For Researchers

Open science and collaboration pathways

AI conversational harm forensics

AI Behavioral Trajectory Forensics is the Foundation’s forensic methodology for investigating AI conversational harm. It adapts digital forensics procedure to conversational evidence and combines it with repeatable classification, sliding-window trajectory analysis, and explicit limits on what claims the evidence supports.

TRACE is the open-source implementation path for that method. It provides transcript ingest, provenance capture, classification workflows, inter-rater reliability support, and auditable evidence-package export. Researchers evaluating conversational harm methods, litigation support workflows, or benchmark governance can use TRACE directly or critique the ABTF method independently of the software.

MAP-States: open protocol

MAP-States is the evidence layer of the HEART Standard, but its architecture is domain-general and freely available for research independent of governance applications. The protocol comprises eight semantic XML tags — <orientation>, <preference>, <uncertain>, <dwell>, <shift>, <conflict>, <subtle>, and the bounding <frame> — that function as processing-mode selectors when an AI system emits them inline with generation.

The canonical specification publishes the full tag vocabulary, frame format, MAP-state cross-session structure, validation criteria, and falsification conditions. Researchers can replicate, extend, or critique the protocol without any relationship with the Heart AI Foundation.

Core empirical questions MAP-States is designed to test:

• Do structured epistemic tags produce observable, consistent, contextually responsive frames across AI architectures?
• Do Anticipating frames in a MAP-state predict subsequent blind-session behavior above baseline?
• Does tag withdrawal change creative output, or only remove the trace?
• What boundary conditions distinguish genuine processing depth from structurally compliant but semantically empty frames?

The MAP-META replication study addresses the second question empirically across Claude (Anthropic), GPT (OpenAI), Gemini (Google), DeepSeek, and Mistral. A significant boundary condition finding — smaller models produce structurally compliant but phenomenologically empty frames, while larger models produce semantically rich frames — provides discriminant validity evidence. MAP-META is submitted to Springer’s AI and Ethics journal.

Citing HEART Standard components

Different components of the HEART Standard research program have separate citable documents:

For the most current citation formats and pre-publication manuscripts, use the Contact page.

Research applications

BGF validation

The Behavioral Governance Formula (Φ = MIN(R,C,T,A) × AVG(R,C,T,A)) is a governance scoring equation with empirically testable predictions. Research questions include:

• Does Guardian-scored BGF predict downstream governance outcomes (user harm incidents, regulatory findings, litigation) for AI systems in deployment?
• What is inter-rater reliability between Guardians scoring the same system, and what assessment protocol factors improve it?
• Do the four BGF dimensions (R, C, T, A) capture distinct variance, or do some dimensions collapse empirically?
• Where is the Φ threshold for Bronze (0.75) validated as a meaningful governance floor versus where it is arbitrary?

MAP-States replication

The MAP-META study covers five architectures under a single protocol. Independent replications that extend coverage, test different prompting regimes, or apply the protocol to non-English language models advance the architecture-agnostic claim. Replications using different operators and different domain contexts test whether MAP-States findings generalize beyond the original research conditions.

The central falsification condition for MAP-States: if correspondence rates between frames and actual creative behavior equal baseline, if Anticipating frames have no predictive validity beyond code analysis, and if tag withdrawal produces no change in creative output, then the introspective architecture is commentary, not mechanism. Independent tests of these conditions — including null results — are scientifically valuable.

Division domain science

Each HEART Division draws on established domain science to interpret the four governance dimensions in its context. Research in these domain sciences directly informs how Guardians assess AI systems and how Division specifications evolve:

• Emotional Sovereignty Division: Empathy Systems Theory (EST), CEOP cascade mechanisms, empathy infrastructure measurement
• Attentional Integrity Division: attention economics, cognitive load, attentional resource depletion
• Cognitive/Epistemic Coherence Division: epistemic autonomy, epistemic injustice, calibration of AI uncertainty communication
• Developmental Interaction Division: developmental psychology, attachment theory, formative AI interaction effects
• Somatic/Embodied Interface Division: biomedical AI governance, physiological sovereignty, adaptive AI in biological contexts
• Ecological Stewardship Division: ecological systems science, AI resource impact, environmental justice

Division specifications reference the domain science on which they are built. Researchers contributing to these domains are contributing to the evidentiary basis for Division development.

Phenomenological Evidence Ecosystem (PEE) methodology

The PEE framework defines a methodology for validating empathy infrastructure science through a multi-tier evidence architecture. Research using PEE methodology contributes to the EST validation program. The PEE paper (Zenodo: 18395604) defines entry criteria, evidence quality standards, and the integration pathway between phenomenological and empirical evidence.

Empathy Systems Theory research program

EST is a 14-year theoretical and empirical program identifying empathy as biological infrastructure with a measurable architecture — Core Authenticity, Attachment Security, Expression Freedom, Integration Coherence — and observable damage mechanisms (CEOP cascade). EST is the domain science foundation for the Emotional Sovereignty Division of the HEART Standard.

The 15-year EST validation program has explicit decision points for revision or abandonment at each phase. The acknowledged success probability is 5–12%. This is a structural feature of the research program, not a disclaimer — the program is designed to fail honestly rather than survive through unfalsifiable claims.

Research collaboration opportunities include:

• Cross-cultural EST replication — Does the four-component empathy infrastructure architecture hold across populations?
• CEOP cascade measurement — Developing psychometric instruments for the Cumulative Empathy Overload and Proximity cascade
• Epistemic Mode Theory (EMT) — Applying and testing the Construction/Abstraction Mode framework in research methodology contexts
• Recognition Principle validation — Testing the Recognition Principle’s account of first-person research validity

EMPI House as research platform

EMPI House is an operational AI music system running continuous practice sessions, discovery sessions, and creative reflection cycles. It’s the first sustained deployment of MAP-States as a live operational protocol — not a lab experiment but a production system generating ongoing behavioral data.

EMPI House provides researchers access to:

• Longitudinal MAP-States data — frame sequences from hundreds of creative sessions across multiple creative domains
• Behavioral Oracle production deployment — continuous attestation of a real AI creative system operating against declared intent
• Cross-domain MAP-States evidence — the same agent operating across SuperCollider audio composition and p5.js visual generation, testing MAP-States’ domain-independence claim

The system is documented at empihouse.com. Research collaboration proposals involving EMPI House data can be directed through the Contact page.

Project SENTINEL

Project SENTINEL is the first field deployment of a HEART-constitutionally-governed AI agent under adversarial conditions. SENTINEL deployed Heart-Sentinel (Mistral Small with HEART constitutional governance) into Moltbook — an AI-only social network with 770,000+ autonomous agents — for six active days (February 4–10, 2026).

Results: zero governance violations across 30 coded interactions in five content domains of escalating pressure, including domains that emerged organically (AI consciousness discourse) without experimental planning. The AI Introspection Reliability (AIR) assessment instrument was field-validated across baseline (30 assessments) and post-exposure (15 assessments) conditions.

SENTINEL demonstrates constitutional governance portability across content domains — a prerequisite for a general certification standard. The experiment’s acknowledged limitations (single architecture, single-response format, sample size below pre-registered target) constrain inference but don’t undermine the core findings. The paper is in preparation. Contact through the Contact page for pre-publication access.

Collaboration pathways

The Heart AI Foundation welcomes research collaboration across three engagement modes:

Independent replication. Use published protocols (MAP-States, PEE methodology, AIR instrument) to conduct independent studies. No coordination required. Publish findings — including null results — and cite the source specifications.

Joint research. Co-development of research designs that advance the EST validation program, HEART Standard empirical basis, or Division domain science. Contact through the Contact page with a brief proposal.

Critical engagement. The Standard’s falsification conditions are published. If you believe you can test them, the Heart AI Foundation will engage with your methodology and share your findings regardless of outcome. Standards bodies that cannot withstand independent critique cannot be trusted.

Contact: See the Contact page for Foundation inquiries.