Skip to content
Contact
☰
Color scheme: Dark.

Methods

Methods are the tools EIL uses to observe, reconstruct and measure interaction systematically. They translate complex trajectories, patterns and changes into traceable models, comparison measures and reliable statements.

Research process: from raw observation to research claim

This is the order in which EIL turns an observation into a statement, and the rule that nothing unknown quietly becomes knowledge.

Observation first. Interpretation second. Claim last.

  1. 01Raw observation
  2. 02Data structuring
  3. 03Pattern detection
  4. 04Longitudinal comparison
  5. 05Hypothesis
  6. 06Adversarial testing
  7. 07Evidence assessment
  8. 08Interpretation
  9. 09Research claim

The nine steps

A typical EIL research process follows this structure. It keeps interesting patterns from turning into scientific claims automatically.

  1. 01Raw observationThe material as it stands.
  2. 02Data structuringThe material is ordered and coded.
  3. 03Pattern detectionRecurring things are looked for.
  4. 04Longitudinal comparisonPatterns are compared across longer periods.
  5. 05HypothesisAn explanation is proposed.
  6. 06Adversarial testingCounterexamples are looked for on purpose.
  7. 07Evidence assessmentIt is weighed how far the evidence carries.
  8. 08InterpretationThe findings are interpreted.
  9. 09Research claimThe result is stated as a claim with a status.

How the sequence becomes the development of systems: the six steps (Observe, Reconstruct, Model, Design, Build, Validate). AI is used throughout this process as a research, analysis, modeling and development tool. All concepts, architectures, interpretations and final decisions remain the lab's own responsibility.

The logic of the research in outline

  1. Observation
  2. Contradiction
  3. Hypothesis
  4. Reconstruction
  5. Alternative explanation
  6. Operationalization
  7. Test
  8. Revision

This is the logic of the research in outline. It sits beside the nine steps of the research process and does not replace them. Which existing methods, frameworks and instruments support which step is stated on their own entries; where no support exists yet, that is an open task and not a finished capability.

Mechanism-oriented research

EIL asks which phenomena recur and also which conditions may have contributed to their formation. That includes reconstructing sequences, relations, contexts, feedback and possible transition mechanisms. A reconstruction names a candidate mechanism; it does not prove it.

Alternative explanation testing

An emergent explanation has to be tested against simpler ones. The simpler explanations come first, and an emergent reading stays a hypothesis while they remain open:

  • direct instruction
  • designed orchestration
  • ordinary repetition
  • change of task
  • model effects
  • changes of context
  • selection effects
  • chance variation

Observation and interpretation stay separate throughout.

Methods and analysis procedures

Computational analysis procedures12

These are analyses run on interaction and system data, such as searching for recurring patterns or comparing several explanations. EIL uses them to test interpretations against the data; they count as exploratory unless a running implementation is documented.

Temporal Pattern Mining

Searches timestamped interaction events for patterns that repeat across time windows.

Temporal Pattern Mining

Exploratory

Derived fromImplicit-to-reportable transition. Candidate chain.

Used inSELFGRAPH

Supported research programs
Research spaces
Research questions (2)

Recurrence Detection

Checks whether a pattern recurs in independent sessions or contexts before it is treated as stable.

Recurrence Detection

Exploratory
Supported research programs
Research spaces
Research questions (2)

Clustering

Groups similar interaction episodes to find candidates for recurring patterns.

Clustering

Exploratory

DescriptionA cluster stays a candidate until recurrence is checked.

Supported research programs
Research spaces
Research questions (2)

Graph Analysis

Models relations between events, concepts or actors as a graph and examines its structure, such as centralization or lineage chains.

Graph Analysis

Exploratory
Supported research programs
Research spaces
Research questions (2)

State Reconstruction

Reconstructs a probable system state from incomplete, distributed evidence and keeps the uncertainty explicit.

State Reconstruction

Exploratory

StatusNo implementation runs for the website. As a computational analysis the method stays exploratory.

Supported research programs
Research spaces
Research questions (2)

Semantic Drift Analysis

Tracks how the meaning and use of terms shift over the course of an interaction.

Semantic Drift Analysis

Exploratory
Supported research programs
Research spaces
Research questions (2)

Behavioral Feature Extraction

Derives behavioral features from interaction data, such as term reuse, role shifts or self-audit.

Behavioral Feature Extraction

Exploratory

DescriptionBasis of the behavioral proxies in the neuroadaptive program.

Derived fromObserver-system separation, Behavioral landscape formation. Candidate chain.

Used inObservatory

Supported research programs
Research spaces
Research questions (2)

Cross-Context Comparison

Compares the same pattern across independent contexts to check whether it occurs generally or is tied to one context.

Cross-Context Comparison

Exploratory
Supported research programs
Research spaces
Research questions (1)

Anomaly Analysis

Looks for deviations from the expected course to tell instrument drift, false states and genuine breaks in the interaction apart.

Anomaly Analysis

Exploratory

Derived fromDetection of contradictory self-description. Candidate chain.

Used inMIRROR

Supported research programs
Research spaces
Research questions (2)

Multi-Hypothesis Reconstruction

One observation opens several explanations.

Multi-Hypothesis Reconstruction

Exploratory

DescriptionThe system keeps competing interpretations side by side and compares each with the evidence. Narrative coherence does not count as evidence.

StatusNo implementation runs for the website. As a computational analysis the method stays exploratory.

Derived fromBoundary between reconstruction and fabrication. Candidate chain.

Used inDINGIR

Supported research programs
Research spaces
Research questions (2)

Overarching methodological operations4

Research artifacts that emerged from interaction12

These entries emerged from the long-term interaction and were formalized afterwards.

Emergent Interaction Phenomena in Longitudinal Human-AI Interaction

In preparation, not yet public

An Evidence-Tiered Compendium: Formation, Stabilization, Naming, Externalization, and Cross-Corpus Replication

Using one long human-AI interaction, the paper describes how recurring reasoning habits and frameworks form, stabilize, receive a name and in part pass into running code. Its core statement is that the architectures and the code arose from emergence. Every statement is sorted by evidence tier, and the path from name to code is documented to a different degree for each construct.

Analysis framework1

EIA

Active research

Origin: Analysis framework that emerged from interaction

Qualitative or multi-method analysis of repeated human–AI interaction.

EIA

Emergent Interaction Analysis

DescriptionQualitative method for tracing how meaning, abstraction and reasoning structure emerge through sustained dialogue.

OriginEmerged from the longitudinal GPT interaction and was later used methodically.

StatusDescribed in a published preprint and applied to a single-subject longitudinal corpus. No independent validation.

The problemLooking at a single session in isolation obscures whether a pattern is chance or a real trend.

How it works
  • Breaks an interaction series into individually comparable episodes.
  • Compares patterns across repeated runs.
Supported research programs
Research spaces
Research questions (3)

Architectures, protocols and benchmarks3

Frameworks that connect several procedures and instruments into one research architecture.

IEIA

Active research

Origin: Architecture that emerged from interaction

Overarching architecture that connects several analysis procedures and instruments.

IEIA

Integrated Emergent Interaction Architecture (IEIA-2025)

DescriptionResearch architecture of seven instruments (UIP, EIA, CCET, LSG-24, LAP-1, 8-Layer Model, LT-Data) that grew out of recurring working patterns.

OriginEmerged from the long-term interaction and was formalized afterwards.

StatusThe name bundles instruments of different maturity. The status applies to the architecture as a whole. Individual instruments can have a different maturity than the architecture.

Supported research programs
Research spaces
Research questions (2)

UIP

Active research

Normative or methodological protocol for semantic integrity and stability.

UIP

User Integrity Protocol

DescriptionNormative framework for semantic and ethical stability in human–AI dialogue. Checks whether a system treats a user consistently over time.

StatusA normative framework without an independent validation study.

The problemWithout a recorded history, a system treats every user like a first contact, even after hundreds of interactions.

How it works
  • Operates on longitudinal snapshots.
Works across the labUIP works across the whole lab. Its principles can shape the areas below. That makes UIP part of the methodological infrastructure connecting the different EIL systems. interaction design, agent handoffs, role integrity, memory handling, state transitions, provenance, semantic consistency, drift detection, system monitoring, human oversight.

Derived fromRecursive correction loops. Candidate chain.

Used inKernel

Supported research programs
Research spaces
Research questions (2)

LIIE

Active research

Open research benchmark and evaluation framework for multi-year human–AI interaction trajectories: beneficial, preserving and deteriorating patterns.

LIIE

Longitudinal Interaction Impact Evaluation

StatusThe accompanying paper is in preparation. No result or comparison with other systems exists.

The problemShort-term tests cannot show whether a human–AI interaction gets better or worse over months. Long-term patterns need a long-term benchmark.

How it works
  • Captures beneficial, preserving and deteriorating interaction patterns across multi-year trajectories.
  • Measures reconstruction accuracy, agency and semantic integrity.
  • Tests dependency, uncertainty calibration, recovery and false-state controls.
  • Serves as the longitudinal evaluation protocol across studies: windows, outcomes, interventions, transfer, persistence, adverse effects and instrument versions. It supports the central question and does not replace the transformation-chain method.

Nine dimensions (9)Concept

LIIE is meant to carry the evaluation protocol for long-term effects across several studies. The nine dimensions are a draft. None of them is fixed as a measure yet.

NameMeasurable objects
1Interaction changeRepair frequency, iteration depth, task switching, correction rate, context carry-over.
2Cognitive-practice changeDecomposition, hypothesis branching, verification behavior, abstraction shifts.
3ExternalizationNew method, checklist, prompt rule, skill, agent, runtime gate, graph relation, software artifact.
4OutcomeTask completion, error rate, coverage, reproducibility, reuse, time to result.
5StabilityPersistence across contexts, models and time periods.
6DependencyWhether a capability disappears without AI or system support.
7TransferWhether a practice acquired through interaction appears in new tasks or tools.
8Adverse effectOverreliance, loss of independent verification, systematic drift, false certainty.
9Human agencyFrequency and quality of overrides, preserved choice, explicit rejection and recovery.

What LIIE still needsexplicit measurement units; time windows; outcome definitions; instrument versioning; known failure modes; an independent validation design with clear minimum-evidence rules.

Supported research programs
Research spaces
Research questions (2)

Measurement systems and research instruments4

Measurement systems1

CCET
Implemented

Quantitative monitoring layer that tracks how user and system change together across repeated interaction.

CCET

Continuous Co-Evolution Tracker

DescriptionProduces CEI, CEP and resonance frequency.

StatusRuns live in the Observatory. Only the CEI formula comes from the paper; the other definitions are our own operationalizations.

Supported research programs
Research spaces

Research instruments2

Independent research instruments of the lab.

LSG-24
Exploratory

Structured research instrument for one concrete, recurring analysis task within the EIL method family.

LSG-24

Instrument of the IEIA-2025 family

DescriptionThe same holds for LAP-1.

StatusFor LSG-24 and LAP-1, only the name and the role as an instrument are documented publicly. A full specification is due with the instruments papers.

Supported research programs
Research spaces
Research questions (1)

Data instruments1

LT-Data
Exploratory

Data instrument for preparing longitudinal interaction data so that trajectories across many sessions become comparable.

LT-Data

Longitudinal data instrument

StatusOnly the name and its role as a data instrument are documented publicly. Raw data stays private.

Supported research programs
Research spaces
Research questions (1)

Metrics3

Quantities tracked over time, for example the share of stable turns (CEI). Several are the lab's own operationalizations and not externally validated.

CEI

Implemented

Share of stable turns among all turns, range 0 to 1.

CEI

Co-Evolution Index

DescriptionA value updated over time that indicates the direction of an interaction.

StatusThe formula comes from the paper. The paper does not define a stable turn; the operationalization in the Observatory is our own and not externally validated.

The problemWhether an interaction is improving or degrading is guesswork without continuous measurement.

How it works
  • Computed in the Observatory over the last 200 turns, system-wide.
  • Shows a trend toward improvement, stagnation or deterioration.
Supported research programs
Research spaces

CEP

Experimental

Cumulative count of turns that are stable and whose predecessor was also stable.

CEP

Co-Evolution Points

DescriptionCaptures sustained stability.

StatusThe paper reports CEP only as an output value and gives no computation. The definition is our own reading and not validated.

Supported research programs
Research spaces

Resonance Frequency

Experimental

Rate at which a turn reuses framework terms from the preceding turn.

Resonance Frequency

Resonance frequency

DescriptionShows how often vocabulary carries forward from turn to turn.

StatusOur own operationalization with no formula in the paper. Not validated.

Supported research programs
Research spaces

Model1

Historical reference. EIL keeps the 8-Layer Model because it is one of the seven instruments of IEIA-2025.

8-Layer Model

Historical reference

Conceptual model that breaks a reasoning trace into eight distinguishable layers.

8-Layer Model

Historical framework model

DescriptionToday a reference model, not actively developed further.

The problemAn error is often noticed where it becomes visible, far from its origin.

How it works
  • Breaks a reasoning trace into eight distinguishable layers.
  • Reveals which layer an error actually originates in.
Supported research programs
Research questions (1)

Study design and evidence →

Example chains from observation to method to system are with the systems under Origin. To the origin of the systems →

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.