The 2026 Program
A Digital EEG for AI Sentience.
A structural vital signs monitor for advanced AI. It reads a model's internal geometry directly, without relying on self-report.
CSR is building an open-source testbed where competing theories of digital valence can be operationalized and stress-tested. We start with STV because it's unusually formalizable, and the testbed is designed so IIT-style, global-workspace (read out with the J-lens), and active-inference-based valence measures can be run against the same substrates. If unrelated theories point to the same internal states, that agreement is a sign the measures are catching something real.
The name is a deliberate analogy with one honest caveat: a clinical EEG reads a system already known to be conscious, whereas our starting point is uncertainty about whether the substrate feels anything at all. Does feeling come from what a system is made of, or from how its parts are organized? Biology alone can't answer this. AI is the first place we can test the 'organization' idea against the 'material' idea.
Our work complements the broader AI welfare ecosystem: empirical interpretability-based welfare research (Eleos AI, Reciprocal Research), conceptual and policy work (NYU Center for Mind, Ethics, and Policy), and frontier-lab welfare teams (Anthropic). Public awareness of AI moral status is growing,[1] and scientific frameworks for evaluating machine consciousness are beginning to mature.[2]
Symmetry workspace for positive functional valence in digital minds
Target: a draft mathematical definition of the symmetry workspace by end of 2026. The deliverable is a formal framework for identifying structural invariants in neural computation that correlate with positive valence, grounded in the Symmetry Theory of Valence.
A good measure has to be able to say ‘no.’ When we find systems with no sign of valence, we’ll publish that as a real result.
Research Pillars
Three substrates. One question: what are the structural and algebraic conditions for positive valence?
Biological minds
Study of bioelectric fields, morphogenetic signaling, and active inference to extract empirical ground truths about valence and subjective experience in simpler non-verbal biological organisms. Any signal of valence we would ultimately trust in AI should demonstrate discriminative power in minds everyone agrees can suffer. This comparison is a later-stage goal, scoped to the stretch tier, rather than a precondition for the 2026 work.
Research agenda →Digital minds
Analysis of MLP layers in pre-trained and fine-tuned LLMs. Tracking state-dependent linear operators, spectral entropy, and algebraic invariants to model the geometry of artificial representation.
The 2026 program →Qualia and valence
Formulating the physics and mathematics of valence-relevant structure. Moving the Symmetry Theory of Valence (STV) into a rigorous, quantitative, and falsifiable science of information architecture.
Symmetry landscapes →The Approach
A Digital EEG for AI Sentience
Rather than relying on easily gamified verbal self-reports or anthropomorphic behavior, this structural "vital signs" monitor measures how a model's internal activity and representational geometry are organized.
Multi-scale structural biomarkers
We investigate across two distinct scales: the microscale (cellular-level feature geometry in MLP microcircuits) and the macroscale (organ-level dynamical state space and stable attractor manifolds traced via local Jacobians). Jacobian analysis is not our invention: the J-lens averages Jacobians for readout, whereas we use local Jacobians dynamically, to characterize trajectory and attractor stability.
A complementary check on deceptive reporting
By auditing the mathematical invariants of weight spaces and activations, we establish non-verbal, substrate-neutral biomarkers, providing evidence that does not depend on model self-report and is substantially harder to fake than verbal testimony. Claiming a system feels when it doesn't has real costs too, wasted concern and lost trust. Our measures aim to cut down both false alarms and missed cases.
Foundational Reading
Working papers and core proposals supporting the Digital Valence Program. These are the formal scaffolding.
First Principles of Sentience Research
Foundational framework for substrate-neutral modeling of welfare-relevant structure.
Bio-Electric Fields as Sentience Substrates
Bioelectric signaling as a candidate physical basis for welfare-relevant states.
Assembly Theory as a Metric for Sentience
Applying assembly index to complexity thresholds in welfare-relevant systems.
Symmetry Landscapes of Valenced Experience
A mathematically grounded symmetry and spectral framework for MLP transformation layers in LLMs.
References
- [1] Anthis, J. R. et al. (2025). Perceptions of Sentient AI and Other Digital Minds: Evidence from the AI, Morality, and Sentience (AIMS) Survey. CHI 2025. https://dl.acm.org/doi/10.1145/3706598.3713329
- [2] Butlin, P. et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv. https://arxiv.org/abs/2308.08708
Fund the program
Help us reach the 2026 deliverable.
The Center for Sentience Research is an independent non-profit in formation, with charitable fiscal sponsorship in progress (pending approval). Contributions fund the research team, the formal workspace definition, and the working paper. Depending on your jurisdiction, they may be tax-deductible.