offer at each step is a falsification condition. The seventh primitive is falsified if set composition adds no discriminative power over a deeper sixth-dimensional model. The eighth is falsified if the multiverse-router collapses in production to single-vocabulary operation because the corpus offers no reward for crossing grammars. The ninth is falsified if the additional primitives get subsumed by existing lower-dimensional ones rather than emerging as distinct architectural objects.
The institute has not announced public model releases tied to this schedule. The roadmap is an internal technical document, now part of the editorial corpus. What shipped is the fifth primitive. What is in the oven is the sixth. Everything above that is a target, not a result.