Papers and downloads
Every paper, specification and report, with the page it is presented on and the PDF where one exists. Sizes are as measured on disk. Open releases carry their license.
LTMi-XT v0.1 specification
The open retrieval format: single-claim loci, a four-level breadcrumb, BLAKE2b lattice coordinates, byte-offset provenance, horizon and decay, with the TypeScript reference implementation.
XE5: memory-conditioned specialist mixing
A 147.5M model trained from scratch in which specialists are combined by a learned composer rather than selected by a gate, with the weights conditioned on where a descent through runtime-written memory terminates. A mixture of experts computes w = gate(x); XE5 computes w = compose(x, m), so fixing the input does not fix the weights.
XE4-S: dense composition after a routing collapse
The previous architecture selected among specialists and collapsed past N = 3: top-1 route accuracy fell from 100% at N = 2 to 23.3%, a learned replacement router carried exactly 0.0 bits of mutual information, and enlarging one corpus twentyfold made things worse. This is what was built instead.
Cassandra T2 design document
Successor design for the Cassandra masked-diffusion model: architecture targets, training plan, and its relationship to the released Cassandra T1.
Anchor-token masking, paper
Three masking objectives compared on identical data. Anchor-token masking gives a 1.67× held-out advantage over random-position masking with a 95% interval of [1.51, 1.85]; four mechanism probes end in a V/O ablation. Deposited on Zenodo under DOI 10.5281/zenodo.20496783, CC-BY-4.0.
Anchor-token masking, technical report
The report form of the same study: setup, raw measurement tables, the reverse-asymmetric control, attention and weight-allocation probes, and the reproduction recipe.
Constellation
The lane router behind the lab: six registered lanes, an arbiter with fixed escalation and release rules, and the Agent Channel the lanes talk over, with the rules replayed live on the page.
Stack architecture
The memory, verifier, compute, coordination and deployment layers as one diagram built from the components in the repository.
Magnetic confinement via dual rotating tungsten-lined plate discs
A concept study of plasma compression between two counter-rotating discs, with design targets, a Grad-Shafranov solver and a reconnection model on the page.
Rico Inf Curve, research report
Particle streams in a four-singularity potential well: verification, a batched GPU integrator, a 99,856-particle production run, a dimensional scan, a reversibility experiment and the geodesic module.
EM-D Stream, design specification Rev D
Document of record for the propellantless vehicle: three drive mechanisms, validated and negative results, the field-reversed core, the crewed configuration, the test article and the corrections register.
EM-D Stream, mathematical supplement
Derivations for the field topology, momentum transfer, the radiation bound, the circuit condition and hull loading, written against Rev C.
EM-D Stream, test plan and validation record
Bench article definition, measurement methodology, error budget, acceptance criteria and the complete simulation record including every negative result.
Ternary quantum
Metaplectic anyon language models: three-state logic for model computation on topological quantum architectures.
Digital feudalism
An analysis of the concentration of AI power, platform monopolies and governance structures for equitable access.
SOPHIA custom training methodology
The per-customer fine-tuning service: anchor-token-masked LoRA training, the evaluation harness and data preparation.
LENS-X specification
A declarative specification language for token-level deterministic generation: the open .lensx format, runtime semantics and anchor preservation.
SOPHIA XT technical whitepaper
The full lab in one document: research division, model family, retrieval architecture, multi-agent orchestration and applied products.
AGI alignment
Safety frameworks that treat oversight and intervention as design constraints.
AI and the web
How machine readers change information access and web infrastructure.
Ethical AI
Development frameworks for bias, transparency and human oversight.
Fluid dynamics
Computational fluid simulations with lattice Boltzmann methods and Navier-Stokes solvers.
Physics
Mercury induction, magnetohydrodynamics, differential rotation and field generation in conducting fluids.
Quantum computing
Exploratory work where quantum logic, non-Abelian braiding and model computation meet.
Spatial comprehension
Scene understanding, object permanence and spatial reasoning for embodied systems.
Telemetric pathfinding
Real-time spatial navigation with sensor fusion.
Open releases on GitHub
The LTMi-XT reference implementation and the Cassandra T1 model repository are published under Apache 2.0 with their benchmark corpora, evaluation scripts and checksums.