Preliminary evidence that the 5D epistemic space of AletheionV2 converges toward toroidal topology as predicted by Directional Relational Manifolds theory — and how we measured it.
The Setup
We've been building AletheionV2 — a decoder-only LLM where every token generates not just a vocabulary logit, but a full epistemic tomography: aleatoric uncertainty (q1), epistemic uncertainty (q2), calibrated confidence, intentionality vector, and cognitive state. These 5 scalars live on a learned 5D Riemannian manifold.
The underlying theory — Directional Relational Manifolds (DRM) — predicts that stable DRMs naturally converge to toroidal topology. Specifically, the manifold should have the homological signature of a torus T²:
- H1 = Z² (two independent loops)
- H2 = Z (one cavity)
We decided to test this empirically. Here's what we found.
The Experiment
We applied Riemannian Voronoi tessellation to the 5D epistemic vectors generated by a 1M parameter AletheionV2 model, then ran persistent homology to check the topology.
The pipeline:
- Extract 5D epistemic vectors per token (~285K tokens from WikiText-103)
- Riemannian K-means tessellation (30 seeds, initialized from semantic anchor points)
- Local Tangent Space Analysis (LTSA) per Voronoi cell
- Persistent homology via
ripser(H0, H1, H2) - Compare under different metrics: Euclidean vs learned Riemannian G(x)
We ran this across three training phases:
| Phase | Metric | Description |
|---|---|---|
full_mahalanobis |
Euclidean | Constant metric baseline |
real_geodesic |
G(x) MetricNet | Position-dependent Riemannian metric |
gravitational_objective |
G(x) MetricNet | Extended training |
The Results
Topological Convergence
| Phase | H1 | H2 | ANOVA F (avg) |
|---|---|---|---|
| full_mahalanobis | 48 | 12 | ~260,000 |
| real_geodesic | 29 | 5 | ~900,000 |
| gravitational_objective | 33 | 7 | ~1,029,000 |
| T² target | 2 | 1 | — |
H1 dropped from 48 to 29 — a 40% reduction — when we activated the learned Riemannian metric G(x). The topology is simplifying in the direction the theory predicts.
What the ANOVA is telling us
We ran one-way ANOVA across Voronoi cells for each of the 5 epistemic dimensions:
q1 (aleatoric): F = 658,673 p ≈ 0
q2 (epistemic): F = 879,582 p ≈ 0
q3 (complexity): F = 1,144,089 p ≈ 0
q4 (familiarity): F = 1,040,744 p ≈ 0
q5 (confidence): F = 1,426,255 p ≈ 0
F values above 600K mean each Voronoi cell corresponds to a completely distinct epistemic region. The tessellation isn't arbitrary — every leaf has its own interpretable epistemic identity.
Dimensionality compression
The effective dimensionality of the real model vs null models:
| eff_dim mean | eff_dim median | |
|---|---|---|
| Real model | 3.5 | 3 |
| Null (shuffled) | 3.8 | 4 |
| Null (uniform) | 3.8 | 4 |
The manifold is operating in ~3.5 effective dimensions, compressed below the 5D ambient space. The null models don't show this compression. This is structure, not noise.
What This Means (and What It Doesn't)
What's confirmed:
- The epistemic manifold has real, non-trivial geometric structure
- Activating the learned Riemannian metric G(x) simplifies the topology — the metric encodes meaningful geometry
- Each Voronoi cell/leaf has a distinct, interpretable epistemic profile
- The topological simplification is monotonic and consistent with DRM's prediction
What's NOT confirmed:
- T² has not been reached. H1=29 vs H1=2 target — we're on the trajectory, not at the destination
- This is a 1M parameter model with minimal training. The DRM predicts convergence in stable DRMs — a 1M model with ~600 steps is not stable
The honest interpretation:
We're watching the topology simplify in the right direction. Whether it actually converges to T² requires the full-scale experiment.
Why Toroidal Topology Matters
The DRM paper proves that stable DRMs naturally converge to toroidal topology. If this holds empirically for a neural model, it means:
- The model didn't learn an arbitrary geometry — it learned a specific one that the theory predicts
- The 5D epistemic space has a natural closed structure — epistemic states wrap around rather than diverging
- This connects a mathematical theory of adaptive dimensionality to a concrete neural implementation
From the Geometry of Consciousness paper: a system with a 5D geometric substrate has a theoretical cognitive order ceiling of O_max = 5. If the manifold is toroidal and stable, it means the system is using all 5 dimensions in a structured, non-degenerate way.
The Falsifiable Hypothesis
When we run the full experiment on the 350M model with proper training (5x H200, ~7B tokens per phase):
If H1 ≥ 10 after full training chain (without RLHF): Scale and training alone are insufficient for toroidal convergence — RLHF may be a necessary condition, not just an accelerator.
If H1 < 10 after full training chain: Convergence is driven by scale and training, RLHF is an accelerator.
If H1 = 2 with long persistence bars: DRM empirically validated — direct connection between mathematical theory and neural implementation confirmed.
We'll know in about two months.
The Code
Everything is open source under AGPL 3.0:
Repository: gnai-creator/aletheion-llm-v2 — branch epistemic-foliation
Key scripts:
# Extract 5D epistemic vectors
python scripts/extract_epistemic_vectors.py \
--checkpoint checkpoints/your_checkpoint/final.pt \
--output-dir eval_results/foliation \
--label experiment \
--device cuda
# Run Voronoi tessellation + foliation detection
python scripts/voronoi_foliation.py \
--vectors eval_results/foliation/experiment_vectors.npy \
--checkpoint checkpoints/your_checkpoint/final.pt \
--output-dir eval_results/foliation \
--n-seeds 30 \
--use-metric-net \
--homology-points 1500 \
--device cuda
# Generate visualizations
python scripts/plot_foliation.py \
--results-dir eval_results/foliation
The foliation detection pipeline covers:
- Riemannian K-means with position-dependent metric G(x)
- LTSA (Local Tangent Space Analysis) per cell
- Tangent coherence testing
- Reeb graph via level sets with automatic logit pre-conditioning
- Persistent homology with T² validation criterion (H1=Z², H2=Z)
- Null model comparison (shuffled, uniform)
- Foliation score F ∈ [0,1]
Papers
- AletheionV2: A Decoder-Only LLM with Intrinsic Epistemic System on a 5D Riemannian Manifold
- DRM: Directional Relational Manifolds
What's Next
- Full training chain on 350M (5x H200, ~2 months)
- Backbone → full_mahalanobis → real_geodesic → gravitational_objective → foliation
- The falsifiable hypothesis above will be tested with proper scale and token budget
If you clone the repo and run the pipeline on your own models, I'd genuinely want to know what topology you find. The experiment is straightforward to replicate on any model that produces per-token uncertainty estimates.
Preliminary results on a 1M parameter model. Do not cite as definitive validation of the toroidal hypothesis. Full validation pending 350M training.
The DRM theory predicts this torus.
The neural network is learning it.
350M will tell us if it converges.
Is this the geometry of cognition itself?
Top comments (0)