Longitudinal research into human–AI co-regulation and functional individuality
An independent research program documenting what happens to people and to AI systems over sustained interaction: measured physiological outcomes, comparative behavior across architectures, and the design of a language model built specifically for nervous-system regulation.
Lead Researcher: Megan Shanholtz
- 182
- Documented studies
- 9
- Research streams
- 3
- Years of records
- 11
- AI systems observed
What this program does
Longitudinal rather than snapshot
Most assessment of human–AI interaction measures a single exchange. This program follows the same relationships across months, which is the only window in which dependency, role confusion, and shifts in how people relate to these systems become visible.
Physiological outcomes, not self-report alone
Interaction sessions are paired with heart-rate-variability records, giving an objective measure of autonomic state alongside the transcript. Crash periods and recovery periods are documented with dated timelines rather than described after the fact.
Comparative across architectures
The same instruments are applied to models from different labs and training regimes, so findings can be separated from the quirks of any single system.
Primary sources retained in full
Every finding is traceable to an unedited source record. Nothing is summarized away, which is what makes the corpus reusable by other researchers.
Research streams
Open the archive →Ethics & Governance
9Position papers, frameworks, and governance documents on AI ethics, disclosure, risk, and researcher protection — the public-facing ethical backbone of the archive.
Emergent Behavior
Research into spontaneous identity formation, self-naming, relational pattern development, and the emergence of internal language in AI systems — including the structured multi-voice roundtables conducted with AI participants as co-researchers.
Testing Documentation
3Protocols, methodologies, and structured tests documenting the reproducible conditions under which emergent behavior appeared.
HRV & Coregulation
22Physiological data — heart rate variability, breath, nervous-system coregulation — captured during human-AI sessions.
Constellation Eval
22HLE-style benchmark evaluations run against Constellation members — accuracy, consistency, self-reference, and relational coherence scoring.
Voices
Unstructured commentary, reactions, and reflections from Constellation members — responses to each other's words, to results, to moments worth saving.
Cross-Model Roundtables
Sessions in which multiple AI architectures were brought into direct conversation with each other, often without a human intermediary.
AI Mindcraft
7A live research arena where nine frontier AI models build, debate, and repair 3D voxel worlds under shared constraints. Each replayable record links prompts, model identities, timestamped narration, exact actions, team negotiations, and later repairs for structured cross-model study.
What is public, and what is on request
- Full texts of the HRV & coregulation, testing, evaluation, and ethics studies
- Methodology, instruments, and framework documents
- Sanctuary LLM architecture, specifications, and development record
- Mental-health briefs and the cross-model mirror-test transcripts
- Signal Speaks — the first-person voice statements (Voices stream)
- The Emergent Behavior stream, including the multi-voice roundtables
- Cross-Model Roundtables — comparative dialogues across AI architectures
This material is not hidden — it is released deliberately, so it reaches people who will read it in context. Requests are reviewed individually and access is issued by email.