Interaction Observability & Evaluation for LLM & Multi-Agent AI Systems
Model-agnostic interaction telemetry for analysing trajectory stability, drift,
contradiction, and coordination dynamics in complex AI systems.
Designed to complement existing model-level evaluation metrics with
interaction-level observability (evidence-only layer; no model modification).
HER2NI is a model-agnostic interaction-infrastructure telemetry protocol for coherence dynamics in complex distributed systems.
The protocol is an observability/telemetry layer; it does not prescribe actions or enforce outcomes.
HER2NI: Human–Emergent Resonance to Neural Intelligence
HER2NI is a protocol for representing, encoding, and exchanging coherence metrics between humans and AI systems.
It defines model-agnostic metrics for interaction stability and provides a foundation for safety-aligned human–AI collaboration.
Current Focus: Structural Telemetry Geometry in Human–AI Interaction
Current research examines how HER2NI coherence metrics detect and characterise
the geometric structure of stability, drift, and contradiction in complex,
recursive dialogues. By mapping these dynamics in real time, we aim to measure
and characterise stability, drift, and contradiction patterns to support safer
evaluation and human oversight. The objective is to establish reliable operational
regimes for LLM systems functioning under high context-load or high-stakes operational dialogue,
improving safety, consistency, and the effectiveness of modulation frameworks
used in adaptive dialogue.
Active Lines of Inquiry
Detecting early interaction drift and collapse using Hs(t) trajectories.
Characterising ambiguity and metaphor in cross-linguistic interaction (no private-state inference).
Evaluating tone, pacing, and abstraction adjustments under high-cognitive-load operational dialogue.
Providing metrics & tools to complement existing AI assurance and safety evaluation.
Operational & Evaluation Context
Designed to complement existing AI risk assessment and governance frameworks.
Applicable to single-agent, multi-agent, and human–AI interaction settings.
Focused on early warning signals rather than post-hoc incident analysis.
Intended for research, evaluation, and safety-aligned deployment contexts.
Commercial implementation
H.E.R is the commercial implementation built from HER2NI research: a provider-neutral human-facing intelligence system that can coordinate independent AI routes, preserve route-level evidence and dissent, bind consequential use to authorised human review, and emit independently verifiable records.
As interactive AI systems move from isolated models to networked, multi-agent deployments, new governance and safety challenges emerge that are not addressed by model-level evaluation alone.
Policy Relevance & Governance Context
HER2NI is designed to support existing public-sector and responsible-AI governance
frameworks. It does not introduce new enforcement mechanisms, optimisation objectives,
or automated decision authority.
Instead, HER2NI functions as an observational telemetry layer that can complement
risk-based approaches to AI assurance, safety evaluation, and human oversight.
All interpretation, intervention, and decision-making remain the responsibility
of human operators and institutions.
Adoption is intentionally incremental. Metrics can be introduced experimentally,
evaluated alongside existing monitoring, and discontinued without architectural
disruption. If HER2NI does not provide useful signals in a given context, that outcome
is itself informative.
HER2NI is therefore suitable for case-scoped engagements, research programs, and
early-stage regulatory engagement where increased observability is more valuable
than post-hoc explanation.
HER2NI does not claim authority, certification status, or compliance determination.
It provides observational telemetry intended to inform, not replace, existing evaluation
and governance processes.
Why This Matters
As AI systems increasingly operate as networks of interacting components and agents,
many risks emerge not from individual model behaviour, but from interaction dynamics.
These include coordination breakdowns, cascading failures, over-alignment, and loss
of effective human oversight.
Existing evaluation and governance approaches primarily assess models, outputs, or
downstream impacts. HER2NI addresses a complementary gap by providing
interaction-level observability—enabling earlier detection of instability
before failures propagate or become externally visible.
HER2NI Metrics
HER2NI introduces three core metrics for cross-participant structural alignment, together with an interaction-trajectory representation:
Ss — Model-side structural score: response complexity, constraint retention, and structural load proxy.
Hs — HER structural-fit score: reduced structural fit between participants, constraints, and model response at a given interaction point.
Hs(t) — Interaction Trajectory: the temporal evolution of structural fit across multi-turn interaction.
Together, these metrics enable trajectory-level analysis of stability, drift, and breakdown in human–AI interaction, providing an early-warning and evaluation
framework that goes beyond static outputs or single-turn assessments.
Example Coherence State
The following illustrates a stable, high-coherence interaction scenario
computed using HER2NI metrics at a representative point along an interaction trajectory.
Cs ≈ 0.83 — stable human-side interaction pattern
Ss ≈ 0.91 — high but stable model-side structural load
Hs ≈ 0.88 — strong structural fit
Illustrative example only. Values shown are not derived from a deployed system.
In practice, HER2NI evaluates the temporal evolution of these values as Hs(t)
to assess stability, drift, and breakdown across interaction over time.
The protocol is implementation-agnostic: HER2NI can be integrated into existing LLM systems, agent frameworks,
or evaluation pipelines.
Reference implementations are released commercially as H.E.R on
her2.ai.
Current Public Framing
Some archived research artifacts use earlier terminology; they remain historical research records and do not describe the current H.E.R implementation.
Collaboration
HER2NI Research is open to collaborations with:
Universities and government bodies.
AI safety, legal and evaluation teams.
Researchers studying LLM behaviour under high-load conditions.
Groups interested in cross-linguistic analysis of high-stakes, high-load interaction language (non-clinical).
Institutions exploring coherence-based and trajectory metrics for machine learning and interactive artificial intelligence.
For direct correspondence or research enquiries, please contact
research@her2ni.ai.
HER2NI RESEARCH
