Research
Every study here was pre-registered on the Open Science Framework before data collection, and every deposit carries a DOI and a replication package. Nothing on this page asks to be taken on trust.
Current Research Program
A current research program built on a foundational point of view: every scientific instrument has limits, and the discipline of asking what each instrument can and cannot adjudicate is the move that everything downstream depends on. The Robot in the Dark sets out the point of view itself. The pre-registered cross-linguistic scientific work listed below is its empirical application. The Honest Code book and the Open Honest framework and audit apply it to enterprise software. The Replicators is the metaphysical synthesis that arises when the foundation, the empirical results, and a particular philosophical structure are taken together: a parsimonious account of the origin of complex emergent structure in which many collaborative intelligences operate across substrates and time.
The Robot in the Dark (foundational; manuscript in preparation)
Every scientific instrument has an edge where the light runs out. The most consequential moments in science are not what happens inside the illuminated region but what happens at the boundary. The book traces the pattern through evolutionary biology, physics, mathematics, large language models, and finally consciousness itself, arguing that instruments which name their edges become more authoritative, not less. Subtitle: Instruments, Illusions, and the Limits of Knowledge.
Pre-registered scientific work (empirical application)
A pre-registered program of controlled experiments on cross-linguistic large language model training dynamics. Four deposited preprints; child-scale extension accepted for presentation at EMNLP 2026 Budapest. Detailed below.
Honest Code and Open Honest (operational application)
The same point of view applied to enterprise software: a commercially-published book of principles, a correct-by-construction development framework whose limits are bounded by architecture rather than by hope, and a peer-reviewable measurement instrument whose own thresholds are explicit and falsifiable. Eight pre-registered studies (Papers A through H) sit alongside the working code. Detailed below.
The Replicators (metaphysical synthesis; in preparation)
Complex emergent structure (biological evolution, software systems, language itself) is more parsimoniously explained by many collaborative intelligences operating across substrates and time than by either single-creator or no-creator accounts. The argument arises out of three things taken together: Robot's foundational point of view, the empirical results from the cross-linguistic scientific work, and a particular philosophical structure that includes the convergence-of-signals method (multiple independent instruments meeting at the same boundary), the discount-bias diagnosis (in which our use of consciousness as the measuring instrument leads us to discount any agency that does not look like human consciousness), and the silicon-vs-DNA parsimony argument (if binary silicon transformers can plausibly host intelligence, then DNA codons with sixty-four possible permutations doubly so). Subtitle: Multiple Collaborative Intelligences as the Substrate of Complex Structure.
Pre-registered empirical research
A pre-registered program of controlled experiments on cross-linguistic LLM training dynamics, with all four lead deposits in place, the child-scale extension accepted for peer-reviewed presentation, and a geometric-loss-function extension in pre-registration. Collaborators: David Beauchemin (Laval) on the BabyLM 2026 paper and the EMNLP main-conference extension; Hafedh Mili (UQAM); Edward Levin (VM4AI) on the geometric-loss study.
The Scaling Hypothesis Is Language-Contingent
Pre-registered controlled ablation training identical 125M-parameter transformers on matched English and French C4 corpora. French achieves grammatical competence at 197M tokens; English remains at chance through 4+ billion tokens. Pre-registration: OSF SJ48B. Training logs: github.com/adamzwasserman/fractal-language.
English Considered Harmful
Companion preprint to the language-contingent scaling line.
Geometric Loss Functions and Cross-Linguistic Training Dynamics (in preparation)
Joint work with Edward Levin (VM4AI) extending the language-contingent-scaling line into training dynamics. Tests whether geometric topologies (Polytope, Sphere) applied as loss-function constraints during pretraining can replicate the structural advantages morphologically rich languages provide naturally. Pre-registrations in preparation.
The 70% Rule: When Axiomatic Prompting Helps, and When It Hurts
Pre-registered work on the conditions under which explicit logical axioms in prompts produce qualitatively different outputs on reasoning benchmarks. OSF pre-registrations: 7Z49A, MZF79.
Process Discipline
Methodological preprint on pre-registered empirical practice for AI-assisted enterprise software development.
Finite Testability of Enterprise Software
A 200-repository, 2.2-million-function cross-language survey of mutable-state ratio. The median enterprise codebase in Python, Java and C# is structurally incapable of exhaustive behavioural testing. Paper A of the Open Honest empirical program. Pre-registration: OSF DBSYG.
Process Discipline as the Key Variable in AI-Assisted Enterprise Software Development
A natural experiment: the same team and the same AI tools satisfied 18 of 18 enterprise quality dimensions under a structured SDLC, and 2 of 18 once process authority moved to non-technical stakeholders. Measured across 287 FTE-days.
The Scaling Hypothesis Is Language-Contingent
A pre-registered controlled ablation training identical 125M-parameter transformers on matched English and French corpora. Scaling behaviour differs by language, so the universality assumption is doing work it has not earned. Pre-registration: OSF SJ48B.
English Considered Harmful
Mixing English into a training corpus degrades both perplexity and structural pattern acquisition in structurally richer languages. Three controlled comparisons at 125M parameters.
The 70% Rule
Axiomatic prompting improves LLM classification accuracy by 15 to 24 points where zero-shot performance is below 70%, and degrades it above that threshold. Six tasks, eight models, five providers.
The Robot in the Dark
Instruments, illusions, and the limits of knowledge. A robot maps what its headlamp illuminates and concludes the illuminated patch is the building. The book names that error and traces it through the philosophy of measurement to claims about large language models.