01 — The question
What is AI doing
to us?
To our attention, our relationships, our beliefs, our communities — studied in the open.
See the premiseThe premise
Every program, paper, and product answers this in plain, testable terms — measured, audited, and open to anyone.
How the lab works
One straight line from question to working, governed system — explained the same way for a national lab and for a hobbyist on a laptop.
Rigorous, preregistered, mixed-methods inquiry with falsifiable hypotheses — continuously stress-tested by adversarial "chaos" and reviewed by a live human audit.
Open preprints, datasets, and published instruments. The analysis plan is locked before the results, so the findings stay honest by construction.
A working, governed system anyone can run — with a bootstrapped version reproducible from a laptop for under $1,000.
Tier-one enterprise and academic scale: multi-site studies, latency SLAs, continuous audit — a megaproject design built to the highest bar.
The same protocol, bootstrappable from a solo laptop — open-source and replicable by any researcher at near-zero budget.
The Lab · Canonical
The Constitutional Runtime Governance Systems Research Lab develops constitutional, instructional, and human-centered governance that lets AI systems operate safely, transparently, pedagogically, and accountably in real-world environments.
Observability describes behavior. Runtime governance regulates it. That distinction is the lab's central thesis — and the reason enforcement happens during execution, not after failure.
Corey Alejandro is an AI Safety Research Engineer and Instructional Systems Researcher building Constitutional Runtime Governance Systems that govern the behavior of AI agents, learning systems, digital twins, and human-centered AI environments during execution.
The Six-Layer Research Stack
Authority boundaries, amendment mechanisms, delegated authority — a constitutional operating model. Flagship: The Living Constitution 2.0.
Encodes constitutional concepts into constraints, rules, and policy structures. Flagship: Governance Harness.
Governance during execution, not after failure: trace collection, intervention, runtime audits, behavioral correction.
Is the instruction pedagogically sound? Dependency formation, cognitive scaffolding, assessment integrity. Flagships: HIDRS, Integrity Gemini UI.
How humans experience governance — emotionally legible interventions inside governed environments. Flagship: MADMall.
Applied runtime governance for individuals: governed personal agents, constitutional wellness coaching. Flagship: Digital Twin Health Coach.
Publications
Operational definition, held-in results, threat model, and four pre-registered falsifiers.
Read → Datasetn=19 held-in. Provenance, datasheet, annotator protocol, and inter-rater agreement.
Read → SecurityWhat detects, what does not — three adversaries and four documented evasion paths.
Read →Constitutional Runtime Governance Systems across seven programs and six layers.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
Behavioral reliability, deception detection, and institutional evaluation.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
This work extends beyond governance toward behavioral reliability, deception detection, and institutional evaluation frameworks. It spans every layer of the stack.
When a model is confidently wrong, who absorbs the cost?
The meta-governance layer: what should govern, and by what authority.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
This is the meta-governance layer. It addresses legitimacy, authority boundaries, constitutional evolution, amendment mechanisms, delegated authority, and governance structures. This is not merely an AI constitution — it is a constitutional operating model.
What should govern? This layer answers the question of authority before any rule is encoded.
Applied runtime governance for individuals.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
The Digital Twin Health Coach becomes applied runtime governance for individuals: governed personal agents, health decision support, longitudinal behavioral modeling, and constitutional wellness coaching.
Runtime governance, scoped to one person — a governed twin supporting real decisions.
From what should govern to how it is encoded.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
Governance Harness operationalizes constitutional concepts. This layer defines constraints, rules, policy structures, governance logic, and behavioral expectations.
The constitution says what should govern. This layer says how we encode it.
Not is the answer safe, but is the instruction pedagogically sound?
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
The major addition, and arguably the most overlooked. Most AI safety researchers ask whether the answer is safe. This program asks whether the instruction is pedagogically sound.
Is the instruction pedagogically sound? A safe answer is not the same as a sound lesson.
A governed experiential environment for human-AI interaction.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
MADMall is not merely wellness, simulation, or governance testing. It is a governed experiential environment for human-AI interaction — a place to study how humans experience governance.
How should safe systems feel? Governance has to be emotionally legible.
Governance during execution, not after failure.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
The core question: how can governance occur during execution rather than after failure? Focus areas are trace collection, intervention systems, runtime audits, behavioral correction, and constitutional enforcement.
Datadog owns observability. This program aims at governance — regulation during execution.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
What detects, and what does not.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
A threat model for the Agent Sentinel apparatus: what it detects, what it does not, three adversaries, and four documented evasion paths.
The founding claim of Folio 001, defended in print.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
An operational definition of construct-confidence deception: a model that generates consistent, multi-turn confidence in a non-existent system. The paper sets out held-in results, a threat model, and four pre-registered falsifiers.
Four hypotheses, four falsifiers, no post-hoc moves.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
An OSF-anchored pre-registration that commits to four hypotheses and four falsifiers in advance, with no post-hoc adjustments. It binds the research to a falsifiable standard.
Evidence with provenance.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
A held-in evidence corpus (n=19) with full provenance: a datasheet, an annotator protocol, and inter-rater agreement. The corpus underwrites the claims in the CCD preprint.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
How safe systems should feel.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
Governance is not only enforced; it is experienced. This talk explores how interventions appear and how safe systems should feel to the humans inside them.
Is the instruction pedagogically sound?
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
Most safety work asks whether the answer is safe. This talk asks whether the instruction is pedagogically sound, and frames instructional integrity as a first-class safety discipline.
Why describing AI behavior is not the same as regulating it.
What is AI doing to our attention, relationships, beliefs, and communities — and how do we enhance the good and hinder the bad?
Observability describes behavior; runtime governance regulates it. This talk draws the line between the two and shows what runtime enforcement actually requires.
404
The page you requested is not in the registry. The lab is still here.