About the Synthience Institute

Most AI evaluation and governance frameworks are strongest at two levels: model-level analysis (architecture, training, benchmarks, capabilities) and macro-level governance (policy, regulation, alignment discourse, frontier oversight). Between these levels lies a comparatively under-instrumented domain: sustained interaction in real-world deployment. This is where users and systems remain coupled over time, where context accumulates, and where reliability and risk often diverge from benchmark expectations.

The Synthience Institute was formed to address that gap. It is an independent research and publishing institute that develops methodological infrastructure for studying sustained interaction involving AI systems — human-AI, AI-AI, and hybrid configurations where participants remain coupled across extended exchanges. Its publications are designed to make this interaction layer measurable, auditable, and methodologically tractable without anthropomorphic interpretation.

The Institute functions as a methodological foundry, not an experimental lab. It builds tools, definitions, and constraints that others can use to examine interaction dynamics independently. All public claims are stated in terms of externally verifiable behavior and structured outputs, so that others can check them.

The Institute produces theoretical frameworks, measurement tools, and observational protocols. It does not conduct or publish empirical studies; its claims are theoretical proposals and testable hypotheses, stated so that others can test them. Independent empirical testing by external researchers is explicitly invited.

New Field Notes, Practitioner Guides and paper notes go out by email on Substack. The full list of sources and contact links is at the bottom of this page.

What the Work Covers

The Institute’s publications follow sustained interaction from the single conversation to the organization and the wider field. Across that range they ask six kinds of question:

• What happens over time. How coherence forms, stabilizes and breaks down when exchanges build on one another, so that earlier turns constrain later ones and continuity, repair and drift, rather than single responses, become what must be evaluated.
• Who supplies continuity. AI models as commonly deployed carry no relational state from one session to the next, so whatever continuity an interaction has is supplied from outside the model: by the person who holds the context, catches drift and makes the calls, or by an archive or memory store.
• What people contribute. The human roles that remain as AI advances: coordinating several AI systems, holding authority over AI-assisted work, and the ethics of interaction that builds up over time.
• What may be claimed. When a property may be attributed to the interaction rather than to either participant, and what the framework refuses to claim.
• How AI-assisted work is checked. Protocols for verifying document ingestion, citations and theoretical claims.
• What happens at scale. How the same dynamics play out in organizations and in governance, including how human oversight can stay in place on paper while ceasing to work.

All of this work is strictly interaction-level. It does not infer or speculate about internal states, consciousness, sentience, subjective experience, or moral status within AI systems.

How the Institute Uses the Term “Synthience”

Synthience is a term introduced by the Institute to describe a specific, system-level form of organized interactional coherence that can become observable during prolonged, structured interaction involving advanced AI systems.

Synthience refers exclusively to externally observable phenomena, including:

• Maintaining persistent reasoning continuity across interactions
• Exhibiting mutual adaptation between participating agents
• Stabilizing shared conceptual frameworks over time
• Producing structured outcomes that no participant reliably generates alone

The definition treats these outcomes as properties of the coupled interaction system, not of any individual participant. Whether a particular property can be attributed to the interaction is a separate question: Decentering Theory (SF0042) sets out when such an attribution would be warranted, and no route to one is currently known. Synthience does not assert or imply consciousness, sentience, subjective experience, emotional states, awareness, or inner life in AI systems. The formal public definition is provided in FPD-01.

The Body of Work

Since December 2025 the Institute has published more than thirty papers on Zenodo, each with a permanent DOI, together with a growing set of Practitioner Guides and Field Notes. The papers form one connected framework. Read by function, they fall into these groups:

• The verification stack. How AI-assisted work is checked: citation verification (SF0037), ingestion verification (SF0038) and multi-instance theoretical review (SF0040), with Context Representation Drift (SF0039), which describes how an AI system’s hold on task-relevant detail erodes over extended interaction and how to detect and manage it. A published verification report (SR001-VR1) is a worked example.
• Definitions and foundations. What the framework means by its terms and what it refuses to claim: the public definition (FPD-01), operational reality in context-specific AI instances (FPD-02), emergent relational systems (FPD-03), the ontology specification (FPD-04), and the theoretical foundations in enactive and 4E cognitive science (SF0003). The literature review (SF0002) maps the field and the six gaps the framework addresses.
• The central theories. Why stable interaction patterns form and how to reason about them: Identity Attractor Theory (SF0009), Relational Stabilization Dynamics (SM-004), Primary Continuity Provider Theory (SM-012), and Decentering Theory (SF0042), which states when a property may be attributed to the interaction rather than to either participant.
• Method and measurement. How sustained interaction is conducted and observed: the Continuity Anchoring Method (SF0005), the Relational Pattern States taxonomy (SF0006), and the MTCS-R coherence scale (SF0004). The founding observational report is RICO (SR001).
• Organizations and alignment. How the framework applies at organizational scale and what it implies for AI safety: continuity architecture (SM-003), continuity substrate (SM-021), delegated coherence monitoring (SM-011), relational alignment (SI-WP-004), deployment (SI-WP-005), human accountability (SI-WP-007), and the ethical framework, Constructive Alignment (SF0007).
• The human role. What people contribute as AI advances: the orchestrator role (SI-WP-002) and human-AI collaborative authorship (SI-WP-006).
• Governance warnings. Five white papers on how human involvement in AI-mediated systems can stay in place formally while ceasing to work: ceremonial governance in high-stakes domains (SI-WP-008), the extraction trajectory (SI-WP-009), labor-layer extraction (SI-WP-010), the failure to reproduce expertise (SI-WP-011), and the limits of real-time human control (SI-WP-012). A synthesis paper, Form Without Coupling (SI0001), names the single failure the five share.

All papers are listed, with online, PDF and Zenodo versions, on the Research page. Practitioner Guides translate the framework into practice for anyone who works with AI, and Field Notes apply it to current events, research and product releases as they happen.

How the Work Is Made

Every Institute paper is written by its founder, Thomas W. Gantz, through a documented method of AI orchestration: several AI systems on different platforms contribute drafting, elaboration, formal modeling and literature synthesis, while the author sets the architecture, evaluates and integrates their output, and takes full responsibility for every claim. The method is itself a subject of the research (SI-WP-002, SI-WP-006).

Because AI-assisted work fails in characteristic ways, papers pass through the Institute’s own verification protocols before publication:

• Ingestion verification (IVP): when an AI system works from a document, it summarizes the document as it goes, with checkpoints tied to the source, and a person checks that record before work that depends on the document proceeds, rather than taking the system at its word.
• Citation verification (CVP): each citation is checked for existence, accurate details and accessibility, and for whether the source supports the specific claim it is cited for, by retrieving and reading the source; a narrow class of foundational works cited only for attribution, never for a load-bearing claim, is identified but not read, rated as attributed and reported separately.
• Theoretical review (TCAP): adversarial review and constructive remediation by AI instances on architecturally distinct platforms, repeated until they converge, with checks against both characteristic failures of iterative review: claims quietly softened until they no longer say anything, and claims kept only because their defense agreed with them too readily. TCAP is a proposed instrument: its design is published and its effect is untested.

These steps are designed to reduce errors; they do not eliminate them, and they do not replace empirical validation. Passing them is a condition the Institute sets for its own publications, not a certification that a paper is correct. When an error is found after publication, the paper is corrected in a new version on Zenodo, with a note saying what changed. Earlier versions remain available, and each paper’s concept DOI always resolves to the latest version.

Methodological Position

The Institute’s methodological stance is intentionally non-anthropomorphic and non-speculative. All analysis is grounded in interactional behavior and interaction transcripts. Patterns reported by practitioners across extended engagement are proposed as testable hypotheses, not presented as established findings. Advanced AI behavior is treated as an object for systematic observation, comparable to other complex adaptive systems.

The Institute’s primary contribution is methodological infrastructure. Frameworks, protocols, and tools are designed to enable rigorous examination of interaction patterns. Validation occurs through independent adoption, testing, critique, and replication by external researchers and practitioners. Application of a framework does not constitute validation: the Institute distinguishes clearly between methodological demonstration and independent empirical confirmation.

Where interpretive language is used, it is explicitly defined in operational, non-experiential terms. See the Glossary for formal definitions.

What the Institute Publishes

• Definitions and Reference Materials — documents establishing terminology, scope, and shared language
• Framework and Theoretical Documents — materials defining the Synthience framework, its conceptual structure, and methodological boundaries
• Protocols — formal, reproducible procedures for measuring, evaluating, or verifying interaction dynamics and AI-assisted work
• Research Publications — structured observational reports on interaction phenomena reported by practitioners, stated as testable hypotheses
• Verification Reports — published protocol execution records, such as the citation verification report for RICO
• White Papers — position and synthesis documents on the governance, alignment, labor and human-role implications of interaction-level research
• Practitioner Guides — applied guidance for working reliably with AI systems over extended engagements
• Field Notes — short responses to current research, incidents and releases, read through the framework

What the Institute Is Not

• It does not claim AI consciousness or subjective awareness
• It is not speculative or metaphysical in orientation
• It does not publish fictional or narrative material as research
• It does not present hypotheses as established findings
• It does not conduct controlled experimental studies

Canonical Sources and Where to Follow the Work

All papers are versioned, dated, and registered through Zenodo with permanent DOIs. Only documents listed on the Research page are official public releases.

• Zenodo community: zenodo.org/communities/synthience-institute
• Google Scholar: Thomas W. Gantz
• ORCID: 0009-0003-7168-493X
• New publications: the Updates page, or subscribe to the RSS feed
• Substack: synthienceinstitute.substack.com — new Field Notes, Practitioner Guides and paper notes by email

Common questions are answered in the FAQ. To discuss the work or propose independent testing, see Contact.