Your agent is wrong about your own numbers one time in five.

The answer was in the files the whole time. We gave Claude Sonnet 5 and Claude Opus 5 everything we publish, with grep and cat to search it, and asked what our own figures say. Across 56 questions and four runs each, one answer in five came back wrong. Then we changed two things, and neither was the model or the prompt: the same files became a searchable store, and each model got a single tool that returns the figure we had already settled on. The scores went from 79% and 80% to 93% and 95%. Every run with both beat every run without them, on both models, and the work finished around three times faster while reading a quarter as much.

That store is a VS Code extension. It indexes the folders you already have open, keeps them current as you save, and gives whichever model you are already using five tools: search the store, read a whole file into ordered notes, look something up with citations, keep a decision, add another folder. It also puts your own servers in the editor's model picker, so an open-weight model drives the same chat and the same agent mode. The index and the text stay on your machine, nothing is routed through us, and there is no key to give anyone.

Sunstone North is the engineering and product studio behind the SRT (Semiotic-Reflexive Transformer) research program. We build small, auditable instruments that read meaning directly out of frozen language models, and everything you can run on your own hardware is free to run.

The VS Code extension, Sunstone, is on the Visual Studio Marketplace, and its source is at github.com/space-bacon/sunstone-vscode. For anything else, write to burton@sunstonenorth.com.

What is measured

• Localisation. On all 500 instances of SWE-bench Verified the gold file is named first 229 times, where the same bug report pointed at another repository lands 5 and a length-normalised text search lands 58. Ceiling 477 of 500. django is 46.2% of the set and per-repository recall runs 0.265 to 0.688. Every ranked list is published.
• Context spent. With the register an agent reads 85k characters of tool result per run against 396k without it on Claude Sonnet 5, and 116k against 441k on Opus 5, four runs an arm. It answered more, not fewer, and finished 2.7x and 3.5x faster.
• Portability. 97% head-space agreement when inference moves from datacenter CUDA to Apple Silicon, against a 99.96% same-runtime ceiling. Moving a head across costs 3 points of image-to-text R@1, 0.640 against 0.670, weights unchanged.
• Footprint. A 22 MB cross-modal retrieval head, nearly 3,000 times smaller than the frozen model it reads.

What you can run

Black Window puts a model, a search index and a reader in one browser tab. Point it at your folders and it indexes them where they already sit: 229 to 896 passages a second on WebGPU, rows kept at 196 bytes in int4, 2.8 million passages searched in 36 milliseconds. Nothing to install, no account, nothing uploaded, and it keeps working once you are offline.

Sunstone is the VS Code extension, macOS, on the Visual Studio Marketplace. Your own model servers appear in the editor's own model picker, and the folders you weave are indexed on your machine and searched by the model through a tool. A prompt past the model's window folds into the index instead of being refused: one fact 154,699 characters past the cut came back verbatim in 32.1 seconds, measured on GLM-4.7-Flash at a context of 131,072. Keys live in the editor's secret store, never in settings.

The Lab replays four guided chapters built from real precomputed artifacts: introspection, cross-modal retrieval, steering, and substrate invariance.

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. Tiers differ in latency and cost, never in capability.

The instruments

• Sunstone Head: 22 MB cross-modal retrieval from a frozen multimodal LLM. R@10 0.967 against a 1,000-image held-out pool at chance R@1 about 0.001, and R@1 0.416 on the 5,000-image Karpathy split, 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 prototypes were enough to separate one corpus under one loss with no labels given, which is a property of that run rather than a count of attractors in the model. New maps train on your own corpus at adapter cost.

Every claim ships with its controls

A number without a baseline is marketing. Every metric is normalised against an unrelated-text floor, bounded by a paraphrase or oracle ceiling, carries a shuffled-pairs control that must collapse to chance, and is centred before any cosine is reported.

Fifty-nine corrections filed, eleven of them against ourselves. On 21 September 2026 we re-ran the text-search baseline our own localisation result is measured against. Our version had no inverse document frequency, no length normalisation and none of the per-instance exclusions the retriever gets, and it answered 99 distinct paths where the benchmark has 373 distinct gold files. Repaired and re-run over the same 500 instances it scores 58 rather than 35, so the margin we had been quoting as 6.54 times is 3.95 times, and we state it as +0.342 absolute instead. The retriever's own figures never touched that code and did not move.

Research

• Finding the file: localisation on SWE-bench Verified: all 500 instances, every ranked list, the permuted-query floor, and the repaired baseline.
• The SRT program paper: consolidated findings, prepared for arXiv.
• The Treachery of Signs (SSRN 5987495).
• The SRT architecture (SSRN 6349978).
• SRT-NLA: activation verbalization (Stage 4).
• The SRT repository: code, checkpoints, and controls.

Where this lands

• 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 across a fleet, 3B to 235B, bf16 to 4-bit.
• Security and red-teaming: assess what the model reveals about its own internals.

The instruments are free. The fitting is not.

Free: Black Window; the Sunstone extension, licensed for unlimited internal use including commercial development on your own or your employer's code through September 2030; heads, adapters and probes as open weights; every candidate list behind every retrieval number; and the correction record.

Paid: fitted heads, ready now, where you embed locally and send 384-dimensional vectors and about 2.1 MB comes back; a fixed-scope instrumentation audit on one production model; and custom instruments and fleet work built on your corpus. We quote after a pilot, from the pilot's own numbers, because seat density and per-call cost are things we have not measured on your workload yet.

Contact

burton@sunstonenorth.com · GitHub · Hugging Face

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