Skip to content
Contact
☰
Color scheme: Dark.

Systems

Systems are the tools and objects of study with which EIL does its research: they represent states, reconstruct what is hidden, measure change or test distributed intelligence.

Open Observatory →

How systems, methods and metrics differ

Research system
A system built to investigate a research question. Its research purpose is stated. Example: SELFGRAPH
Agent and intelligence system
A system through which distributed intelligence is studied: specialized agents, their coordination and their failure behavior. Example: MIRROR
Infrastructure
Shared foundation that makes longitudinal (long-running) and distributed research possible. It is a prerequisite of the research. Example: Kernel
Method
A defined procedure by which research or analysis is carried out. Example: LIIE
Metric
A quantity or instrument that measures something over time. Example: CEI

Why EIL builds its own systems

EIL builds its own systems because the research questions often cannot be studied cleanly with existing standard tools.

This raises questions such as:

  • How does a human-AI system change over months or years?
  • How can states, drift, memory and recursion be made visible?
  • How are hidden states reconstructed from incomplete evidence?
  • How are emergent cognitive structures, perspective models or multi-agent intelligence studied?
  • How are observation, evidence, provenance and interpretation kept apart?

Generic tools are often not enough for this. For that reason, instruments, measurement systems, models, runtime structures and specialized systems arise from the research itself.

EIL builds its own systems when existing tools do not map a research question precisely enough. Recurring problems, observations and analysis requirements give rise to measurement, reconstruction, state and intelligence systems of its own, which then become part of the research themselves.

  1. Interaction
  2. Pattern / function
  3. Formalization
  4. System
  5. Reintegration
Reintegration leads back to interaction.

Origin

The systems arose within the research. Some of their functions trace back to recurring working practices, forms of analysis and structures that became visible in longitudinal human-AI interaction and were then formalized.

This makes the systems at once a result, an instrument and a new part of the research: externalized structures are technically represented, put to use and examined again in later interactions.

EIL studies which patterns arise in long-term interaction. EIL equally studies what happens when parts of this recurring practice are externalized and computationally represented.

The transitions from practice to externalization, from externalization to system and from the system back to practice are themselves objects of research.

As a development logic, EIL describes this sequence: research question, observation, phenomenon, reconstruction, formalization, system design, experiment, new observation. The sequence is a reading, not a protocol. Starting points can differ, for example an observation in practice, an open research question or a technical problem. Not every system arose exactly this way. Which sequence is documented for an individual system is stated with its origin, as far as the evidence reaches.

Observation, explanation and evidence remain separate steps. A reconstruction is an explanation and not yet a documented sequence of events. Observatory, SELFGRAPH and DINGIR are instruments of research and at the same time objects of testing, as far as the evidence supports it. The Intelligence Workbench is described as a concept; its effect has not been studied. The multi-agent systems are instruments whose coordination and failure behavior are themselves studied.

Examples: from observation to system

The assignments are candidate chains. They are reconstructed from dated evidence and are not yet confirmed findings. Individual recurring reasoning and interaction strategies were carried over in abstracted form into methods, metrics and system structures. A person and their thinking are not copied this way. Only an abstracted part was carried over.

  • PhenomenonObserver-system separation, Behavioral landscape formationMethodBehavioral Feature ExtractionSystemObservatory

    The separation of observer and system and the formation of a behavioral landscape were named in the interaction and then implemented in code (register: confirmed in code, with limitation).

  • PhenomenonRecursive correction loopsMethodUIPSystemKernel

    Repeated manual corrections of drift and fabrication were described as integrity rules. Candidate: guards in the kernel now enforce such rules in the system.

  • PhenomenonDetection of contradictory self-descriptionMethodAnomaly AnalysisSystemMIRROR

    Recognizing contradictory statements across long contexts was first done by hand. Candidate: MIRROR performs this check as a role of its own.

  • PhenomenonBoundary between reconstruction and fabricationMethodMulti-Hypothesis ReconstructionSystemDINGIR

    The boundary between reconstruction and fabrication appeared in real analysis tasks. Candidate: DINGIR keeps competing hypotheses side by side and marks open relations.

  • PhenomenonImplicit-to-reportable transitionMethodTemporal Pattern MiningSystemSELFGRAPH

    The transition from the implicit to the reportable shows how concepts and decisions change over time. Candidate: SELFGRAPH represents such trajectories and separates the observed from the inferred.

The research cycle behind it →

Shared technical foundation

The foundation of the systems on this page. It sits outside the research programs.

Eight shared primitives and the one-kernel rule

Concept

There is exactly one shared infrastructure layer for state, memory, events, evidence, provenance, identity, skills, guards and audit and replay. All research methods use that one layer. Frameworks and metrics stay versioned methodological instruments above that layer.

NameMeaning
EventA timestamped interaction, decision, mission, system event or observation.
StateAn explicit state snapshot with epistemic status and source.
EvidenceTraceable support or counterevidence linked to source material.
ProvenanceSource, time, transformation path, version of algorithm or instrument, and human override.
Instrument versionThe version of a metric or protocol, so that a change of measurement can be told apart from a change of person or system.
OutcomeObservable downstream effect of an intervention, an interaction or a system action.
InterventionA deliberate change by human, AI or system that could influence outcomes.
Epistemic statusObserved, derived, inferred, hypothesized, supported, contested, rejected or unknown.

When the kernel raises a drift alert

The kernel detects drift and surfaces it through monitoring and feedback. It distinguishes four triggers:

  1. a tool error
  2. reaching an iteration limit
  3. a triggered rejection
  4. a divergence between a claimed and an actually executed action

ImplementedApplies to both components.

Research systems

Experimental tools that make the research observable, reconstructible and computationally testable.

  • Measurement environment

    Observatory

    Prototype

    The Observatory makes long-term changes in human–AI interaction, emergent patterns and technical system states open to study.

    Open Observatory

    Observatory

    Prototype

    Research system / Measurement environment · Research system · Observation

    Observation environment for human, AI, interaction, emergence, system state and neuroadaptive signals.

    In shortThe Observatory makes long-term changes in human–AI interaction, emergent patterns and technical system states open to study.

    What it investigatesHow humans, AI, interaction and system state change over time, and whether a pattern is more than a chance swing.

    Research purposeMakes longitudinal interaction data readable across six measurement dimensions, so that every visualization answers a concrete research question.

    RoleLongitudinal meta-observation. It is meant to hold events, baselines, state, emergence lifecycle, evidence, provenance and replay.

    Does not becomeA generic dashboard or a personal life graph.

    How EIL uses itMakes longitudinal interaction data readable. Every view is tied to at least one research question.

    Connected research spacesHuman–AI Interaction & Co-Evolution, Emergence, Adaptation & Complex Dynamics, Measurement, Evidence & Research Infrastructure, Human–Machine Interfaces & Neuroadaptive Systems, Human Cognition, Behavior & Externalization

    OriginDerived from longitudinal interaction research: questions of trajectory, drift and recurrence arose in multi-year human-AI interaction. Observer-system separation and behavioral landscape were named there and then implemented.

    Development statePrototype. Individual views run on real trajectory data, the full scope is not yet implemented.

    StatusIn development. Individual views run on real trajectory data; the full range of dimensions is not yet implemented.

    Public limitsShows only what was actually logged. Raw data stays private, aggregate views are public.

    Public limitsShows only what was actually logged. Raw data stays private; only aggregate views are public.

    Supported research programs

    Research questions (4)
    • How does repeated interaction change both the behavior and the structure through which later interaction takes place?
    • Does a human–AI interaction improve, preserve or deteriorate over months and years?
    • When does a single signal become a pattern, and when does a pattern become a stable structure with a functional effect?
    • Which signal types can meaningfully be linked to interaction trajectories, and how would they be calibrated against established behavioral measures?

    The problemWithout a shared observation layer, longitudinal research stays tied to single sessions and one-off analyses, and a pattern cannot be told apart from a chance swing.

    Architecture
    • Six measurement dimensions: human, AI, interaction, emergence, system, neuroadaptive.
    • Every visualization is bound to at least one research question.
    • Current and future neuroadaptive channels are kept apart.
    Spec
    • Role in research: Measurement instrument for longitudinal human–AI interaction. The separation of observer and system is also a subject of study: it was named in the interaction and implemented in code, and whether it holds is open.
    • What can be observed and tested: Trajectories of the human, AI, interaction, emergence and system dimensions across sessions and time, with neuroadaptive channels kept separate. The test question is whether a pattern is more than a chance swing.
    • How results lead to new questions: A pattern in one view raises the question of whether it persists across further sessions and longer periods. In this way an observation becomes a test.

    Open Observatory →

  • Longitudinal representation

    SELFGRAPH

    Prototype

    SELFGRAPH represents how states and patterns change over long periods and keeps what was observed apart from what was inferred.

    SELFGRAPH

    Prototype

    Research system / Longitudinal representation

    Longitudinal computational representation of observed state, inferred state, temporal change, genealogy and pattern evolution.

    In shortSELFGRAPH represents how states and patterns change over long periods and keeps what was observed apart from what was inferred.

    What it investigatesHow state, identity and continuity can be represented over long periods, with uncertainty and evidence attached to each statement.

    Research purposeStudies how state, identity and continuity can be represented across long timeframes, with uncertainty, evidence and externalization as first-class parts. Governing rule: OBSERVED ≠ INFERRED. SELFGRAPH follows temporal structure in place of a static profile. It is a bridge between EIL's longitudinal Human–AI research and future neuroadaptive systems.

    RoleResearch instrument for longitudinal personal intelligence. It is meant to hold projects, ideas, decisions, open loops, states, concept evolution, evidence and outcomes.

    Does not becomeA world model or a clinical and diagnostic engine.

    How EIL uses itServes as the temporal representation layer of the longitudinal human–AI research.

    Connected research spacesHuman–AI Interaction & Co-Evolution, Emergence, Adaptation & Complex Dynamics, Human Cognition, Behavior & Externalization, Distributed, Agentic & Collective Intelligence, Computational Interaction & Cognitive Modeling, Measurement, Evidence & Research Infrastructure

    OriginBuilt as a research instrument from the question of which reasoning and working operations appear and change when in long interaction. The origin in detail is being reconstructed from dated evidence.

    Development statePrototype. Phase 1 is implemented and tested at its core. Multi-year trajectories are a research goal. No result exists yet.

    StatusPhase 1 core implemented and tested. Multi-year trajectories are a research goal. No result exists yet.

    Public limitsNo clinical claims. Person-level graphs are private.

    Public limitsMakes no clinical claims and derives no psychological interpretation from telemetry alone. Inferred state stays revisable and is never presented as observed. Person-level graphs are private.

    Supported research programs

    Research questions (4)
    • How are they represented: as method, model, memory structure, agent or computational system?
    • How do they change after externalization?
    • How stable are they?
    • Does a human–AI interaction improve, preserve or deteriorate over months and years?

    The problemWithout an evolution graph, a system loses touch with earlier states, terms and open threads in every session, and continuity has to be re-guessed each time.

    Architecture
    • Typed nodes and edges with temporal reference, confidence and provenance.
    • Observed and inferred state are marked separately throughout.
    • Genealogy: where a term or pattern came from and how it changed.
    • Uncertainty and evidence attach to each individual statement.
    Spec
    • Governing rule: OBSERVED ≠ INFERRED
    • Represents: Observed state, inferred state, temporal change, genealogy, pattern evolution, uncertainty, evidence, externalization, identity, state continuity
    • Role in research: Instrument for the temporal representation of the longitudinal human–AI research. Also a subject of study: which representation can carry continuity and identity over long periods at all is open.
    • What can be observed and tested: How concepts, decisions and patterns change over time, and whether inferred state holds up against observed state. The related origin chain is a candidate chain, not a confirmed finding.
    • How results lead to new questions: Inferred changes remain revisable. A question put to the trajectory thereby becomes a test, not a finding.

    Observed Inferred

    Observed
    Directly present in the data.
    Inferred
    A plausible interpretation that goes beyond what was observed.

  • World-model reconstruction

    DINGIR

    Active research

    DINGIR reconstructs a hidden state and its possible causal structure from incomplete evidence.

    DINGIR

    Active research

    Research system / World-model reconstruction · Research system · World state

    Collaborative, experimental world-model and reconstruction system for hidden state, causal chains and state transitions.

    In shortDINGIR reconstructs a hidden state and its possible causal structure from incomplete evidence.

    What it investigatesHow world states, causal chains and state transitions can be reconstructed from incomplete traces.

    Research purposeReconstructs a hidden world state and its causal structure from incomplete evidence, keeps competing hypotheses in parallel and marks open relations. DINGIR is the only system on this site that models a world state.

    RoleResearch on world model and external state. It is meant to hold entities, relations, events, observations, hidden states, unresolved edges and causal hypotheses.

    Does not becomeSELFGRAPH or a generic EIL module.

    How EIL uses itKeeps competing hypotheses side by side until evidence separates them, and marks open relations.

    Connected research spacesReconstruction, Perspective & World Models, Human Cognition, Behavior & Externalization

    OriginPuts questions from the Perspective Modeling & Cognitive Reconstruction program into practice: reconstruction from incomplete traces without mixing interpretation and observation. Origin: jointly with RFI-IRFOS.

    Development stateActive research. No published validation result exists.

    Public limitsReconstructions are hypotheses with a status. They are not asserted as facts.

    Public limitsReconstructions are hypotheses carrying a status (observed, inferred, hypothesis, simulated, derived, unknown, contradicted), not factual claims. Unresolved relations stay open. World model and actor model remain separate levels.

    Supported research programs

    Research questions (3)
    • Which causal chain and which counterfactual paths best explain a state transition?
    • What structure, state or relationship most completely explains the available traces when direct observation is not possible?
    • What information was available to an actor, and which assumptions, relationships and goals are probable given it?

    The problemA state that cannot be observed directly stays a guess without reconstruction.

    Architecture
    • Chain: evidence, hidden state, causal structure, transition, possible futures.
    • Competing hypotheses stay in parallel until evidence separates them.
    Spec
    • Role in research: Instrument of reconstruction research. Its outputs are hypotheses with a status and a starting point for testing, not findings. The boundary between reconstruction and fabrication is also a subject of study. By current evidence it appeared in real analysis tasks, as a candidate and with a reservation.
    • What can be observed and tested: Whether competing hypotheses are separated by new evidence and which relations remain open. The related origin chain is held as a candidate.
    • How results lead to new questions: Open relations show where evidence is still missing. Targeted questions for the data can be derived from them.

  • Working environment

    Intelligence Workbench

    Concept

    The Intelligence Workbench is a working environment in which research objects, evidence and decisions stay connected and inspectable.

    Intelligence Workbench

    Concept

    Research system / Working environment · Research system · Workspace

    A research environment for working directly with structured intelligence objects. The goal is to make complex reasoning persistent, inspectable and operational.

    In shortThe Intelligence Workbench is a working environment in which research objects, evidence and decisions stay connected and inspectable.

    What it investigatesHow complex reasoning can become persistent, inspectable and operational.

    Research purposeThe architecture is object-centric: projects, research objects, evidence, relationships, agents, missions, artifacts, history and provenance are the units of work. Research decisions stay connected to context, supporting evidence, counterevidence and assumptions, so it stays traceable later what a decision was based on.

    RoleWorking surface for research, building and papers. It is meant to hold objects, projects, relations, missions, artifacts, evidence, search and history.

    Does not becomeThe Observatory, SELFGRAPH or a second runtime.

    How EIL uses itThe units of work are research objects, evidence, relations and provenance. Counterevidence and assumptions stay attached to each decision.

    Connected research spacesHuman Cognition, Behavior & Externalization, Distributed, Agentic & Collective Intelligence, Computational Interaction & Cognitive Modeling, Emergence, Adaptation & Complex Dynamics

    Development stateConcept. So far only the description is public.

    StatusConcept. The Workbench is the object-centric research environment of EIL. Its public description stays at this concept level.

    Public limitsInternal working environment.

    Public limitsInternal working environment. It is described publicly at the level of the concept.

    Supported research programs

    Research questions (2)
    • What cognitive functions externalize?
    • Which collective phenomena arise unplanned, such as spontaneous specialization or hidden centralization?

    The problemResearch decisions often cannot be traced back to their basis later. Context, supporting evidence and assumptions get lost between sessions and tools.

    Spec
    • Role in research: Working environment in which research decisions stay connected to supporting evidence, counterevidence and assumptions. It is described as a concept and not yet a running tool. Its effect on research decisions has not been studied.
    • What can be observed and tested: Whether a decision can later be traced back to its evidence and assumptions. No results exist yet.
Intelligence systems

Multi-agent systems

EIL develops multi-agent systems as distributed intelligence architectures. Specialized artificial intelligences take on different functions of thinking, analysis, review, reconstruction, security and execution, and are joined into one shared system through roles, handoffs, state, evidence and runtime structures.

The research examines how intelligence is distributed functionally, how specialized units take on different perspectives and tasks, check one another, pass on evidence, detect contradictions and coordinate decisions. The central question is when capabilities at the level of the whole system arise from this specialization.

What comes from where stays separate. The capability of a single agent comes from the underlying language model. Roles, handoffs, review steps and the holding of state belong to the system architecture. An example: when one agent reviews the output of another, the review is a decision of the architecture, while its quality depends on the model.

The current architecture includes roughly 15 specialized lead agents and a pool of 292 specialized sub-agents across research and operational contexts. The eleven entries below show a selection of them. The number of agents describes the build state. It does not count as a research result.

How the agents are built (lead agents, sub-agents, roles)

Specialized lead agents and sub-agents.

Specialized lead agents and sub-agents operate through designed roles, handoffs (transfers between agents), state, evidence and runtime structures. What is studied is how several specialized agents work together and where that breaks down.

A single model in one chat window holds no shared state, no handoffs and no audit trail across multiple steps. Once a task needs several specialized perspectives, parallel research or continuity across days or weeks, that model breaks down. The lead/sub-agent architecture keeps explicit track of what each agent knows, who hands off to whom, and what happens when a role fails or drifts.

What is studied are these dimensions: Specialization, Orchestration, Coordination, State, Failure, Repair, Authority, Distributed cognition.

NameMeaning
SpecializationHow tasks divide among components with different roles.
OrchestrationWho sequences the work and how much overview a single component holds while doing so.
CoordinationHow components align their actions without everything passing through one point.
StateWhat each component knows and how that knowledge is kept beyond a single run.
FailureHow breakdowns show up and whether they are noticed at all.
RepairHow a system recovers and who initiates the recovery.
AuthorityWho may decide, override or stop, and how that is recorded.
Distributed cognitionWhether a result belongs to the group of components and to no single one of them.

The roles of the agents partly arose from observing recurring working operations, such as reviewing, contradiction detection, decomposing and synthesizing. They are studied as externalizations of working functions. A functional correspondence is claimed. A biological correspondence is not claimed.

Lead Agents · Specialists · Sub-Agents · Shared Context · Runtime / Evidence

JARVIS

Experimental

Intelligence system / Orchestration · Distributed intelligence · Orchestration

An experimental research-agent interface for bounded research workflows. It supports interaction with research material, previous conversations, notes, simulations, tools and structured research signals. The system is instrumented so that agent interaction itself can become research data.

In shortJARVIS coordinates several specialized agents in bounded research workflows and holds the overall status.

FocusOrchestration, Coordination

What it investigatesServes as a study object for orchestration and coordination: who hands off to whom, who holds the state, and where hidden centralization arises when one orchestrator sees everything. Test question: does a silent failure of an agent become visible in the shared status?

Development stateRuns as a research setup. Claims about the effectiveness of the orchestration are not yet evidenced.

Public limitsThe agents check and report; they do not act on their own in outside systems. Internal configuration is not public.

Supported research programs

Research questions (2)
  • How does distributed intelligence alter responsibility, authority and failure?
  • Which collective phenomena arise unplanned, such as spontaneous specialization or hidden centralization?

The problemWithout central orchestration, every agent has to know on its own when it's its turn and who to hand off to. Past a handful of agents that stops being coordinable, and a silent failure goes unnoticed.

Architecture
  • Distributes subtasks to the matching specialized agents.
  • Gathers intermediate results and tracks overall status.
  • Detects when an agent stops delivering or gets stuck.
  • Keeps track of what each agent knows and who hands off to whom.
Spec
  • Role: Multi-agent orchestrator
  • Input: Subtasks from a workflow
  • Output: Consolidated status + result

MIRROR

Experimental

Intelligence system / Failure and repair · Distributed intelligence · Failure and repair

Detects contradiction, drift, regression and divergence from previously validated system states.

In shortMIRROR detects contradiction, drift and the silent loss of a previously validated system state.

FocusFailure, Repair

What it investigatesStudies failure and repair in agent groups: a run can be technically error-free and substantively wrong. MIRROR therefore looks for the silent loss of a previously reached correct state. To do so, MIRROR compares the current state of a run with the last validated state and formulates counter-hypotheses.

Connected research spacesDistributed, Agentic & Collective Intelligence, Emergence, Adaptation & Complex Dynamics, Human–AI Interaction & Co-Evolution, Measurement, Evidence & Research Infrastructure

OriginCandidate for an externalized working function: contradiction detection that was earlier done by hand. Meant functionally, with no correspondence to a biological module.

Development stateRuns as a research setup. Detection rates are not published.

Public limitsA finding is a pointer to a deviation. It does not count as proof of an error. No public false-alarm or hit statistics exist.

Supported research programs

Research questions (2)
  • Which collective phenomena arise unplanned, such as spontaneous specialization or hidden centralization?
  • How can a measured trajectory be checked as genuine and not the product of instrument drift or a false state?

The problemA team can believe for weeks it's on plan while reality has already moved on: nobody reports that on its own.

Architecture
  • Compares plan versus actual state, rule compliance, targets versus outcomes.
  • Monitors team velocity versus commitments.
  • Detects quiet workarounds, hollowed-out metrics and overlooked deadlines.
  • Delivers an early warning ahead of any post-mortem, with a running integrity score.
  • A run can be technically error-free and still substantively wrong; regression is therefore identified through triggered error messages and through the silent loss of a previously reached correct state.

JANUS

Experimental

Intelligence system / Schema and provenance · Distributed intelligence · Schema and provenance

Guard for schema and provenance: checks whether handoffs between agents match the expected format and a traceable source.

In shortJANUS checks whether handoffs between agents match the expected format and come from a traceable source.

FocusAuthority, Coordination

What it investigatesStudies authority and coordination at the handoff points: how an agent knows an input is valid and where it came from when no human sits in between.

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsPublic detail is intentionally limited.

Supported research programs

Research questions (1)
  • How does distributed intelligence alter responsibility, authority and failure?

The problemHow does an agent know an input is valid and who it came from when no human sits in between? Public detail is intentionally limited.

DAEDALUS

Experimental

Intelligence system / Evidence and logic · Distributed intelligence · Evidence and logic

Maintains an evidence matrix and checks chains of inference before a finding is passed on.

In shortDAEDALUS maintains an evidence matrix and checks chains of inference before a finding is passed on.

FocusDistributed cognition, Evidence

What it investigatesStudies distributed cognition: when several agents supply partial findings, a checkpoint decides whether the chain of inference holds. This makes visible where a group result rests on an unchecked assumption. Test question: would the group reach the same result without the checkpoint?

Connected research spacesDistributed, Agentic & Collective Intelligence, Emergence, Adaptation & Complex Dynamics, Reconstruction, Perspective & World Models

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsPublic detail is intentionally limited. The check does not replace expert review.

Supported research programs

Research questions (2)
  • Which collective phenomena arise unplanned, such as spontaneous specialization or hidden centralization?
  • Which causal chain and which counterfactual paths best explain a state transition?

The problemWhen several agents supply partial findings, a checkpoint decides whether the chain of inference holds. This shows where a group result rests on an unchecked assumption.

DELTA

Experimental

Intelligence system / Role formation · Distributed intelligence · Specialized agent

Specialized member of the agent group on which role formation and delegation are studied; public detail is intentionally limited.

In shortDELTA is a specialized member of the agent group on which role formation and delegation are studied.

FocusSpecialization, Authority

What it investigatesServes as a further specialized member of the group on which role formation and delegation are studied: which tasks end up with a specialist, and how that changes the distribution of authority.

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsThe role is not described in public detail.

Supported research programs

Research questions (1)
  • How does distributed intelligence alter responsibility, authority and failure?

The problemWhich tasks end up with a specialist, and how does that change the distribution of authority? The role is not described in public detail.

NYX

Experimental

Intelligence system / Adversarial simulation · Distributed intelligence · Adversarial simulation

Simulates attacks on agent groups and checks attack surface, dependencies, configuration and exposed credentials.

In shortNYX simulates attacks on agent groups in its own test environments and checks attack surface and configuration.

FocusFailure, Repair

What it investigatesStudies how an agent group fails and recovers under attack: which agent can be manipulated, and whether the group result flips unnoticed as a result.

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsRuns exclusively against own, purpose-built test environments. Findings are ranked by severity. A finding does not count as proof of exploitability.

Supported research programs

Research questions (2)
  • How does distributed intelligence alter responsibility, authority and failure?
  • Which collective phenomena arise unplanned, such as spontaneous specialization or hidden centralization?

The problemSecurity gaps often go undetected until they are exploited: without continuous scanning, you rely on luck or rare manual audits.

Architecture
  • Monitors open ports, exposed services and TLS certificate health.
  • Matches dependencies against newly published CVEs.
  • Detects misconfigured access rights and unused accounts still holding live access.
  • Ranks every finding by severity, CVSS-style.
Spec
  • Role: Security & vulnerability scanning
  • Output: Prioritized findings list, ready to hand to a dev team

ROBERT

Experimental

Intelligence system / Specialization and delegation · Distributed intelligence · Specialization

Researches and qualifies leads and assembles context for outreach, coordinated across several sub-roles.

In shortROBERT researches and qualifies possible contacts from public signals; a human decides who is approached.

FocusSpecialization, Authority

What it investigatesServes as an example of specialization and delegation in a clearly bounded task field: sub-roles gather signals, a human decides who is contacted. The study concerns where the boundary of delegation lies.

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsDoes not replace human decisions. Uses only publicly available signals. An effect on response rates is not evidenced.

Supported research programs

Research questions (1)
  • How does distributed intelligence alter responsibility, authority and failure?

The problemA first outreach message without real context reads generic and gets ignored, and doing research on every lead by hand doesn't scale.

Architecture
  • Gathers publicly available signals about a company: size, sector, current activity, relevant contacts.
  • Turns them into something a first outreach message can actually be specific about.
  • Coordinated across several specialized sub-roles.
  • Doesn't replace the human decision of who to reach out to, but supplies the basis for it.

ARGUS

Experimental

Intelligence system / Continuous monitoring · Distributed intelligence · Continuous monitoring

Continuously monitors market, press and competitive signals and reports changes with history and source.

In shortARGUS continuously monitors market, press and competitive signals and reports changes with history and source.

FocusState, Continuity

What it investigatesStudies state and continuity in a permanently running agent: what it has to retain over days and weeks so that a report stays classifiable against earlier reports.

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsReports rest on public sources and can be incomplete. No hit rate is published.

Supported research programs

Research questions (1)
  • Which collective phenomena arise unplanned, such as spontaneous specialization or hidden centralization?

The problemCompetitive moves often only become visible over weeks, by which point they already have consequences: a weekly digest is structurally too late.

Architecture
  • Monitors competitor announcements, mentions, press, sentiment shifts, leadership changes and product launches.
  • Detects emerging market movements, narratives forming before mainstream coverage and early signs of a market shift.
  • Delivers a same-day alert, with history and sourcing.

ATLAS

Experimental

Intelligence system / Prioritization · Distributed intelligence · Prioritization

Prioritizes and orchestrates signals across multiple specialized agents.

In shortATLAS merges reports from several agents into one prioritized order.

FocusOrchestration, Coordination

What it investigatesServes as a study object for orchestration and coordination at the level of results: how reports from several agents are merged into one ranking, and where contradictions between agents go unnoticed.

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsWhat is studied is the merging of agent reports as a research object. No evaluation of the prioritization has been published.

Supported research programs

Research questions (2)
  • How does distributed intelligence alter responsibility, authority and failure?
  • Which collective phenomena arise unplanned, such as spontaneous specialization or hidden centralization?

The problemTen agents deliver ten separate alerts: without prioritization, nobody knows what to act on first or where findings conflict.

Architecture
  • Monitors what every agent is finding, which alerts overlap and where agents depend on each other.
  • Detects conflicting reports, duplicate alerts and bottlenecks between agents.
  • Condenses ten separate alerts into one ranked list, with a clear next action.

KOPERNIKUS

Experimental

Intelligence system / Specialization · Distributed intelligence · Specialized agent

Identifies funding and capital sources that match the system under study.

In shortKOPERNIKUS identifies funding and capital sources that match the system under study.

FocusSpecialization

What it investigatesServes as a specialized agent on which specialization is studied: how narrowly a role can be defined, and which handoffs it needs so that other agents can reuse its result.

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsThe agent checks and reports; it files no applications.

Supported research programs

Research questions (1)
  • How does distributed intelligence alter responsibility, authority and failure?

The problemGrant databases list thousands of programs, most of which don't actually apply to the case at hand, and sifting through them by hand costs more time than it saves.

Architecture
  • Monitors open grant programs, investor activity, new funding rounds and application deadlines.
  • Detects genuinely matching programs, changing eligibility rules and closing deadlines.
  • Delivers a short list of real options, ranked by fit and deadline.

LYNX

Experimental

Intelligence system / Specialization and failure · Distributed intelligence · Specialized agent

Checks whether and how an organization can be found online.

In shortLYNX checks whether and how an organization can be found online.

FocusSpecialization, Failure

What it investigatesServes as a specialized agent on which specialization and failure behavior are studied: how a narrowly scoped checking agent detects a gradual degradation, and who notices when it reports nothing.

Development stateThe system runs as a research setup and is being tested under real or controlled conditions.

Public limitsThe agent checks and reports; it changes no websites.

Supported research programs

Research questions (1)
  • How does distributed intelligence alter responsibility, authority and failure?

The problemRankings and visibility often degrade long before the traffic loss becomes obvious: without ongoing monitoring, it's noticed only after the damage is done.

Architecture
  • Monitors Google rankings, keyword performance and technical SEO signals like Core Web Vitals and broken links.
  • Checks how a brand shows up in AI search results and in classic search.
  • Detects pages competing against each other for the same keyword, and content search engines can't properly index.
  • Delivers a prioritized fix list.
Spec
  • Role: Visibility & SEO intelligence
  • Output: Technical SEO audit + prioritized fix list

Connected research spaces Distributed, Agentic & Collective Intelligence, Emergence, Adaptation & Complex Dynamics (applies to all entries except MIRROR, DAEDALUS, whose spaces are named in the detail window)

All agents here are experimental. The descriptions are not evaluated evidence of effect. The agents are research objects and not products, and no published evaluation exists for any of them.

System state, memory, security & integrity

The remaining components of the shared foundation, grouped by task. They sit outside the research programs.

ImplementedApplies to all components except Replay and Recovery.

State & continuity

Evidence & provenance

Security & control

Restoration

Status on this page

  • ConceptPublicly described, but not yet implemented as a running system.
  • PrototypeAn internal version exists and can already be used or tested.
  • Active researchThe system or approach is currently being investigated; no published validation result exists yet.
  • ExperimentalThe system runs as a research setup and is being tested under real or controlled conditions.
  • ImplementedTechnically implemented in research operation and actually in use.