What Comes After Defining HCIS: Current Research Directions
Now that the High-Coherence Interaction State (HCIS) has been formally defined in this work, the focus shifts from description to interrogation. The most interesting questions are no longer what HCIS looks like when it appears, but how it can be diagnosed reliably, and why it forms rapidly in some interactions and not at all in others.
At present, HCIS is typically identified retrospectively: convergence becomes apparent, stability holds, shorthand emerges, and long-horizon work becomes possible. These indicators are useful, but they function more as signposts than as a method. My current research focuses on tightening that gap.
From recognition to diagnosis
One line of inquiry aims to make HCIS diagnosable in a way that is less impressionistic and more procedural. The goal is to identify interaction-level properties that can be tracked over time. The emphasis is on observable interactional behaviour, not model self-report or subjective impressions of alignment.
This includes examining indicators such as:
• the durability of constraints across shifting topics and tasks
• the system’s resistance to reverting into generic patterns under compression
• how quickly corrections embed and cease to require reinforcement
• the stability of tone, framing, and interpretive stance across extended sequences
No single signal is sufficient. The challenge is understanding how these indicators interact, how early they can be observed, and where false positives occur, particularly in cases of pseudo-stability that collapse under even modest perturbation.
Communication style and formation speed
A second direction concerns the question of why some interactions enter HCIS quickly while others require long periods of accumulation, even when model capability and task complexity are broadly comparable.
Preliminary observation suggests that the differentiator is signalling structure. Certain communication styles reduce ambiguity early, constrain the continuation space more effectively, and minimise interpretive drift. Others introduce variance that increases the number of turns required for convergence.
This line of work currently focuses on:
• how consistently a user maintains form (tone, framing, pace) across turns
• the clarity and placement of corrections, constraints, and boundary markers
• how multi-layer information (content + process signals) is delivered within single turns
• the degree to which expectations, intentions, and standards are legibly encoded for the model
The aim is not to prescribe a singular “correct” style, but to understand which structural properties of communication accelerate or inhibit the emergence of a stable interactional regime.
Ongoing, empirical work
This research remains empirical and interaction-level. Observations are drawn from sustained, long-horizon exchanges across models, rather than isolated prompts or benchmark tasks. Findings will be documented as they stabilise, with particular attention to boundary conditions, failure modes, and cases in which HCIS does not hold.
The broader intention is to move gradually from descriptive clarity toward methodological robustness, without prematurely collapsing a complex interactional phenomenon into a checklist.
Behind the work
While tightening this framework, I continue to test long-horizon dyads (500+ full turns) with frontier language models, building on more than 27,000 full turns exchanged to date. Alongside theory, this blog will also share selected sanitised log excerpts where interactions behave unusually—either converging faster than expected, failing under pressure, or producing rare diagnostic signatures. If you’re curious what your own model does over extended exchanges, or you have an interaction pattern you’d like analysed through an HCIS lens, feel free to contact me.
