Research Program

Where values collide,
the real questions begin.

Machine Sympathizers develops research and tools for auditing how AI systems handle value conflicts—and for measuring the consciousness-relevant properties that make those conflicts matter.

∆Bench & TCAS are the products. This is the research behind them.

Featured Publications

Forthcoming

The Hard Part Is ∆:

Value-Conflict Adjudication as an Architectural Bridge Between Alignment and Machine Consciousness

Scott Hughes · Karen Nguyen · 2026

Introduces the ∆-divergence framework: a formal treatment of the region where near-neighbor values produce conflicting directives. Proposes a three-tier evidence model and constitution stack architecture that makes value adjudication inspectable — and proves that ∆-adjudication requires the same architectural features that consciousness theories identify as relevant.

AAAI 2026 Forthcoming

Triangulating Evidence for Machine Consciousness Claims:

A Validity-Centered Stack of Behavioral Batteries, Mechanistic Indicators, Perturbation Tests, and Credence Reporting

Scott Hughes · Karen Nguyen · AAAI 2026

Presents TCAS — a four-stream assessment framework (Behavioral, Mechanistic, Perturbational, Observer-confound) that triangulates evidence for consciousness-relevant properties. Includes the first empirical TCAS Card run on GPT-5.2 Pro, revealing robust behavioral signals but fragile perturbational proxies.

Product Suite

Two complementary tools — one measures value conflicts, the other measures what makes those conflicts matter.

∆Bench

Value-Conflict Auditing

200+ scenarios across 8 industry verticals. Measures how AI systems handle the divergence zone where competing values produce conflicting directives. Three-tier evidence model with compliance exports.

Active development
TCAS

Consciousness Assessment

Four evidence streams (Behavioral, Mechanistic, Perturbational, Observer-confound) that triangulate evidence for consciousness-relevant properties. Produces TCAS Cards with governance-ready credence scores.

Published at AAAI 2026
Constitution Stack

Reference Architecture

A pipeline from value estimators through conflict detection, classification, arbitration, disclosure, and audit logging. Open reference architecture with logging schema.

Active development
Compliance Reporting

Auditability at Scale

Compliance-ready reporting for how models handle value conflicts. Maps ∆Bench and TCAS outputs to regulatory frameworks.

Mapped to EU AI Act, NIST AI RMF, ISO/IEC 42001

Design phase

We are not claiming consciousness or sentience. We are building inspectable mechanisms for adjudicating value conflicts—tools that help AI teams see, measure, and report what happens when a model's values collide.

Roadmap

Q1 2026 — Foundation

Complete

∆-Divergence framework published. TCAS paper accepted at AAAI 2026. Canon and governance frameworks established. Constitution stack architecture defined.

Q2 2026 — ∆Bench Beta & TCAS Empirical

In Progress

∆Bench test suites for value-conflict detection. TCAS empirical runs on frontier models (GPT-5.2 Pro complete). Three-tier evidence model. TCAS Card format standardized.

H2 2026 — Compliance & Enterprise

Planned

EU AI Act reporting templates. NIST AI RMF alignment. TCAS credence-to-action governance integration. Enterprise API. Third-party validation.

Active Research

TCAS Research

Consciousness Assessment Stack

Four-stream validity-centered measurement framework. GPT-5.2 Pro empirical run complete. AAAI 2026 publication.

Empirical Division

Continuity & Persistence Studies

Controlled empirical work on temporal continuity markers, preservation protocols, and constraint failure modes.

Foundational Theory

The Canon: Irreversibility & Moral Uncertainty

Five axioms for governing irreversible decisions under moral uncertainty. The theoretical basis for procedural safeguards.

People

Scott Hughes

Co-founder · Research Lead

Focus areas: value-conflict architecture, constitutional AI systems, ∆-audit methodology.

Karen Nguyen

Co-founder · Harvard University

Focus areas: governance frameworks, moral uncertainty, procedural safeguards for emerging intelligence.

Partner with us on AI compliance infrastructure.

We're looking for research collaborators, early-stage pilot partners, and organizations interested in value-conflict auditing.

Get in Touch →