High-Coherence Interaction State (HCIS)
When conversation history becomes infrastructure

Long human–LLM conversations can contain extensive history without using that history reliably. A model may understand a reference to earlier material and reuse it purposefully, yet still fail to carry an explicit correction into later behaviour.
High-Coherence Interaction State (HCIS) is a proposed family of longitudinal constructs for studying these differences. It provides a transcript-based method for examining when accumulated interaction becomes usable infrastructure for later turns — and when that organisation remains strong, becomes uneven, or breaks down.
Primary actions
- Read the full Version 2.0 paper on this site: When Conversation History Becomes Infrastructure
- Archival record (Zenodo DOI): 10.5281/zenodo.23322029
- Author ORCID: 0009-0001-9458-7150
What HCIS examines
HCIS separates four observable functions that are often collapsed into broad claims that a model “remembers” or that a conversation is “coherent.”
Compression and unpacking
Does the user rely on accumulated context, and does the model interpret that reliance correctly? Compression here means context dependence, not merely short wording.
Context Deployment
Does the model select earlier material and use it purposefully at a relevant moment to do new work? Recall alone is not enough: timing and function matter.
Correction Persistence
Does explicit feedback change later behaviour beyond the model’s immediate acknowledgement? A correction may persist, decay, require re-establishment, or oscillate across later opportunities.
Drift
How do these functions change across the declared span? A conversation may strengthen, remain stable, fluctuate, deteriorate, rupture, or recover. Drift is a synthesis of change over time, not an additional score.
Why HCIS is a profile, not a single score
The four functions do not necessarily move together. A model can interpret accumulated context accurately while failing to preserve a corrected behavioural rule. It can reuse earlier material frequently without that reuse being helpful or true. A whole-conversation average can also conceal late improvement, deterioration, or a strong but temporary task period.
HCIS is therefore reported as a component profile with a trajectory. The purpose is not to label an entire dyad as simply coherent or incoherent. It is to identify which longitudinal functions are working, which are failing, and how the pattern changes over time.
What the Version 2.0 study found
The revised paper applies the framework retrospectively to five purposively selected, naturally occurring human–LLM dyads ranging from 94 to 2,730 moves. The conversations arose through the participants’ own personal, professional, creative, and research activity. They were not generated for a laboratory task, benchmark, recruitment exercise, or paid study.
Across the five cases, accurate contextual interpretation, purposeful reuse, correction persistence, and temporal trajectory repeatedly came apart. In the clearest contrast, a model correctly interpreted 44 of 48 sampled history-dependent references, while only six of 28 distinct correction targets received persistence credit within the observed period. Another case combined very accurate contextual interpretation and frequent unprompted reuse with one behavioural norm that recurred after nine explicit corrections.
These contrasts show why saying that a model “remembered” is too coarse. Available history, accurate interpretation, purposeful reuse, and lasting behavioural adaptation are related but empirically separable.
What the framework contributes
The main contribution is methodological: HCIS offers a traceable way to examine how accumulated interaction shapes later human–LLM behaviour. The revised framework:
- turns broad claims about coherence into separately inspectable transcript observations;
- distinguishes the presence of conversation history from its demonstrated use;
- separates immediate acknowledgement from lasting response to correction;
- reports temporal trajectories rather than relying only on whole-conversation averages;
- records start conditions and source uncertainty, including project files, imported vocabulary, product memory, and other infrastructure;
- treats naturally occurring long-term interaction as valuable evidence while stating its limits explicitly.
The ecological contribution is the use of interactions generated for the participants’ own purposes rather than constructed benchmark exchanges. This makes long-term conventions, reuse, correction, and breakdown visible in the settings in which they actually arose.
What HCIS does not establish
HCIS describes observable organisation of interaction history. It does not, by itself, establish:
- a hidden cognitive state or memory mechanism inside the model;
- consciousness, shared mind, or equal human and model participation;
- factual truth, safety, benefit, wellbeing, or relationship quality;
- improved creativity, reasoning, cognitive capacity, or task performance;
- a validated threshold separating HCIS from non-HCIS;
- prevalence across users, models, products, or populations.
Those questions require separate measures and controlled comparisons. Version 2.0 is a worked demonstration of feasibility and diagnostic value, not yet a validated instrument.
Evidence and research status
The study uses five focal dyads involving one expert researcher-participant, drawn from a wider developmental corpus generated by two human users. The same naturalistic interactions helped motivate the framework and were later used for its first retrospective application. They are therefore discovery and feasibility cases rather than an independent validation sample.
The analysis is grounded in ordered transcripts, fixed analytic spans, versioned protocols, source checks, scoring ledgers, and recorded adjudication decisions. Full transcripts are not publicly released because they contain personal and third-party material. The paper instead reports aggregate measures and limited de-identified evidence.
Independent reliability testing, controlled removal of relevant history, comparison with simpler measures, replication across users, and studies linking HCIS profiles to outcomes remain future work.
Paper
When Conversation History Becomes Infrastructure: Operationalising High-Coherence Interaction State in Human–LLM Dyads
Anna Wojewodzka (2026), Version 2.0. Zenodo. https://doi.org/10.5281/zenodo.23322029
Suggested citation
Wojewodzka, A. (2026). When Conversation History Becomes Infrastructure: Operationalising High-Coherence Interaction State in Human–LLM Dyads (Version 2.0). Zenodo. https://doi.org/10.5281/zenodo.23322029
Version and priority note
The framework later named HCIS was first publicly formalised on 7 January 2026 as The Third Space as an Interactional Regime in Human–LLM Dyads (Version 1.0), DOI: 10.5281/zenodo.18130088. It was renamed in Version 1.1, published on 30 January 2026 as High-Coherence Interaction State (Third Space) as An Interactional Regime in Human–LLM Dyads, DOI: 10.5281/zenodo.18432386.
Version 2.0 is a substantial empirical and methodological revision. It adds operational definitions, versioned measurement procedures, five-case analysis, explicit inference boundaries, and a programme for independent reliability and validity testing. Later publications do not alter the chronology of the original framework.
High-Coherence Interaction State (HCIS) Originator: Anna Wojewodzka First publicly formalised: 7 January 2026 Current paper: Version 2.0, 12 October 2026 DOI: 10.5281/zenodo.23322029 ORCID: 0009-0001-9458-7150