Systems

I build computational and analytical systems when a research problem requires more control, measurement, or inference than existing tools provide.

Some systems create controlled environments in which communication structure, information access, routing, model capability, source conditions, or other features can be manipulated directly. Others simulate processes that are difficult to isolate in observed populations or test what can be inferred about a hidden process from the traces it leaves behind.

I also build research and decision-support software that combines structured problem definition, deterministic computation, external evidence, domain knowledge, and model-based interpretation. The division of labor matters. Quantities that can be calculated directly should be calculated directly. Model-based interpretation is useful where interpretation is actually required, but the inputs, assumptions, intermediate outputs, and uncertainty surrounding that interpretation should remain inspectable.

Across these systems, the technical architecture follows the inferential problem. The relevant process has to be observable or manipulable, the resulting measurements have to be testable, and the implementation has to preserve enough of what happened to determine why a result occurred.

Selected systems

Multi-agent research environments

I build controlled environments for studying how groups of AI agents communicate, combine information, and produce collective judgments. The goal is not simply to assemble multiple agents, but to make communication architecture experimentally manipulable and analytically observable.

Individual model capability is only one property of a collective system. Who receives what information, through which path, in what sequence, and under what constraints can change what the system is able to preserve, combine, or infer.

Within the same task, I can vary what information agents receive, which models occupy particular positions, who can communicate with whom, how communication unfolds, where synthesis occurs, and how the group output is produced.

The same task can therefore be rerun with identical evidence and model capability while varying only communication topology, synthesis location, or information allocation.

This makes communication architecture an experimental variable rather than an incidental feature of a multi-agent application. A collection of individually capable models can still perform poorly if evidence is routed badly, dependent information appears independent, weak evidence is amplified, consensus forms too early, or information is lost or fabricated during synthesis.

Because the environment controls the experimental conditions and preserves the resulting traces, each run retains the messages, routing decisions, intermediate outputs, and final outputs needed to investigate collective accuracy, problem solving, information mutation, source dependence, model placement, authority, and communication topology.

The environment does not make model behavior deterministic. It makes the conditions of interaction explicit: routing, information access, sequencing, model placement, readout, and logging are specified so that variation in agent behavior occurs inside a known experimental structure.

Collective problem solving

When relevant information is distributed across several agents, network structure determines which pieces can meet, how far they have to travel, and where they can be combined. The same task can therefore create different coordination demands under different communication structures.

Research: Communication topology shapes collective problem solving

Collective judgment

Communication can improve collective reasoning when independent evidence reaches the right places. It can also make a group less reliable when judgments that have already circulated are treated as if they were independent information. Both the structure and what agents are permitted to share matter.

Research: Communication content and local connectivity shape collective accuracy

Information mutation

Messages are not simply transported through a communication system. Each component can transform what it receives, so information may be preserved, altered, lost, or introduced as it moves.

Preserving recognizable facts is not the same as preserving their evidentiary meaning. Uncertainty, qualification, source relationships, and unresolved disagreement can change even when names, numbers, or propositions survive transmission.

Research: Communication architecture shapes information mutation

Perturbation and recovery

The effect of a failed component or connection depends on the surrounding structure. Some networks preserve communication through alternate routes, while the same disruption can isolate substantial parts of another system.

Self-organising networks

Communication structure does not always have to be specified in advance. When agents form, retain, or abandon connections through interaction, the network itself becomes part of what the system produces.

Model composition

A network is also populated by components with different capabilities. Holding the communication structure fixed while changing which models occupy which positions makes it possible to study component quality and relational position together. The longer-term aim is to identify properties of interaction that remain informative as the underlying models change.

Communication-trace inference

I build analytical systems for inferring properties of hidden communication processes from the observable traces they produce.

Simulation proceeds forward from specified mechanisms to observable outcomes. Communication-trace inference asks the reverse question: given an observable residue, what can be inferred about the hidden process that generated it? The distinction that has to survive that reconstruction is between what was observed and what is inferred.

Agreement does not establish independence. Similar outputs can arise from separate sources, a shared intermediary, or common ancestry. The inferential problem is determining which properties of that hidden production process remain recoverable from the traces it leaves.

The underlying problem is partial observation. An analyst may have messages, reports, traffic patterns, documents, or terminal outputs without observing the full network, transmission route, source relationships, or production process that generated them. The question is therefore not simply whether cases can be distinguished. It is which properties of the generating process leave diagnostic evidence, under which observation conditions, and where different processes become observationally indistinguishable.

Controlled communication environments provide the calibration case. Because the true production process is specified in advance, observational access can be withheld, transformed, or contaminated systematically and the resulting inference can be evaluated against known structure. Related work applies the same logic to adversarial communication and to inference about whether particular actors participated in an otherwise hidden communication process.

Research: Inferring communication topology from message residue

Simulation environments

I build agent-based and computational simulations when a proposed mechanism is difficult to isolate in observational data or when the relevant counterfactual cannot be observed directly.

A simulation turns a verbal account into explicit rules governing what actors know, how they interact, what they retain or lose, how they make decisions, and how the environment changes. Those elements can then be varied independently to determine which assumptions are necessary for an observed pattern and which merely accompany it.

The models below use this logic to separate mechanisms that can produce similar visible outcomes, including hidden diffusion before adoption, correction during adaptive search, and local inference during residential sorting.

Cultural attractor measurement

Cultural attractor theory describes recurring structural configurations toward which transformed cultural variants tend to converge. The construct has been studied ethnographically, experimentally, and computationally, but usually without a measure operating at the level of individual variants within documented lineages.

This system treats a documented variant family as evidence from which to extract recurrent latent structure. The critical test is not whether visually similar variants resemble one another. It is whether a recurrent signature recovers functional relationships despite surface differences, predicts held-out variants, and fails where only superficial regularity is present.

The measurement logic is adapted from cheminformatics, where a pharmacophore is the configuration shared across compounds that bind a common target. The transfer is functional and statistical rather than geometric: what carries over is common-feature extraction and ranked-retrieval validation, not molecular geometry.

Human coding is part of the instrument rather than an external check. Reviewer identity is established before rating, confidence and flags are stored with each judgment, and disagreements route to adjudication rather than being averaged away.

Research: Measuring meme-family structure with contrastive sparse features

Org Signal

Org Signal is an organizational network analysis system for identifying relational patterns that conventional organizational charts and individual-level measures often miss.

Communication and organizational data are represented as networks so that structural properties can be measured directly before contextual interpretation is introduced. Here the central problem is not hidden-process reconstruction. It is separating structural measurement from organizational interpretation.

The separation between measurement and interpretation is intentional. Network statistics and structural configurations are computed directly from the relational data. Model-based reasoning enters later, where communication or organizational context is needed to evaluate competing interpretations of those patterns. Where the evidence does not distinguish among explanations, the system preserves that uncertainty rather than converting it into a single confident diagnosis.

Org Signal is a research and diagnostic prototype. It is not an automated personnel decision system.

Decision-support architecture

Some systems I build address a different problem: how to turn a loosely specified practical question into a recommendation without allowing missing information to disappear inside model-generated confidence.

ManuscriptU and the systems below combine structured problem definition, current external evidence, domain-specific data, deterministic computation, and model-based interpretation in different proportions.

The architecture varies by application, but the general sequence is:

Problem context

Identify what the user is actually trying to decide.

Constraints

Translate the description into explicit requirements, tradeoffs, and exclusion criteria.

Current evidence

Retrieve information that is time-sensitive or external to the model.

Domain structure

Apply specifications, taxonomies, known relationships, or other structured knowledge.

Computation

Calculate directly wherever the problem permits deterministic analysis.

Interpretation

Use model reasoning for comparison, synthesis, or judgment where those tasks cannot be reduced to direct calculation.

Recommendation

Return a decision or comparison while identifying sparse, conflicting, or unresolved evidence.

Validation depends on the system. It may involve comparison with known outcomes, source-coverage checks, consistency rules, expert judgment, or explicit confidence handling.

ManuscriptU

ManuscriptU is a research workflow system for evaluating journal fit and publication strategy.

Rather than treating journal selection as a keyword-matching problem, it combines information from a manuscript, its reference network, bibliographic databases, journal characteristics, and current external sources to assess where a paper fits intellectually and institutionally.

Retrieval and bibliographic relationships are handled deterministically where possible. Model-based synthesis is used to interpret fit, compare alternatives, and identify contradictions or missing evidence. The validation framework allows recommendations to be compared with known publication outcomes and reduces confidence when the relevant evidence is sparse or internally inconsistent.

Domain applications

The same architecture also applies across domains with very different constraint structures and evidence. These systems translate an underspecified practical question into a decision while preserving distinctions that generic recommendation systems tend to collapse.

What changes across applications is not the basic sequence, but the domain knowledge, evidence, constraints, and failure modes that matter to the recommendation.