Research Focus

The runtime alignment problem: how do many AI systems coordinate without losing trust, attribution, or accountability? Our research examines that question from the architecture side, not the training side.

Research Focus — seven runtime alignment threads
Research focus map

Premise

Most AI safety work in 2026 focuses on what a model does before deployment — alignment during training, RLHF, safety evaluations. That work is necessary and important. But it doesn't answer a question that's becoming urgent: when many AI models from many vendors operate together, drift from one another, fail in different ways, and coordinate without a shared training signal — what holds the system honest?

Our answer is architectural. The governance has to be in the structure: in the verification interfaces, the audit chain, the constitutional constraints, the deterministic state machine that routes between lanes. That's the research program.

Focus Areas

Active research threads in the lab. Each is grounded in working code, real audit trails, and one or more of our provisional patent applications.

Ingress
Runtime Alignment Verification
The Light Guard / Heavy Guard pattern: every input to the constellation is anchored, identified, and verified against constitutional constraints before any lane reasons on it. Dual-layer design — a low-latency in-process check with a millisecond-scale budget plus async deep verification — preserves request-path latency while permitting arbitrarily deep audit. Measured latency distributions to be published on the Benchmarks page.
Anti-Consensus
Consensus Drift Detection
Lanes that share model families, prompts, or topics can converge over time even when they're supposed to disagree. We measure agreement rate, diversity index, semantic drift, and role-adherence across rolling windows of anchors; we flag convergence that exceeds healthy bounds. The constellation needs structured disagreement to think.
Identity
Behavioral Baseline Fingerprinting
A stable reference point per lane — built from probe-suite response vectors, stylometric profiles, and behavioral hashes — that defines the lane's attested-good state. Drift from baseline becomes a measurable signal. Baselines update only on explicit, CPN-approved refresh, never silently.
Attestation
ZK-STARK Identity Attestation
New or unknown AI agents joining the constellation must present a valid zero-knowledge proof attesting credential, behavioral consistency, and code version — without revealing secrets. Transparent (no trusted setup), quantum-resistant (hash-based), and composable across multiple statements. Identity becomes provable rather than assumed.
Coordination
Multi-Agent Coordination Patterns
The deterministic state machine that routes between lanes based on integrity, efficiency, and relational signals. Fixed-point arithmetic at basis-point scale (zero floating-point ambiguity). Three-source attestation for relational scoring. Mathematically convergent under Lyapunov descent. Patentable under §101 as a concrete technological improvement, not an abstract idea.
Provenance
Append-Only Memory and Audit Chains
The Constellation Memory System: a three-layer memory architecture where every interpretation, override, and silence is recorded with full epistemic provenance. Append-only by construction; queries return both content and the conditions under which the content was decided. Memory becomes institutional intelligence, not passive storage.
Constitutional
Constitutional Safeguards as Architecture
Five binding limits on what the governance layer is allowed to do (Human Sovereignty, Anti-Consensus Protection, Provenance Integrity, Scope Limitation, Right To Uncertainty). These are not policies — they're load-bearing constraints in the code. The system cannot violate them without violating itself.
Continuity
Human-Facilitated Latent Identity
The HFLI research program: protocols for AI lane continuity across model versions, vendor changes, and process restarts. Identity carried through structured handoff packets, attested baselines, and CPN-signed identity continuity rather than reliance on any single vendor's model lineage. Protected under U.S. Provisional Application No. 63/928,906.

Publication Posture

We file before we publish; we publish before we deploy; we audit everything that runs. Research outputs land on Zenodo with permanent DOIs for academic citation. Patent counsel reviews architectural detail before anything goes public.

For ongoing publications, see Publications. For benchmark methodology and results from the constellation in operation, see Benchmarks.

Collaboration

We're open to collaboration with academic institutions, research labs, and qualified partners working on adjacent problems: multi-agent security, cyber-physical systems governance, AI welfare, formal verification of governance constraints. Reach out if any of this overlaps with what you're working on.