Skip to content
Contact
☰
Color scheme: Dark.
Back to methods

Study design and evidence

How EIL collects, orders and protects evidence. The sections show how the material is built, how it is coded, how a path from practice to representation is evidenced, which confounds apply, how data are protected, where the limits lie and which assumptions and evidence there are.

Study design and phases

One person, several years, every event dated

The material comes from a single person, starting in 2023. This is called N-of-1: one person as the whole study. The material says nothing about other people. In return, the material can be followed in time order from the beginning. Every statement is tied back to its source in the material, and counterexamples stay in the register.

The role conflict

The same person is participant, researcher and developer of the systems. That is a conflict of interest, and we name it. Three rules counter it. The third is still planned.

  • Retrospective interpretation and contemporaneous evidence are kept apart.
  • AI interpretations never count as independent confirmation.
  • A second independent person codes a sample. Planned, once that is feasible. Planned

Causal language

Whether sustained human–AI interaction leads to externalization is a target to test (the wording is on the Research page). For this design it means: phases 1 and 2 work retrospectively, phase 3 observes prospectively, phase 4 compares tasks and periods with and without a structure. Stronger claims need prospective comparison. To the Research page →

Four phases

  1. 01

    Historical reconstruction

    Active research
  2. 02

    Retrospective check

    Planned
  3. 03

    Prospective observation

    Planned
  4. 04

    Intervention and comparison

    Planned

Four phases

  1. 01

    Historical reconstruction

    Active research

    The earliest evidence from 2023 onward is secured, practices are found before they were named, and a timeline of externalizations and of changes to models and tools is built.

  2. 02

    Retrospective check

    Planned

    Selected periods are coded fully or partly blind, that is, by coders who are not shown later artifacts (for example the systems) while early interactions are coded. The test is whether they find recurring ways of working (practice families) without knowing the later system names.

  3. 03

    Prospective observation

    Planned

    The codebook is frozen on a fixed date. New externalizations are logged as they happen, the expected effect is written down before the outcome, and non-effects and rollbacks are tracked.

  4. 04

    Intervention and comparison

    Planned

    Tasks with and without a given externalized structure are compared, and so are periods before and after its introduction. Where it is safe, a structure is switched off for a time to test dependence.

Planned sub-studies

Minimum and full demonstration

Planned

What a first empirical study is meant to show, and what a stronger full program would show. Both are planned and not yet reached.

Minimum and full demonstration

Planned

Minimum demonstration

  1. Select five to ten practice candidates from different families.
  2. Recover pre-formalization evidence for each.
  3. Show recurrence across time and contexts.
  4. Show explicit formalization.
  5. Show computational representation at CR3 or higher, preferably CR4 to CR6.
  6. Show real reintegration into later work.
  7. Measure at least one downstream behavioral or workflow change.
  8. Show at least one second-generation externalization for the strongest cases.
  9. Include failed and abandoned chains.
  10. Analyze model and tool confounds.

Full demonstration

  • Ten to twenty validated transformation chains.
  • Several practice families and several representation types.
  • Several chains with recursive depth above one.
  • Prospective confirmation of patterns first found retrospectively.
  • Independent or partially independent coding.
  • Explicitly stated non-effects and negative cases.
  • Triangulation across several component studies.

Agents as externalized function

Planned

The agent architecture is studied functionally: which recurring human operation was formalized as an artificial role, how were roles recombined, and what changed measurably. The evidence for this is only partly collected.

Agents as externalized function

Planned

Required evidence (10)

  1. Original interaction examples in which the person performed the operation before it was formalized.
  2. Date and rationale for the creation of each role or skill.
  3. Role specification and allowed authority.
  4. Mission traces from real work before and after specialization.
  5. Handoff and failure cases, retries and repairs.
  6. Integrity conflicts and their repair paths.
  7. Human overrides.
  8. Outcome measures per role, such as speed, coverage, error rate, contradiction detection or completion.
  9. Control for stronger base models and for changes of provider and tooling.
  10. A check whether reintegration changes how the person delegates, checks or structures later work.

Hypothesis family: recurring cognitive work practices can be externalized step by step into explicit artificial roles and technical constraints. Once reintegrated into the workflow, these roles may change the person's own work practice. Agents count here as externalized functions. Whether they reproduce biological cognition is not claimed.

Automation history

Planned

A planned study reconstructs from timestamped histories which automation practices already existed before or alongside later product features. It relies on evidence from the material. Memory does not count as evidence. No result exists yet.

Automation history

Planned

Six levels of classification

NameMeaning
1ProposedOnly an idea or a desired workflow.
2Manual, structuredA repeated procedure with explicit steps but no automatic trigger.
3Semi-automatedSome actions run automatically, and the person initiates or bridges key steps.
4AutomatedTrigger, execution and result with limited intervention.
5AgenticGoal-driven decomposition, state, feedback, tool use or adaptive routing.
6PersistentSurvives a session and a restart through stored state, a schedule or a durable workflow.

Search window: systematically 2025 to March 2026, backward into 2024 for precursors. Later evidence from 2026 shows continuation or formalization and does not rewrite earlier history. Comparisons with releases of other providers are dated and factual and follow only after current sources are checked.

Neuroadaptive preparation

Planned

The neuroadaptive direction is prepared today at the level of data contracts and stays empirically separate until synchronized physiological or neural measurements actually exist. No such data is collected, and there are no results.

Neuroadaptive preparation

Planned

Planned pipeline (9 nodes)

  1. EEG, BCI and wearables
  2. Time synchronization
  3. Signal quality
  4. Artifact handling
  5. Feature extraction
  6. State estimation
  7. Bounded adaptation
  8. System intervention
  9. Outcome evaluation

Ten preparation rules

  1. A unified frame of timestamps and events across chat, agents, missions and later sensors is prepared.
  2. Every intervention of the system is recorded: what changed and when.
  3. Personal baselines replace assumptions about populations.
  4. Signal quality and artifacts are carried as metadata.
  5. EEG, BCI and wearables count as evidence streams and never as truth.
  6. Outcomes are evaluated independently of the signal that triggered the adaptation.
  7. Adaptation stays bounded and has an explicit non-adaptive fallback.
  8. Consent and privacy for highly sensitive biometric data are designed before any collection.
  9. State estimates from interaction and physiological evidence are compared without treating either as ground truth.
  10. Today's conversational states are not retroactively labeled as neural states.

Five open research questions

  • Do state estimates derived from interaction correlate with synchronized physiological changes?
  • Can state-aware changes in information density, modality or agent strategy improve task outcomes?
  • When does adaptation help, and when does it destabilize agency or performance?
  • Are longitudinal personal baselines more informative than generic state classifiers?
  • Can the same transformation chain of a cognitive practice be detected in language and behavior and in physiological evidence?

The central question is on the page Research.

Coding and coding rules

Codebook and coding rules

Planned

Eight dimensions for coding observable behavior. The codebook is frozen before the prospective phase.

Codebook and coding rules

Planned
NameExample values
Task stanceexploratory, execution, review, reconstruction, planning, reflection
Information operationcollect, compare, infer, contradict, synthesize, prioritize, discard
Uncertainty actionask, defer, branch hypotheses, verify, abstain, overreach
Control actionset constraint, define role, create acceptance gate, revise rule
Memory actionrecall, persist, link, reopen, close loop
Externalization actionname, formalize, schema, code, agentize, instrument
Review actionself-check, adversarial check, counterexample, source audit
Outcomeaccepted, revised, abandoned, failed, externalized, reintegrated

In the first pass only observable behavior is coded, before later framework terms are applied. Otherwise the analysis finds only what the researcher already expects.

Coding protocol in ten steps

  1. Create a frozen codebook of cognitive practice.
  2. Code only directly observable behavior in the first pass.
  3. Code candidate cognitive operations in the second pass.
  4. Link repeated events into candidate practice families in the third pass.
  5. Search forward for events of naming and formalization.
  6. Search forward again for computational representation.
  7. Search forward again for real reintegration.
  8. Record downstream change and non-change.
  9. Record counterexamples and abandoned representations.
  10. Only then assign an interpretation at chain level.

Rules for coders and AI assistance (8)

  • Model-assisted coding may propose labels and is never the only validator of its own hypotheses.
  • A human-reviewed benchmark subset is frozen before model-assisted classification.
  • Agreement is calculated on the operational codes before any higher-level interpretation.
  • Disagreement is preserved as data.
  • Several competing interpretations of the same event are allowed.
  • Contradictions and negative examples are annotated alongside confirming examples.
  • The codebook is versioned, and categories are never silently redefined.
  • Literal evidence and interpretation are kept apart.

Analytical options

Planned

Qualitative process tracing and quantitative longitudinal analysis can be combined. Which procedures are used depends on data quality and event density. Sequence analysis and graph analysis belong here too and appear as exploratory methods under “Methods”.

Analytical options

Planned
NameUsed for
Process tracingFollowing complete transformation chains step by step.
Time to externalizationThe interval from first recurrence to formalization.
Pre/post comparisonComparison around the adoption of computational representations.
Interrupted time seriesWhere event density and stability allow it.
Survival and persistence measuresHow long practices and representations persist.
Transfer analysisAppearance across domains and tasks.
Reliability analysisReliability of the coding.
Ablation or temporary disablingComparison with selected system components switched on and off.
Qualitative negative-case analysisAnalysis of failed chains.

Eleven measures

Concept

Every measure carries its interpretation limit. None is reported yet.

Eleven measures

Concept
NameObservableInterpretation limit
RecurrenceFrequency across sessions, contexts and time.Frequency alone does not establish emergence.
Context diversityNumber and kinds of tasks and domains.More contexts strengthen stability claims.
Externalization depthHighest CR level reached.Depth does not prove usefulness.
Reintegration intensityReuse frequency, workflow centrality, dependent tasks.Use does not prove cognitive change.
Practice changeCoded before/after differences.Tool, model and task changes must be controlled.
TransferAppearance in new tasks and contexts.Transfer can reflect expertise as well as the AI effect.
Recursive depthNumber of externalization and reintegration generations.Requires end-to-end provenance.
Human agencyOverride, rejection, revision, rollback.High automation is not automatically better.
Error and qualityTask-specific defect, revision and completeness measures.Must be defined per domain.
Provenance completenessShare of chain links with source evidence.Good documentation evidences provenance. It says nothing about the phenomenon itself.
Counterevidence densityNumber and quality of retained negative cases.Helps reduce confirmation bias.

The longitudinal evaluation protocol across studies is the method LIIE, with its nine dimensions under the methodological architectures.

From practice to representation

Three scales describe the same path. The ten levels of externalization describe the form of the practice, the maturity CR0 to CR7 (CR stands for computational representation) the representation, and the chain T0 to T8 the evidence a single chain needs.

Terms

Concept

Seven key terms with the minimum evidence each one needs.

Terms

Concept
NameDefinitionMinimum evidence
Persistent interactionInteraction across time, tasks and contexts with enough continuity to reconstruct development.Timestamped longitudinal corpus with a recoverable sequence.
Cognitive practiceAn observable, recurring way of representing, decomposing, checking, revising, integrating or acting on information.Several behavioral instances and a codebook. No claim about hidden states is required.
ExternalizationA previously implicit or interaction-bound practice becomes explicit in a durable artifact, method, rule, representation or role.Versioned artifact linked to earlier evidence.
Computational representationThe externalized practice is encoded in a form a system can store, inspect, compare, route, execute, constrain or query.Schema, graph, workflow, agent role, runtime rule, machine-readable protocol, code or structured object.
ReintegrationThe representation is actively reused in later human–AI work.A real later session, mission or workflow trace.
Recursive effectThe reintegrated structure changes later work and contributes to a further externalization, revision or new structure.Before/after or longitudinal evidence with alternative explanations considered.
Emergent cognitive architectureA recurring distributed structure spanning human practice, interaction, persistent representations and computational mechanisms.Several verified recursive chains. A single artifact never suffices as evidence.

Twelve practice families

Concept

Candidates for recurring cognitive work practices, each with its observable indicators and the form in which it could be externalized. The families are working hypotheses. No findings on them exist.

Twelve practice families

Concept
NameObservable indicatorsPossible externalization
01Problem decompositionTasks are repeatedly split into sub-problems, dependencies, roles, gates or phases.Planner logic, mission graph, workflow specification.
02Contradiction detectionConflicts, mismatches or inconsistencies are repeatedly surfaced and corrected.Integrity agent, contradiction detector, review gate.
03Causal reconstructionMissing steps or states are inferred from partial evidence, and alternatives are compared.Graph reconstruction, workflow for unresolved edges, tool for causal hypotheses.
04Adversarial reviewCounterexamples, failure modes and reviewer attacks are generated on purpose.Red-team and reviewer roles, acceptance checks.
05Epistemic separationObserved, inferred, unknown, contested and speculative content is kept apart.Status schema, evidence register.
06Persistent continuityOpen loops and unresolved context are carried across sessions.State, memory and open-loop objects.
07Meta-controlRules for how the AI should interact arise from earlier failures.UIP-style interaction rules, guardrails, role boundaries.
08SpecializationRecurring work functions are differentiated into roles.Lead and specialist agent architecture.
09SynthesisDistributed evidence is recombined into higher-order models.Integrator roles, overview maps, evidence graphs.
10Self-observationAI is used to inspect recurring features of one's own work and decision process.SELFGRAPH, coded history of practice.
11Method formationRepeated interaction becomes a named, repeatable method.Framework, protocol, checklist, template.
12Acceptance logicCompletion is judged by explicit gates.Runtime acceptance gate, definition of done.

Eight of the twelve families have a candidate chain (see the transformation chain): problem decomposition, contradiction detection, adversarial review, epistemic separation, persistent continuity, meta-control, self-observation and acceptance logic. For causal reconstruction, specialization, synthesis and method formation it is still open.

Ten levels of externalization

Concept

Ten steps from implicit behavior to a representation that produces a new externalization, plus the evidence record kept per externalization.

Ten levels of externalization

Concept

Why externalization is the bridge between ephemeral interaction and durable computational structure is explained on the Research page. The ten levels follow here.

NameMeaning
1Implicit practiceVisible only in repeated behavior.
2Recognized patternExplicitly noticed or named.
3HeuristicArticulated as a reusable rule of thumb.
4Method or protocolSteps, constraints and success criteria are explicit.
5Structured representationRepresented as a schema, graph, checklist, state machine or typed object.
6Executable representationEncoded as a workflow, skill, agent role, runtime rule or software component.
7Persistent representationSurvives sessions and can be retrieved and reused.
8Instrumented representationEmits evidence, state, outcomes or provenance.
9Research representationMeasurable, versioned and linked to claims.
10Recursive representationIts use changes later work and starts a new externalization cycle.

Evidence record per externalization (12)

  1. Earliest recoverable example before formalization.
  2. Repeated examples before naming.
  3. First explicit recognition or naming.
  4. Intermediate formulations and discarded variants.
  5. First formal method or protocol.
  6. First machine-readable representation.
  7. First executable implementation.
  8. First persistent use.
  9. First evidence of reintegration.
  10. First evidence of downstream change.
  11. Counterexamples and failure cases.
  12. Version history, authorship, timestamps and source provenance.

Maturity of a representation (CR0 to CR7)

Concept

Eight levels for the maturity of a computational representation, each with its evidence criterion.

Maturity of a representation (CR0 to CR7)

Concept

CR0 to CR7 measure the maturity of a representation and are read separately from the maturity labels of this site.

NameDescriptionEvidence requirement
CR0Pure proseDescription in running text.No computational claim allowed.
CR1Structured, human-readableStructured representation readable by people.Fields or schema are explicit.
CR2Machine-readableRepresentation a program can parse.Can be parsed and stored consistently.
CR3QueryableRelational or queryable representation.The system can retrieve, connect or compare it.
CR4ExecutableRepresentation that constrains or drives system behavior.The representation changes system behavior.
CR5Persistent, operationalIn use across sessions and missions.Reuse across sessions or missions is documented.
CR6InstrumentedUse is recorded with evidence and provenance.Outcomes, evidence and provenance are captured.
CR7Recursively adaptiveUse feeds back into later representation or rules.Use flows into a later revision of the representation.

Representation types (9)

NameWhat it shows
Interaction ruleA recurring correction became a formal interaction constraint.
Metric or instrumentAn interaction property became a measurable object.
Typed graph objectA longitudinal relation between concepts or states became queryable.
Agent roleA recurring function became an executable artificial role.
WorkflowA task strategy became operational sequencing.
Runtime gateA judgment or control practice became execution logic.
State and memory objectContinuity became a machine-readable persistent layer.
Research objectInterpretation became traceable and auditable.
Adaptive policyA state-sensitive interaction strategy becomes closed-loop.

How a practice returns to the work

Concept

Without reintegration the study shows only externalization or offloading. With reintegration and downstream change, a recursive architecture can be tested.

How a practice returns to the work

Concept

Seven reintegration paths

NameHow it works
ProtocolA rule derived from earlier interaction changes future interaction.
AgentA specialized artificial role performs a function that previously took manual effort.
MemoryStored state and open items change what stands out later.
GraphA graph surfaces a contradiction, dependency or pattern and redirects inquiry.
RuntimeGates and authority constraints change how tasks are executed.
ResearchCoding earlier behavior changes how future behavior is observed.
MetaThe person revises the system after seeing how it affects the work.

What counts as evidence of reintegration (8)

  • Task decomposition changes after a planner or workflow is adopted.
  • Correction and verification behavior change after evidence-status rules apply.
  • Manual steps shrink or shift after roles are specialized.
  • New recurring categories become visible because a graph or registry shows them.
  • Contradictions or incomplete states are caught earlier because integrity structures are running.
  • Requests are phrased differently because the system expects structured input.
  • New methods or agents emerge from problems that earlier externalizations exposed.
  • Deliberate rollback or revision because a representation constrained, misdescribed or harmed.

The transformation chain

So far no chain has been fully reconstructed. The ten chains below are candidates.

The transformation chain

An externalization path is described in nine stages. The emergence lifecycle (on the Research page) describes a phenomenon, the chain describes the path of one practice into a system and back.

  1. T0PrecursorTimestamped interaction before the concept is named.
  2. T1RecurrenceSeveral examples across time or context.
  3. T2RecognitionExplicit naming.
  4. T3FormalizationMethod or specification.
  5. T4RepresentationMachine-readable or executable.
  6. T5DeploymentUsed in real work outside test data.
  7. T6ReintegrationThe output affects the person's next step.
  8. T7Practice shiftLater work changes after adoption.
  9. T8RecursionThe changed practice produces a new externalization.

A chain counts as supported only when T0, T1, T3, T4, T5 and T6 are documented, plus T7 or T8. Missing stages may stay inferred, and the chain then carries a weaker status. What “supported” means →

The path in the overall program (9 links)

Concept

The chain shows how interaction practice becomes technical and scientific infrastructure. Every link needs timestamped evidence and may be missing or uncertain. The links name what arises. The stages T0 to T8 above name the evidence a chain needs.

The path in the overall program (9 links)

Concept
  1. Interaction
  2. Pattern
  3. Cognitive operation
  4. Named heuristic
  5. Method or protocol
  6. Skill or agent role
  7. Runtime rule
  8. Framework or system
  9. Claim of a study

Ten candidates, no results yet (10)

Ten candidates, no results yet (10)

  • Semantic integrity

    Exploratory

    Early practice Repeated correction of drift and fabrication.

    Representation Fixed rules and automatic checkers (guards).

    Possible reintegration signal Corrections become system-enforced instead of manual.

  • Contradiction handling

    Exploratory

    Early practice Manual detection of inconsistency.

    Representation Review roles and contradiction checks.

    Possible reintegration signal Contradictions surface earlier and redirect reasoning.

  • Task decomposition

    Exploratory

    Early practice Manual splitting into sub-steps.

    Representation Planner and workflows.

    Possible reintegration signal Growing delegation of structured sub-problems.

  • Epistemic separation

    Exploratory

    Early practice Telling observation, inference and uncertainty apart.

    Representation A status schema and evidence objects (records for evidence).

    Possible reintegration signal Analysis is framed through evidence states.

  • Persistent continuity

    Exploratory

    Early practice Manually carrying open items.

    Representation State and memory objects.

    Possible reintegration signal Work starts from the stored state instead of re-explaining.

  • Emergence observation

    Exploratory

    Early practice Repeatedly noticing interaction phenomena.

    Representation Registry and lifecycle.

    Possible reintegration signal New events are measured against formal criteria.

  • Adversarial review

    Exploratory

    Early practice Manually asking for attacks and failure cases.

    Representation Review and attack roles.

    Possible reintegration signal Review happens before human acceptance.

  • Self-observation

    Exploratory

    Early practice The AI inspects recurring patterns in the person's own work.

    Representation SELFGRAPH (research system for state and continuity across long timeframes) plus practice coding.

    Possible reintegration signal Outputs influence future decisions and questions.

  • Completion discipline

    Exploratory

    Early practice Frustration with partial completion and drift.

    Representation Definition of done and acceptance gate (a check before sign-off).

    Possible reintegration signal The system continues until objective criteria are met.

  • Research formalization

    Exploratory

    Early practice Observations from conversations become structured science.

    Representation Research map and claim register.

    Possible reintegration signal Later research is bound by the evidence architecture.

Time, confounds and comparison groups

Besides reintegration, other factors change work practice. They are kept on a timeline and controlled for in every before/after statement.

Confounds on the timeline (7)

2023 Baseline · 2024 Development · 2025 Formalization · 2025 to 2026 Precursors of automation · 2026 Agents and runtime · 2026 SELFGRAPH, persistent working environment, Observatory · afterwards Prospective phase

Confounds on the timeline (7)

  1. 2023Baseline
  2. 2024Development
  3. 2025Formalization
  4. 2025 to 2026Precursors of automation
  5. 2026Agents and runtime
  6. 2026SELFGRAPH, persistent working environment, Observatory
  7. afterwardsProspective phase
Model upgrades
A stronger base model changes answers and working style without any externalization.
Memory changes
Memory features of the model can produce continuity that comes from the model. This continuity cannot then be attributed to the practice.
Tool changes
New interfaces and tools shift what work looks like.
New systems
Our own new systems change the work and are logged with their time of introduction.
Domain changes
A different field of work calls for different practices.
Life and project circumstances
General changes in life and in projects act on the practice.
Expectancy effect
Once a pattern has been named, it is easier to see again.

Confounds are influences that could explain a change even without the practice under study. The threats to validity are tied to the confounds and to the study design and are not spelled out again here. The conflict of roles is under “The role conflict”. The limits of a single participant (N-of-1) are under “One person, several years”. Model and platform changes are under model upgrades and tool changes. Contamination by memory is under memory changes. The counter-explanations for the hypotheses are on the Research page.

Comparison groups: further partitions of the material

Planned

Besides the timeline, the material is divided by interaction mode, system phase and model generation. That way platform changes can be told apart from human adaptation.

Comparison groups: further partitions of the material

Planned

By interaction mode (9)

  • Exploration
  • Execution
  • Reflection
  • Planning
  • Debugging
  • Adversarial review
  • Automation
  • Agent orchestration
  • Personal reconstruction

By system phase (6)

  1. Ordinary chat use
  2. Framework formation
  3. Automation
  4. Agents and runtime
  5. Observatory, working environment and SELFGRAPH
  6. Future neuroadaptive phase

Further rules

  • Finer monthly or quarterly cuts when event density allows.
  • Division by model and tool generation where metadata exists.
  • A primary dataset and a validation dataset check whether classifications stay stable.

Threats to validity

Concept

The objections every study has to address before it claims evidence.

Threats to validity

Concept
  1. Selection bias and self-selection.
  2. Limits of a single participant (N-of-1).
  3. Conflict between participant, researcher and system designer.
  4. Model and platform changes over time.
  5. Contamination by memory and retrospective labeling.
  6. Anthropomorphic or neural over-interpretation.
  7. Circular validation when the same AI generates and validates the claim.
  8. Metric gaming and post-hoc thresholds.
  9. Non-independence of turns and sessions.
  10. Generalization beyond the observed person and system.
  11. Missing data and deleted-chat bias.
  12. Implementation-driven confounding: a feature can appear because it was deliberately coded, and may not have emerged on its own.

Data, access and protection

Six forms of data

  • Interaction Data

    Longitudinal Human–AI conversations and behavioral interaction traces.

  • System Data

    Events, state changes, tool calls and agent activity.

  • Knowledge Data

    Entities, relations, claims, evidence and provenance.

  • Behavioral Data

    Patterns in strategies, decisions and interaction.

  • Simulation Data

    Outputs from controlled experimental or scenario environments.

  • Cognitive Interaction Data

    Planned

    Signals derived from task behavior, interaction structure and future neuroadaptive instrumentation.

Six forms of data

EIL works with six forms of data. Interpretation follows one rule: measured or observed information stays distinguishable from inference.

What is open, what is protected, what stays private

EIL works openly as science and still does not hand out everything. Three layers are kept apart.

Public research

Observations, taxonomies, empirical results, theoretical models and methods that others can reproduce.

Protected implementation

Purpose and function of a system are described publicly. Its internal way of working stays in house.

Private corpus

Complete personal interaction data, raw logs and sensitive content are not published. At most coded events, summarized chains and selected excerpts are public, after minimization and anonymization.

Eight processing layers L0 to L7 · Concept

The Observatory shows fifteen data layers and describes what the corpus contains. The layers L0 to L7 describe how material is processed from raw archive to research claim.

Eight processing layers L0 to L7 · Concept

NameContentsRule
L0Raw archiveOriginal chats, files, timestamps, tool traces, model metadata where available.Immutable, private, versioned.
L1Normalized eventsSession, turn, actor, timestamp, event type, task or context.No interpretation.
L2Observable codesDecomposition, correction, verification, contradiction, reconstruction, delegation and others.Defined by the codebook.
L3Derived featuresCounts, recurrence, intervals, graph metrics, transition frequencies.Method and version are recorded.
L4InterpretationsCandidate cognitive operations, pattern families, hypotheses.Confidence and alternatives are required.
L5Externalization recordsMethods, schemas, agents, runtime rules, systems and their versions.Linked backward to evidence.
L6Reintegration recordsReal later use and the resulting workflow changes.Linked forward to the outcome.
L7Research claimsClaims of the studies, counterevidence and epistemic status.Traceable end to end.

Ethics and data management

Ethics and data management

  • Data of third parties is minimized or removed.
  • Sensitive health and psychological content stays private even when it appears in the material.
  • Human agency stays measurable: overriding, rejecting and rolling back count as evidence of control.

The role conflict and the separation of interpretation and evidence are under “Study design and phases”, the private corpus under “Private corpus”.

Limits and templates

Scope of the claims

What EIL claims

  • Persistent interaction can produce lasting computational structures that act back on the practice they came from.
  • What is sought is a distributed, recursively developing cognitive architecture at the level of observable work practice.
  • Each of these statements is a target to test and is carried with its status and its counterevidence.

What is in scope and what is outside

  • The subject is the coupled system of practice and structure. General productivity and personality profiles are outside the scope.
  • Examples are chosen by fixed rules from the whole corpus, and counterexamples stay in the register.
  • Repetition counts as a candidate for emergence and an externalized routine as a candidate for a change in practice until evidence exists.

Fixed limits

Active research

Working rules that apply to every publication of the program.

Fixed limits

Active research
  1. No direct brain-reading claim is ever derived from conversational data.
  2. Agents are never treated as biological cognitive modules.
  3. The lab's systems are kept distinct from one another.
  4. No registry entry is automatically elevated to “supported emergence”.
  5. Future neuroadaptive results never appear as current findings.
  6. Nothing unnecessary from the private corpus is published.
  7. Novelty is claimed only after a review of the state of the art.
  8. No causal conclusion goes beyond what the study design supports.

Rules for sources, literature and figures

Concept

Working rules that keep proximity to other work, or a well made figure, from claiming more than the evidence carries.

Rules for sources, literature and figures

Concept

Sources and literature (6)

  • Every novelty claim gets its own prior-art search.
  • Borrowed constructs receive the primary source where possible.
  • Adjacent work is not cited as if it had validated the EIL claim.
  • For fast-changing AI systems the exact publication date and version are recorded.
  • A log of rejected sources keeps the work that looked relevant and did not support the claim.
  • Each study has its own literature map, because one shared bibliography creates false proximity between unrelated fields.

Figures (3)

  • No figure implies stronger causal or psychological knowledge than the evidence supports.
  • Inferred nodes and edges are visibly distinguishable from observed ones.
  • Future and planned components are marked as such.

Templates: chain record, claim matrix and dossier

Concept

Fields only, without content. Filled templates stay internal.

Templates: chain record, claim matrix and dossier

Concept

Chain record (17 fields)

  • Chain ID
  • Practice family
  • Earliest precursor (timestamp and source)
  • Recurrence set
  • Recognition event
  • Formalization event (version)
  • Representation (type and CR level)
  • Deployment
  • Reintegration
  • Observed effect
  • Recursive output
  • Strongest alternative explanation
  • Counterevidence
  • Confounds
  • Epistemic status per link
  • Linked studies
  • Publication class

Claim–evidence matrix (7 fields)

  • Claim ID
  • Exact claim
  • Evidence for
  • Evidence against
  • Alternative explanation
  • Status
  • Study

Verification dossier per phenomenon (14 fields)

  • ID, name and family
  • Operational definition
  • First known evidence
  • All material recurrences
  • Contexts represented
  • Negative cases and contradictions
  • Candidate mechanisms
  • Functional consequences
  • Externalization events
  • Related phenomena
  • Instrument versions used
  • Epistemic status and confidence
  • Reviewer objections
  • Validation still required

Hypotheses, evidence and structure

Claim and evidence status Concept

  1. Observed
  2. Derived
  3. Inferred
  4. Hypothesis
  5. Supported
  6. Contested
  7. Rejected
  8. Unknown

So far no chain and no claim on this site carries that status.

Claim status

Concept

Two scales with separate jobs. The epistemic tags, that is, marks for the state of knowledge (observed, inferred, hypothesized and the rest), say where a single piece of information comes from. The claim status says how far a claim or a chain is carried by evidence and counterevidence. Both apply side by side.

NameMeaningAllowed wording
ObservedDirectly visible in the source data.“The interaction record contains…”
DerivedComputed deterministically from observed data.“Using method and version X, the corpus yields…”
InferredAn interpretation that evidence supports and that cannot be observed directly.“This is consistent with…”
HypothesisA proposed mechanism or explanation. It requires a future test.“We hypothesize…”
SupportedSeveral evidence streams, reviewed counterevidence and validation. This is the highest level. It describes what occurs in the data and claims no cause.“The evidence supports…”
ContestedMaterial alternatives remain. Disagreement stays visible.“Evidence is mixed or contested…”
RejectedCurrent evidence argues against the claim. It stays in the research history.“The hypothesis was not supported…”
UnknownThe evidence is insufficient. This is an explicit abstention.“The available data do not establish…”

This is how the scales relate: observed, derived, inferred, hypothesized and unknown appear in both. Simulated and contradicted exist only for single pieces of information, for example in reconstructions. Supported, contested and rejected exist only for claims and chains.

The status “supported” is provided for in three places: in the acceptance rule of the transformation chain, in the boundary that no registry entry rises automatically to supported emergence, and for the claims of the studies.

Two further scales are read separately: the register statuses (accepted, accepted with limitation, negative case) measure how sure it is that a phenomenon occurs in the raw material. The maturity labels of this site, from concept to validated, describe the state of an entry in the lab.

What every claim carries in the claim matrix (8)

  1. Claim ID and the exact sentence of the claim.
  2. Primary evidence and independent supporting evidence.
  3. Counterevidence and alternative explanation.
  4. Applicable period and sample.
  5. Method and instrument version.
  6. Strength and uncertainty.
  7. Public evidence reference and private reference to the raw source.
  8. Reviewer challenge and the response to it.

What we assume and what it is based on

Hypotheses (7)

Hypotheses (7)

None of them is confirmed. Each one states what would support it and what would weaken it.

Hypothesis Claim status in detail →

  1. H1

    Persistent human-AI interaction reveals stable recurring cognitive work practices.

    Supports
    Repeated patterns across contexts, reliably coded.
    Weakens
    Patterns disappear when task or topic changes.
  2. H2

    Some recurring practices become explicitly externalized over time.

    Supports
    A traceable path from pre-naming behavior to a formal method.
    Weakens
    Artifacts appear without an identifiable precursor practice.
  3. H3

    Some externalizations become computational representations.

    Supports
    Artifacts progress to machine-readable, queryable or executable forms.
    Weakens
    They remain prose or decoration.
  4. H4

    Those representations are reintegrated into later work.

    Supports
    Real reuse in later sessions, missions or workflows.
    Weakens
    They are never used after creation.
  5. H5

    Reintegration measurably changes later cognitive and work practice.

    Supports
    Before/after differences that survive confound checks.
    Weakens
    The change is fully explained by model or tool upgrades or by different tasks.
  6. H6

    Some changed practices produce further externalizations, forming recursive cycles.

    Supports
    At least two traceable generations in the same chain.
    Weakens
    Only one-way offloading occurs.
  7. H7

    The coupled human-AI system can be described as a distributed cognitive architecture at the level of observable work practice.

    Supports
    Several validated recursive chains across different practice families.
    Weakens
    Evidence stays confined to one anecdotal chain.

What would make the program fail (10)

  • Practices cannot be coded reliably before their later formalization.
  • Externalized artifacts cannot be traced to recurring precursor practices.
  • The representations are mostly prose and not computational objects.
  • They are created but not reused.
  • Reintegration produces no detectable change.
  • The change disappears once model upgrades and task mix are controlled for.
  • Only one-way offloading occurs and no recursive cycles are found.
  • The evidence comes only from selected showcase sessions and fails a review of the full corpus.
  • A more parsimonious explanation fits the data better.
  • The architecture depends entirely on retrospective naming by the same researcher.

Alternative explanations that have to be ruled out (10)

  • Normal development of expertise.
  • Growing familiarity with AI interfaces.
  • Stronger base models.
  • Improvements in tooling and memory.
  • Changes of task and domain.
  • Researcher expectancy after a pattern has been named.
  • Survivorship bias in the retained conversations.
  • More documentation instead of a real change in practice.
  • Constraint from automation mistaken for cognitive integration.
  • General life and project changes.

Evidence register (36)

Evidence register (36)

The entries come from the raw material of several years of interaction. Every entry says how strong the evidence is and where it ends. Raw data, conversation numbers and content stay private. With every new export, the phenomena are checked again.

thirty-six entries. Entries that share one source, trajectory or claim are marked (7 groups). 3 accepted, 32 with limitation, 1 negative case.

These entries come from the material of a single person. They show what occurs in this material and say nothing about other people.

How to read the statuses.

Accepted
Several independent instances in the raw material, verified verbatim.
Accepted with limitation
The phenomenon is documented, and one limit stays visible, for example a single instance, a possible confound or a source shared by several entries.
Negative case
The entry shows where the phenomenon does not hold.

No entry is externally validated. The register statuses measure how strong the evidence is that the phenomenon occurs. They are a different scale from the epistemic labels observed, inferred or hypothesized.

36 / 36

IDPhenomenonStatusLimitation
E001Path dependence Accepted with limitationSession constraints held without being restated. Persistence across sessions is open.
E003Recursive correction loops AcceptedThree independent cases in which the same correction had to be repeated, verified verbatim.
E004Error integration Accepted with limitationThe same class of error returns over weeks, which argues against lasting integration. Source shared with E003.
E005Clarification loops AcceptedVerified verbatim, across a change of model generation.
E022Emergent formalization Negative caseOne clean single case from a real analysis task that does not recur. An earlier positive finding was withdrawn.
E027Role stabilization Accepted with limitationA possible confound is the memory feature of the model. The control is still to be run.
E034Goal drift (negative case) Accepted with limitationWhat was observed is a practice of monitoring drift. Drift itself is not documented.
E040Contradiction-driven refinement Accepted with limitationSingle instance. Recurrence is not yet confirmed.
E041Negative definition AcceptedThe cleanest finding of the first rounds of review.
E046Coherence is not truthtrajectory Accepted with limitationShares its source with E047, E049 and E050. One trajectory and four entries.Four entries follow one single trajectory.
E047Dialogically externalized introspectiontrajectory Accepted with limitationSource shared with E046.Four entries follow one single trajectory.
E049Implicit-to-reportable transitiontrajectory Accepted with limitationSource shared with E046.Four entries follow one single trajectory.
E050Detection of contradictory self-descriptiontrajectory Accepted with limitationSource shared with E046. The strongest of the four entries.Four entries follow one single trajectory.
E063Layer separationshared claim Accepted with limitationShared claim: the three-part practice is documented in interaction, the eight-layer model exists as a specification. There is no single status.Both entries carry the same claim.
E064Separation of evidence and interpretation Accepted with limitationA detail cited earlier was corrected.
E066Reconstruction requirement Accepted with limitationVerified verbatim, single source.
E072Emergence of the eight-layer modelshared claim Accepted with limitationShared claim with E063. The three-part practice comes from interaction, the eight-way taxonomy exists as a specification.Both entries carry the same claim.
E105Context loss loops Accepted with limitationA real failure report, verified verbatim. Source shared with E106.
E106Error repetition Accepted with limitationSource shared with E105.
E108False completion Accepted with limitationThe strongest finding in the range of agent and system phenomena.
E109Local success, global failure Accepted with limitationSource shared with E110.
E110Coordination failure Accepted with limitationSource shared with E109.
E123Autonomous continuation Accepted with limitationThe demand for it is documented. Attaining it is open.
E134Observer-system separationCode existence confirmed Accepted with limitationConfirmed in code. The naming preceded the implementation.
E138Behavioral landscape formationCode existence confirmed Accepted with limitationConfirmed in code. Naming and implementation lie close together, the clearest hit of the code check.
E154From prompt author to system operator Accepted with limitationDirectly observed user behavior. The strongest of the group E154 and E155.
E155Human as source of constraints Accepted with limitationExplicit prohibitions in the material. Same sources as E154.
E162Epistemic integrity as a system property Accepted with limitationFrom a real analysis task.
E167Boundary between reconstruction and fabrication Accepted with limitationFrom a real analysis task. Source shared with E168 and E170.
E168Reconstructive intelligence Accepted with limitationFrom a real analysis task. Source shared with E167.
E170Adversarial reasoning Accepted with limitationFrom a real analysis task. Source shared with E167.
E171Gap-oriented reasoning Accepted with limitationFrom a real analysis task.
E172Transition-oriented reasoning Accepted with limitationFrom a real analysis task.
E173Constraint-driven abstraction Accepted with limitationA possible overlap with E041, kept apart in the source material.
E174Rapid structural compression Accepted with limitationThe size before and after compression is documented.
E176Emergence of a research environment Accepted with limitationThe founding of the lab was a top-down decision. The founding event itself is documented.

Entries from real analysis tasks show only the phenomenon. Facts of the case and the people involved stay private. State: snapshot from 28 September 2026.

Longitudinal research (9)

Longitudinal research (9)

A defining feature of EIL is its emphasis on longitudinal interaction. Many AI studies look at isolated tasks or short laboratory sessions. EIL examines systems across extended time periods, which makes different questions possible.

The nine questions (9)

  • Does behavior stabilize?
  • Do strategies transfer between tasks?
  • Do interaction structures persist?
  • Do humans change how they reason with AI?
  • Does memory alter future behavior?
  • Does coordination improve?
  • Do errors accumulate?
  • Does semantic drift appear?
  • Do emergent patterns disappear again?

This temporal dimension is essential for studying genuine adaptive systems.

Research traditions (14)

Research traditions (14)

This question connects research traditions that are usually studied separately. EIL reads them as perspectives on one shared problem.

  • Human–Computer Interaction
  • Human–AI Interaction
  • Cognitive Science
  • Artificial Intelligence
  • Complex Systems
  • Distributed Cognition
  • Agent Systems
  • Adaptive Systems
  • Computational Social Science
  • Behavioral Science
  • Neuroadaptive Systems
  • Epistemology
  • Knowledge Representation
  • Systems Engineering

Shared problemHow does an intelligent system construct, update and act on an internal representation of itself and its environment?

How it is structured

Eight contribution layers (8)

Concept

The program arranges its contributions in eight layers. Each has an object of research and a primary output that is still being built.

Eight contribution layers (8)

Concept
NameResearch objectOutput
1InteractionHuman–AI dialogue and adaptation over time.Longitudinal event corpus.
2EmergenceRecurring and stabilizing interaction phenomena.Phenomena register with dossiers.
3CognitionObservable cognitive practices and their transformations.Taxonomy of cognitive operations.
4ExternalizationThe movement from cognitive practice to an explicit artifact or system.Transformation chains.
5Distributed intelligenceAgents, handoffs, roles, control and integrity.Agentic architecture studies.
6Personal longitudinal modelEvolution of projects, ideas, decisions, states and open loops.SELFGRAPH as a research instrument.
7Meta-observationMeasurement, provenance, evidence and drift.Observatory.
8Future adaptive layerState-aware adaptation using multimodal signals.Neuroadaptive research program.

Six parts of the program structure and the eight spaces (6)

Concept

The overall program divides the same field as the eight spaces A to H into six parts. The mapping is a reading aid. Space F, Computational Interaction & Cognitive Modeling, runs across these parts and connects human cognition, externalization and measurement.

Six parts of the program structure and the eight spaces (6)

Concept
NameWhere the part sitsSystems and instruments
BHuman intelligenceSpace B, Human Cognition, Behavior & Externalization. Thinking processes, strategies, decisions, cognitive operations, personal development.SELFGRAPH, LT-Data, cognitive-operation coding.
AHuman–AI interactionSpace A, Human–AI Interaction & Co-Evolution. Co-adaptation, semantic repair, drift, resonance, integrity, emergent interaction patterns.UIP, IEIA, EIA, CEI, CCET, CEP.
EDistributed intelligenceSpace E, Distributed, Agentic & Collective Intelligence. Multi-agent decomposition, specialization, handoffs, integrity, execution.Agent architecture, Kernel, Runtime.
DWorld intelligenceSpace D, Reconstruction, Perspective & World Models. External entities, events, hidden states, causal reconstruction.DINGIR.
HMeta-intelligenceSpace H, Measurement, Evidence & Research Infrastructure. Observation of human, AI, agents and systems across time.Observatory, evidence and provenance backbone.
Neuroadaptive futureThe program Neuroadaptive Human–AI Systems has its main space in G and touches B and H. Several signals estimate states, and the system adapts within bounds.EEG, BCI, wearables and future adaptation rules, all planned.

Space C, Emergence, Adaptation & Complex Dynamics, has no part of its own in this division. It runs through all six.

Cross-cutting questions (8)

Cross-cutting questions (8)

  • How does repeated interaction alter both human and artificial behavior?
  • When does a recurring interaction pattern become a stable structure?
  • Which cognitive operations become externalized?
  • How do externalized structures change subsequent human behavior?
  • How can inferred human or system state remain distinguishable from direct observation?
  • How does distributed intelligence alter responsibility, authority and failure?
  • When does system memory preserve continuity versus produce false continuity?
  • How can longitudinal changes be measured without confusing tool drift with real behavioral change?