Research preview Introducing
Hermeneusthe interpreter

A model that is itself an expert in mechanistic interpretability. It reads another model’s internals — and tells you, in plain words, what it learned and where it lives.

the meaning of
scroll
§ 01The readerLive · one tongue per thread

Ask it anything. In any tongue.

Hermeneus reads the model’s internals and answers in plain language — a full conversation in one language, switching only when a new thread begins. It always opens in English.

Hermeneus — sessionEnglishlive
Speaks 10 languages this sessionEN · ES · FR · DE · 日本語 · 中文 · العربية · РУ · ΕΛ · LA
§ 02How it worksReads internals → explains

Point it at a model. Three moves to an answer you can sign.

01 · read

Reads the internals

Pythia’s paired traces, decomposed into sparse, named features — layer by layer.

activations · graph-flow
02 · locate

Finds what fires

It locks onto the feature, the layer, the magnitude — learned, not memorized.

localize · attribute
03 · explain

Explains it back

Plain language — your language — with the receipt attached.

plain words · signed

Request early access. Invites roll weekly.

Hermeneus is in research preview. We’re onboarding teams shipping fine-tuned models who need to prove what changed — not guess.

StageResearch preview
AccessPrivate beta · rolling invites
Runs onPythia · the instrument
LabDelphi · New York
⟦ request access ⟧

On the list.

We’ll reach out from research@delphilab.ai with a preview invite. Check your inbox.

No spam · we read every request · γνῶθι σεαυτόν