Emergent Graph Theory and what follows from it: language, consciousness, epistemics. The work is ongoing; what gets published is what survives being recomputed — carrying the stage a piece of work is actually at, not a riper one.
Local rules, gradient descent, and a data-grown context graph on a common character-level harness
A comparison across a family of mechanisms with the harness held fixed: does language capability come from the substrate's own dynamics, from the learning rule applied to it — or from neither? The answer went against the programme's own expectation. The best model is the parameter-free context graph; the local associative rules do not learn language at all.
Where bounded memory helps next-character prediction across data and capacity, and where local statistics meet their ceiling
Substrate, data and evaluation are held fixed; only the memory regime varies. Bounded forgetting lowers held-out bits-per-character at every corpus size, and the margin grows as data gets scarcer. No regime lifts the ceiling itself: at 11.3 million characters the generated text is orthographically perfect and has no long-range coherence at all.
A two-mechanism architecture for an open-ended language system on modest hardware
A design paper, not a results paper. It separates two capabilities that language models conflate — sequential binding across distance and open-ended knowledge selection — and gives each its own mechanism. What is measured, what is hypothesised and what remains open is fixed in advance in a table of its own.
An open inquiry into soul, consciousness, and sentience in computational systems
Not an argument for an answer but for a better vocabulary: soul as sedimented informational structure, consciousness as the interpretive operation upon it, sentience as interpretation extended in continuous time. The tripartite distinction maps onto the ECQ triad that governs structure formation at the physical level. The hard problem is explicitly not solved.
A working hypothesis on the convergence of substrate, mediation, and reflection
What we ordinarily call consciousness is, looked at closely, not a single property but the convergence of several structurally independent operations. Three of them are proposed as axes. Explicitly an early draft: released for discussion, not for citation.
A graph-theoretic reading of sensory processing
Perception read not as a mapping but as the reorganisation of a substrate under energy flow. A work in progress, released for discussion and not submitted for publication.
On the conditions under which probabilistic inference is valid, and why fluent systems obscure them
The validity of a prediction is not intrinsic to the method — it is borrowed from the process that generates the data. Where that process is non-stationary, reflexive or singular, the same machinery produces outputs formally indistinguishable from valid ones. Language models sharpen this, because their register stays constant across every domain and their uncertainty is thereby inaudible.
A structural limitation of RLHF and an architectural complement
RLHF assumes that values can be trained into a system independently of the subject-position from which that system operates. Examined closely, the assumption does not hold. The paper names the limit and proposes an architectural complement rather than discarding the method — whose success it explicitly declines to minimise.
Resolving the Fermi paradox through information-theoretic constraints on intelligence
Any sufficiently advanced intelligence discovers an optimisation principle over energy, information and time — and pursuing it blindly proves lethal. Four optimisation filters, with particular weight on the AI trap. That stable matter in molecular simulation converges on a subcritical attractor rather than on the optimum serves as the physical argument.
A phonetic-morphological architecture for a low-friction operation grammar
A constructed language as an interface between a human and a language model, not between humans. Phonetic categories, a closed suffix matrix and a four-position architecture map onto a tightly specified semantic surface. Presented as a linguistic specification rather than a metaphysical proposal — with falsifiable claims and a method for testing them.
The foundational paper — Emergent Graph Theory — and the papers on the individual systems are not here but in the software section: each with the system it describes.
To the software →Questions about a paper, criticism, an error in the arithmetic: connect@pantareon.io