Introducing Latent Geometry Lab & The New Research Hub
Measuring latent models (self-other-world) using LLMs as my experimental platform. Turning philosophical concepts into testable predictions using geometry.
Today I’ve released my new research hub that’s focused on bridging the gap between philosophical concepts and measurable properties: https://robman.fyi
This hub contains detailed information on all of my research so far, including the story behind why I take a geometric approach and why I’m using LLMs as my experimental platform.
The research program currently has three interconnected layers:
Curved Inference (Measurement)
Methodology for measuring residual stream geometry in LLMs: curvature (how sharply the system reorients), salience (how much it moves), semantic surface area (integrated work).
What we’ve measured:
Concern bends inference trajectories predictably (CI01)
Intent appears as structured surface area patterns - detectable before behavioural shifts (CI02)
Self-models appear to require a resistant non-zero curvature; models accept 3× perplexity increases rather than flatten completely (CI03)
Deictic competence emerges when self-other-world axes separate
Each measurement started as a FRESH prediction (see below), got operationalised into a metric and then tested empirically.
PRISM (Experiments)
PRISM (Persistent Recursive Introspective Self-Model) - a lightweight scaffold separating private deliberation from public output. Tests FRESH predictions (see below) using LLMs as experimental platforms.
Key results (1,271 trials across 3 models):
Hidden theatre (internal arbitration without surface conflict) in 36-53% of trials
Surface replies compress reasoning by ~70% before speaking
Significant style shifts between internal thought and external output
Pre-thought scenarios improve user alignment whilst reducing surface conflict
Model-specific fingerprints stable across topics
Operationalises phenomenological concepts from the Minimal Phenomenal Experience Project as measurable continua.
FRESH (Theory)
FRESH (Functionalist & Representationalist Emergent-Self Hypothesis) - the foundation for the whole program is a geometric framework treating consciousness as traversal through role-space under specific constraints.
Key claim: Subjective experience isn’t a mysterious extra ingredient. It’s what traversal through properly structured role-space looks like from inside. Identity isn’t substance but conserved shape of motion (GIP-S: Geodesic Identity Principle - Shape).
Unlike most consciousness theories, FRESH makes geometric predictions that can be measured if you have the right instruments.
Current Focus: Latent Deictic Models
Right now, the program centres on understanding how self-other-world models emerge and function geometrically.
These three models (self, other, world) don’t exist in isolation. They co-emerge because language demands stable deictic anchoring - orientation in person, time, place, discourse. How does this happen geometrically? When does it happen during training? What minimal architecture supports it?
Why it matters:
AI safety: Understanding computational self-models matters for alignment and deception detection
Consciousness science: Operationalises phenomenological concepts of perspectival structure
Interpretability: Provides tools beyond linear probes for measuring latent structure
This synthesises all three layers: FRESH provides the theoretical framework for deictic structure, Curved Inference measures when axes separate, PRISM tests predictions about register boundaries.
Read the full article: ”Parrot or Thinker: A Functional Account of ‘Thinking’ in LLMs”
What’s Available Now
On the hub - https://robman.fyi:
Complete theoretical framework (FRESH)
Published measurement methods (Curved Inference I-III)
Experimental results (PRISM)
Tools and replication guides
Papers, preprints, and articles
Audience-specific starting paths
On GitHub - https://github.com/robman/FRESH-model:
Full Curved Inference pipeline (capture → compute → analyse)
PRISM scenarios and metrics (code requires ethics agreement)
Example datasets and analysis notebooks
Everything designed to be falsified, not just demonstrated
Who This Is For
This research sits at the intersection of mechanistic interpretability, computational phenomenology, and AI safety - built for researchers who believe theory and tools should inform each other.
If you’re working on:
Mechanistic interpretability: Geometric methods detect hidden reasoning, self-models, deception signatures that linear probes miss
Consciousness science: FRESH makes phenomenological concepts falsifiable using LLMs as instruments
AI safety: Understanding where conflict gets resolved, detecting surface calm masking internal tension
Philosophy of mind: Making theories of self-models and agency empirically testable
Applied AI: Understanding what your model is actually doing beyond surface correlations
Then this program has tools, methods, or frameworks you can use.
The hub provides tailored entry paths based on your background: Start Here
What Happens Next
This newsletter documents an active research program. You’ll get updates when there’s something substantial:
New experiments and results
Tool releases and tutorials
Papers and preprints
Methodological breakthroughs
Open questions and collaboration opportunities
No filler. No hype. Just falsifiable predictions, working implementations, and geometric methods that turn philosophical concepts into testable claims.
Collaboration welcome on:
Applications to AI safety problems
Extensions to multimodal/embodied systems
Alternative operationalisations of FRESH predictions
Philosophical implications and critiques
Explore the research hub: https://robman.fyi
See the complete program: Research Program Overview
Access the tools: GitHub
Thanks for being here...
Rob