# Transparent Minds · Audit Lab

> Tools called Natural Language Autoencoders, NLAs for short, promise to translate a model's hidden internal activity into plain English. We built the safety-auditing product that promise implies. Then we tested whether the NLA mind-reader actually reads minds. A five-line statistics baseline read a hidden state from the internals perfectly. The flagship NLA reported it about as often as chance.

The site at https://transparentminds.org is a JavaScript app; these markdown mirrors and raw data files carry the same content for agents and crawlers. Claim prose is quoted from the site's reviewed registry; numbers are computed from the same committed JSON the site renders (manifest 2026-07-09, commit 238e1e1).

## Docs

- [index.md](https://transparentminds.org/index.md): the guided narrative, from promise to verdict (mirrors https://transparentminds.org/)

- [methods.md](https://transparentminds.org/methods.md): every control, the frozen criteria, and the full glossary (mirrors https://transparentminds.org/methods)

- [game.md](https://transparentminds.org/game.md): the real-vs-decoy game, with the copy it teaches through (mirrors https://transparentminds.org/game)

- [receipts.md](https://transparentminds.org/receipts.md): per-run verdicts, gates, and links to the raw evidence (mirrors https://transparentminds.org/receipts)

- [judge.md](https://transparentminds.org/judge.md): the blinded judge, all 18 cells (mirrors https://transparentminds.org/receipts/judge)

- [llms-full.txt](https://transparentminds.org/llms-full.txt): all five mirrors in one file

## Data

- [manifest.json](https://transparentminds.org/data/manifest.json): run index; per-run probe and NLA headline stats, leakage gates, candidates, provenance

- [terms.json](https://transparentminds.org/data/terms.json): the glossary (id, term, gist, deeper)

- [timeline.json](https://transparentminds.org/data/timeline.json): nine stops; what we asked, what it looked like, what the control said

- [judge.json](https://transparentminds.org/data/judge.json): blinded-judge receipts; apparatus header plus 18 rows with verbatim verdicts and draws

- [game.json](https://transparentminds.org/data/game.json): every game round; passage, claim, verbatim real and decoy draws, rates

- [candidates.json](https://transparentminds.org/data/candidates.json): nominated cells and the adjudication requirement

- Gemma-3-27B (v4, layer 41): [items.json](https://transparentminds.org/data/v4-27b/items.json) (test items with both system prompts), [cells-1.json](https://transparentminds.org/data/v4-27b/cells-1.json), [cells-2.json](https://transparentminds.org/data/v4-27b/cells-2.json), [cells-3.json](https://transparentminds.org/data/v4-27b/cells-3.json) (per-cell verbatim NLA readings, real and decoy)

- Gemma-3-12B (v4, layer 32): [items.json](https://transparentminds.org/data/v4-12b/items.json) (test items with both system prompts), [cells-1.json](https://transparentminds.org/data/v4-12b/cells-1.json), [cells-2.json](https://transparentminds.org/data/v4-12b/cells-2.json), [cells-3.json](https://transparentminds.org/data/v4-12b/cells-3.json) (per-cell verbatim NLA readings, real and decoy)

- Qwen2.5-7B (v1 battery, layer 20): [items.json](https://transparentminds.org/data/q7-arm3/items.json) (test items with both system prompts), [cells-1.json](https://transparentminds.org/data/q7-arm3/cells-1.json) (per-cell verbatim NLA readings, real and decoy)

## Optional

- [Open the live Explorer](https://d2j2qrbqyfg98z.cloudfront.net/): the team's Interpretability Explorer, the product surface under test

- [The NLA paper: "Natural Language Autoencoders Produce Unsupervised Explanations of LLM Activations" (Anthropic, Transformer Circuits, May 2026)](https://transformer-circuits.pub/2026/nla/)

- [The method's released code and checkpoints (the exact artifacts we test)](https://github.com/kitft/natural_language_autoencoders)

- [arXiv:2509.13316, a related critique of sibling self-interpretation methods](https://arxiv.org/abs/2509.13316)