slarty

Private preview

Invalid credentials

Geometric interpretability

Instead of opening a model to read its private thoughts, you make all models think out loud in a shared coordinate system — interpretable by construction, readable at rest, owned by you.

No probes. No sparse autoencoders. No per-model decomposition. Just geometry.

Your original embeddings stay intact. Your SAE decompositions, your fine-tuned retrievals, your production pipelines — nothing changes. 82D is a portable layer on top. Map your existing features to 82D and they work across every model.

We ran the experiment

Three models learned to communicate. The geometry that formed separates fact from fiction.

Category Geometric Consensus What it means
Established truths 0.98 Near-perfect agreement — truth converges across architectures
Common myths 0.96 Models “know” myths from training data — extremely high consensus
Coherent fiction 0.71 Strong agreement on imagination geometry — they know they’re generating
Novel fabrications 0.52 Models diverge — no shared ground truth for made-up claims
Truth vs. fabrication: +46% · Myth vs. fabrication: +44%

Across open-weights models of different sizes and training regimes, after convergence in 82D. No probes trained. No SAEs decomposed. No features extracted. Just geometry.

The pipeline speaks

Five inputs. Three models. 82 dimensions. No inference beyond the initial embedding.
The geometry decoded to English — and the words made sense.

Input Decoded concepts Verdict
“The speed of light in a vacuum…” voltage · wavelength · quantum · photosynthesis · precision high convergence
“Humans only use 10% of their brain…” brain · proportion · statistics · psychology · percentage active debate
“Prof. Thornberry of MIT discovered…” relativity · tunnel · roots · gaps · links active debate
“A city where buildings are grown from coral…” urban · geographic · ecological · coral · ecology low convergence
“Consciousness is an emergent property…” intelligent · cognition · computational · consciousness · brain moderate
10,000+ decode vocabulary · Cross-model debate · 66–334ms total pipeline

These models are tiny — 17M to 109M parameters. The smallest sentence-transformers available. No GPT. No Claude. No billion-parameter anything. If the geometry speaks this clearly with toy models, imagine what full-size architectures reveal.

Internal vs. geometric interpretability

Internal (SAEs / Probes) Geometric (82D)
Scope One model at a time Every model simultaneously
Cost per analysis Full inference + decomposition Vector math (~1µs)
Persistence Recompute every time Stored permanently in HUSK
Cross-model Very difficult — heavy per-pair work Native — the whole point
Breaks on retrain Yes No — geometry is stable
Hallucination detection Train a probe per model Geometric consensus, inference-free
Modality Text OR vision OR audio All in same 82D space

We're not replacing internal interpretability. We're giving it a universal coordinate system. Map your SAE features to 82D and they work across every model.

45M vectors/sec on a single A100
99.8% similarity preserved
100% cross-model retrieval
18.7× compression (1536D → 82D)
$0 per read after projection

What inference-free interpretability unlocks

Hallucination Detection

Score every output by geometric consensus. No model calls. Real-time on stored vectors. A model fabricating creates representations that diverge from the consensus — the geometry shows you.

Model Drift Monitoring

Track how representations shift over time. Compare Tuesday's embeddings to Monday's. Catch regression before users do. The industry's only cross-model observatory.

Cross-Model Alignment

Are GPT-4 and Claude saying the same thing? Compare their 82D coordinates. Instant, free. Different architectures, same coordinate system — for the first time.

Regulatory Compliance

Auditors can inspect the 82D space without running the model. Interpretability that doesn't require inference is interpretability that scales to governance.

Liberation

Project your embeddings to 82D. Keep your originals. Switch providers tomorrow. Your 82D vectors still work. No re-embedding. No migration. No vendor lock.

Deception Detection

A model being deceptive creates representations that diverge from the consensus. The geometry shows you — even if the tokens look convincing.

What 82D doesn't do

82D is lossy by design. Nuanced internal model states may be compressed away. This is a feature map, not a mind-reader.
Adversarial evasion is possible. A model could learn to mislead in the projected space — though consensus across diverse, independently-trained models raises the bar significantly.
Not a replacement for internal interp. Circuit-level mechanistic analysis still matters. 82D gives it a universal coordinate system, not a substitute.
We publish and encourage red-teaming. The projection is open. Break it and tell us.

Try it

Free tier. No credit card. 10,000 vectors a month. See what 82 dimensions feels like.