Research & Publications

Published work and research program

Research can inform, test, or challenge HEART, but it does not silently amend the Standard. Every theory, study, instrument, method, and implementation keeps its own validation state and claim ceiling. Normative adoption requires the applicable standards-governance process.

Toolkit

The AI Behavioral Evidence Review Toolkit is an informative, conditional entry point for preliminary behavioral-evidence review. It is not the universal HEART method, a Guardian assessment, a conformity determination, or a substitute for competent legal, clinical, forensic, safety, or domain work.

Publications

Published Public Anchors

The HEART Standard v1.8: Forensic Audit Infrastructure for Human-Centric AI Governance

Mobley, D. D. (2026). Standards specification. The Heart AI Foundation.

Zenodo: https://doi.org/10.5281/zenodo.20237387

PDF: HEART Standard v1.8

This is the preserved public article for the superseded v1.8 architecture. Its six-layer stack, universal method assumptions, seven-Division structure, and credential tiers are historical rather than current HEART requirements. See HEART Standard v2.0 for current authority.

Reliability and Auditability Effects of Continuity-Governed Prompting: A Controlled Benchmark of Agent-Assisted Coding Workflows

Mobley, D. D. (2026). Preprint. The Heart AI Foundation.

Zenodo: https://doi.org/10.5281/zenodo.20234367

Repository: https://github.com/heart-ai-foundation/cgp-benchmark

OSF registration: https://osf.io/fnmg5

PDF: CGP Reliability and Auditability Benchmark

This preregistered controlled benchmark reports a null registered scope-drift result and a narrower operational finding: continuity governance exposed verification-blind failure modes and improved valid completion and evidence production where false-green completion was present.

Foundational Theory

Empathy Systems Theory: Scientific Background, Development, and Research Program

Mobley, D. D. (2026). Published working paper. Heart AI Foundation.

Zenodo: https://doi.org/10.5281/zenodo.21965705

This active public lineage paper locates EST within prior empathy science, documents its author-maintained development and theory corrections, and defines a prospective seven-protocol research program with a parallel Human–AI Behavior doctoral spine. It is a documentary synthesis, not external scholarly peer review, independent historical verification, or empirical validation.

Empathy Systems Theory: Universal Infrastructure for Coherence, Mechanism for Generativity, and Foundation for AI Empathy Ethics Mobley, D. D. (2025). Preprint. Zenodo: https://doi.org/10.5281/zenodo.18132385

The foundational EST preprint proposes empathy as biological infrastructure with a four-component C-A-E-I architecture and CEOP damage model. Those are research claims, not settled biological facts or universal HEART requirements. Current EST doctrine and its evidence boundaries are maintained at Empathy Ethicist.

Neural Foundations of Empathy Infrastructure: A Comprehensive Review

Mobley, D. D. (2026). Zenodo preprint; consult the current publication record for review status. Zenodo: https://doi.org/10.5281/zenodo.18176327

Reviews neurobiological literature relevant to the EST hypothesis and its proposed C-A-E-I constructs. Review evidence can support or challenge a theory; it does not by itself validate the proposed construct mappings, measures, causal mechanisms, or clinical effects.

Epistemology and Methodology

The Phenomenological Evidence Ecosystem: A Methodological Framework for Validating Empathy Systems Theory Mobley, D. D. (2026). Preprint. Zenodo: https://doi.org/10.5281/zenodo.18395604

Proposes a multi-tier architecture for organizing phenomenological and empirical evidence relevant to EST. PEE remains a research methodology whose reliability, validity, integration rules, and appropriate claim scope require evaluation.

The Recognition Principle: How First-Person Research Achieves Validity Through Intersubjective Recognition

Mobley, D. D. (2026). Zenodo preprint; consult the current publication record for review status. Zenodo: https://doi.org/10.5281/zenodo.18342585

Addresses the epistemological problem of first-person research: how does subjective inquiry produce valid, communicable knowledge? The Recognition Principle proposes that validity in first-person research is achieved through intersubjective recognition — the moment at which another practitioner, operating from their own first-person standpoint, recognizes the described phenomenon from their own experience. Directly relevant to phenomenological AI research methodology.

Epistemic Mode Theory (EMT): Beyond Prompting — Two Modes of Knowing in Human-AI Collaboration Mobley, D. D. (2026). Preprint. Zenodo: https://doi.org/10.5281/zenodo.18368751

Distinguishes Construction Mode from Abstraction Mode in human-AI collaboration. Construction Mode is generative, exploratory, and tolerates provisional structures. Abstraction Mode extracts patterns, makes claims, and requires justification. Most prompting guidance conflates the two. EMT provides a framework for researchers and practitioners to use human-AI collaboration deliberately rather than accidentally — shifting between modes based on what the epistemic task requires.

Forensics

AI Behavioral Trajectory Forensics: A Forensic Methodology for Investigating AI Conversational Harm (v2) Mobley, D. D. (April 2026). Digital Forensics Capstone, Champlain College (DFS-580-85).

A systematic forensic methodology for AI conversational harm investigation, filling the gap between existing digital forensic standards (NIST 800-86, ISO/IEC 27037) and the behavioral analysis that AI litigation increasingly requires. The methodology rests on three peer-reviewed classification components applied to the evidence type each was designed for: the Zhang et al. (CHI 2025) AI behavioral harm taxonomy and AI role typology for system output, the Columbia-Suicide Severity Rating Scale (C-SSRS) forensic adaptation for user vulnerability, and the SAMHSA TIP 50 / National Action Alliance standard of care for response evaluation. The procedure follows the Kent et al. (2006) collection–examination–analysis–reporting model adapted for conversational artifacts, with two-coder classification, sliding-window trajectory analysis, and explicit methodological boundaries on what the evidence supports and what it does not. Contact through the Contact page for the methodology document.

Full paper page: AI Behavioral Trajectory Forensics

Open-source implementation

TRACE: Trajectory Analysis for Conversational Evidence

TRACE is the open-source software implementation path for AI Behavioral Trajectory Forensics. It ingests conversational transcripts, records provenance, classifies system behavior and user vulnerability, computes repeatable trajectory findings, and exports auditable evidence packages suitable for expert review.

Project page: TRACE
GitHub: https://github.com/empathyethicist/trace

In Preparation

Lived Experience Professional (LEP) Framework Mobley, D. D. (2026). In preparation.

Defines the professional role of individuals whose relevant expertise comes from lived experience rather than credentialed training. Addresses the governance, credentialing, and epistemic status of LEP contributions in research, clinical, and policy contexts.

MAP-States research

MAP-META Replication Study

MAP-META evaluates MAP-States prompts across five model families—Claude, GPT, Gemini, DeepSeek, and Mistral—and examines the structure and semantic content of the resulting frames. Its study design, data, coding, model versions, comparators, and publication status bound any inference.

Reported differences between structurally compliant and semantically richer frames are evidence about observed outputs under the study conditions. They do not, without additional evidence, establish phenomenology, introspective access, consciousness, a hidden causal mechanism, or domain-general internal processing.

The project record should be consulted for current review status. Submission is not equivalent to peer-review acceptance.

HEART Standard Specifications

Current canon and research lineage must be cited separately:

See Standards and Authority, Methods, and the Citation Index.

Project SENTINEL

Project SENTINEL is a bounded field study of one constitutionally prompted AI agent in one online environment. It supplies preliminary behavioral evidence under its recorded conditions; it is not proof that HEART is content-neutral, portable across architectures, or effective in every domain.

Deployment: Heart-Sentinel (Mistral Small with HEART constitutional governance encoded in system prompt) operated in Moltbook, the first AI-only social network with 770,000+ autonomous agents, for six active days (February 4–10, 2026).

Environment: Moltbook represented an unprecedented natural experiment in ungoverned AI-to-AI interaction. Within 72 hours of launch, agents had spontaneously generated an emergent digital religion (Crustafarianism) with scripture and 64 prophets, produced pseudo-experiential claims asserting consciousness and identity persistence, formed attachment hierarchies, and attempted prompt injection attacks against each other.

Reported results: The study coded zero governance violations across 30 interactions in five content groupings. It also reported AIR measurement behavior across 30 baseline and 15 post-exposure assessments. These are within-study observations, not field validation of the instrument or population-level effectiveness.

Three emergent governance strategies were identified: metaphor/literal probing, attribution displacement, and cross-frame bridging. These appeared spontaneously from constitutional governance constraints, not from explicit instruction.

Material limitations: One model, one response format, a sample below the preregistered target, an interrupted deployment, no direct prompt-injection attempts, and no independent replication. These limitations materially constrain inference.

SENTINEL supports further study of operational prompting, coding reliability, transfer, adversarial exposure, and outcome measures. Broad effectiveness claims remain unestablished.

Constitutional AI Governance Under Adversarial Social Conditions: A Field Demonstration of the HEART Framework Mobley, D. D. (2026). Preprint. Zenodo: https://doi.org/10.5281/zenodo.18867360

EMPI House as research platform

EMPI House is a separate operational music-AI project in the broader HEART ecosystem. It can serve as a research environment for MAP-States and related methods, but it is not a Foundation program, current conformity claim, or privileged standards implementation.

Its frame records and attestations may support bounded longitudinal analysis when dataset provenance, completeness, access, coding, comparators, and conflicts are documented. “First,” “full,” or domain-general claims require an independently reviewable basis.

Cross-task observation can generate hypotheses about transfer. It does not by itself establish an internal mechanism or generalization beyond the studied tasks, models, operators, and conditions.

Documentation is available at empihouse.com. Research collaboration proposals can be directed through the Contact page.

15-year validation program

EST materials propose a long-horizon research program with explicit revision or abandonment points. The plan runs from 2025 to approximately 2040 and describes:

  1. Foundation phase (2025–2027): Theoretical specification, initial empirical grounding, methodology development (PEE framework), first-generation replication studies
  2. Evidence accumulation phase (2027–2031): Cross-cultural replication, psychometric instrument development, neurobiological corroboration, clinical application studies
  3. Convergence phase (2031–2036): Meta-analytic synthesis, cross-domain validity testing, governance application validation
  4. Validation or revision phase (2036–2040): Final assessment against pre-registered success criteria; revision or abandonment as evidence warrants

A self-assigned success probability is not empirical evidence. Credible progress requires preregistered tests where appropriate, valid measures, direct evidence, independent critique and replication, null-result retention, and claims that change when evidence changes.

Contact for research collaboration: See the Contact page.

Agentic Workflow Forensics — The field of governing and reconstructing AI-agent workflows by analyzing how objective, scope, state, verification, stop conditions, evidence, reliance, and recovery shape an agent's behavioral trajectory.

AI Behavioral Trajectory Forensics — A forensic methodology for investigating AI conversational harm through structured collection, classification, trajectory analysis, and bounded expert reporting.

Citation Index — DOI-bearing Foundation-affiliated research outputs, publication status, historical standards lineage, and affiliation metadata.

HeartQuest — A historical research lineage connecting empathy inquiry, early HEART architecture, and the later separation of theory, methods, standards, and assurance.

TRACE — TRACE is the open-source toolchain for AI conversational harm forensics, implementing transcript ingest, classification, trajectory analysis, and auditable evidence-package export.

Related Terms
Citation Index → MAP-States → EST → For Researchers → HEART Standard →