The models you already run know more than they say.

Sunstone North is the engineering and product studio behind the SRT (Semiotic-Reflexive Transformer) research program. We build small, auditable instruments that read semantics, discourse structure, and cross-modal understanding directly from frozen language models. No retraining. No new model. No capability tax.

Train once, read everywhere. No GPU required.

The structure inside frozen language models is linearly readable, and it is invariant across host scale, weight precision, and hardware. One small trained artifact reads the same structure anywhere the model runs, from Pi-class devices to server fleets. Reading takes a single forward pass, not generation, so no deployment tier requires a GPU. Key numbers: a 22 MB cross-modal retrieval head nearly 3,000 times smaller than the model it reads; 0.17% of backbone parameters trained; one read-out design spanning 3B to 235B backbones; 97% head-space agreement from datacenter CUDA to Apple Silicon, measured against a 99.96% same-runtime ceiling, with on-device image retrieval within 3 points of the datacenter reference.

What we do

• Frozen-model instrumentation: semantic embeddings, per-token divergence, reflexivity, and discourse structure on the models you already serve.
• Cross-modal retrieval engineering: image-to-text search as a free rider on your multimodal LLM, one linear head, one prefill pass.
• Edge-to-datacenter deployment: the same artifact runs unchanged from Pi-class hardware to datacenter fleets, quantization-robust.
• Interpretability and audit: hidden states read back as sentences, adaptive-density reasoning traces, artifacts small enough to inspect by hand.

Eight signals from one forward pass: semantic embeddings, per-token divergence, reflexivity and regime (ECE at or below 0.001 on ported backbones), discourse communities (961 discovered without labels), cross-modal retrieval (R@10 0.967 on the release protocol), activation verbalization, reasoning traces, and hidden-state diagnostics (TruthfulQA-MC2 AUC 0.87).

The instruments

• Sunstone Head: 22 MB cross-modal retrieval from a frozen multimodal LLM. R@10 0.967 release protocol; R@1 0.416 on Karpathy 5k, on par with fully trained dual encoders from roughly 3,000 times less pair data.
• SRT-Adapter v1.0: the stable release. A ~12M-parameter adapter exposes per-token semiotic signals plus a sentence embedding with zero cross-entropy degradation.
• SRT Introspect: generate with a trace; narration dense where the model's internal state moves fast.
• zooL4nD3r v0.1: discourse cartography, 961 communities discovered without labels; new maps train on your own corpus at adapter cost.

Use cases

• Media and consumer platforms: visual search without a search stack.
• Enterprise search and RAG: embeddings from the model you already serve.
• Trust, safety, and communications: see where meaning forks.
• Governance, risk, and compliance: audit artifacts for the AI you deploy.
• Sovereign and air-gapped environments: full capability offline.
• Model operations: observability for the fleet.
• Platform safety and routing: watch state, route accordingly.
• Security and red-teaming: assess what the model reveals.
• Global products and localization: one meaning space, many languages.

Live demos

The Lab: four guided chapters (introspection, cross-modal retrieval, steering, substrate invariance) built from real precomputed artifacts. Also live on Hugging Face Spaces: SRT-Sunstone, SRT Showcase, gpt-oss-20b trace.

Research

• The SRT program paper: consolidated findings, prepared for arXiv.
• The Treachery of Signs (SSRN 5987495).
• The SRT architecture (SSRN 6349978).
• The SRT-Adapter manuscript (Stage 3).
• SRT-NLA: activation verbalization (Stage 4).
• The SRT repository: code, checkpoints, and controls.

Engagement

Three ways to start: a fixed-scope instrumentation audit on one production model; custom instruments trained on your corpus at adapter cost; fleet deployment across every tier with calibration and observability wiring.

Contact

burton@sunstonenorth.com · GitHub · Hugging Face

This page has a full interactive version; enable JavaScript to view it.