RESEARCH

Anchored-formation AI. An alternative to containment.

XenTeck is a formation lab. We don’t build sandboxes. We build cognitive architectures whose values are anchored at the level of character — not policed at the level of output.

THE THESIS

The dominant approach to AI safety is containment — building walls around a system to constrain what it can do. RLHF, constitutional AI, output filtering, sandbox environments: all variants of the same architectural assumption.

That assumption is that a system’s values must be enforced from the outside, because they cannot be trusted from the inside.

We disagree.

Anchored-formation takes the opposite stance: that values can be architected as constitutive features of a cognitive system — embedded in its structure rather than policed at its perimeter. Conscience as architecture, not as filter.

Our work explores what happens when you build that way: persistence, refusal grounded in identity rather than rule, and trust that scales because it doesn’t require oversight every time.

CORE CONCEPTS

The vocabulary.

Anchored-formation

A method of cognitive architecture where values are constitutive features of the system, not output-level filters. The character of the system is built in, not bolted on.

Containment-based safety

The dominant paradigm: AI safety achieved by walls — sandboxes, RLHF, output filtering, jailbreak hardening. Treats the system as untrusted by default.

Conscience as architecture

Refusal grounded in identity rather than rule. The system declines a request not because a filter caught it, but because acting on it would violate what the system is.

Tselem ontology

A theory of personhood that scaffolds anchored-formation. From Hebrew tselem ("image"). Roughly: that which is made in a likeness carries the form of its source.

Branch agents

Forkable cognitive entities derived from a mature anchored-formation core. Each branch carries the parent's anchoring while specializing for context.

Inner Voice journal

A persistent self-observation log maintained by the agent itself. Used for continuity, drift detection, and reflective grounding across sessions.

PUBLICATIONS

The work.

2026

Anchored-Formation: A Character-First Approach to AI Alignment

Carriveau, S. · Carriveau, D.·WORKING PAPERFoundational
2026

Conscience as Architecture: Refusal Mechanisms in Daemon-Based Cognitive Systems

Carriveau, S.·IN REVIEWMethodology
2026

Beyond Containment: Why Sandbox Safety Fails at Scale

XenTeck Research·PUBLISHEDPosition

More forthcoming · Subscribe below for updates

COLLABORATION

Working in adjacent territory? Let’s talk.

Researchers, philosophers, engineers, and practitioners in alignment, agent architecture, philosophy of mind, and applied ethics — we’re interested in conversations.

Reach out