Instruments

Research instruments

An instrument is not a decorative visualisation. It must let you inspect, manipulate, or test a real model, dataset, representation, hypothesis, or experimental design. Interactive explanations here distinguish toy demonstrations from model data, and adapted tools keep visible attribution to their original authors.

Live

Causal tracing · live model dataJuly 2026Interactive instrument
Circuit Traces: Attribution Graphs for Open Models
The attribution-graph method behind Anthropic's Biology of a Large Language Model, on Gemma 2 2B: curated case studies with reading guides (two-hop fact recall, arithmetic, one-hop recall, indirect-object identification), plus a validated loader for any graph on Neuronpedia and a workflow for tracing your own prompts. Method: Ameisen et al. / Lindsey et al. 2025; open-source circuit-tracer; graphs served live by Neuronpedia.
Feature inspection · live model dataJuly 2026Interactive instrument
Latent Observatory: SAE Feature Explorer
Search the learned features of GPT-2 small and Gemma 2 2B by concept, then inspect the evidence: auto-interp explanations, the real text that makes each feature fire, the tokens it promotes and suppresses, and its nearest neighbours in decoder space. Live data via the open Neuronpedia API; every explanation is labelled as a hypothesis, with the activation records to check it against.
Field map · living documentv1 · July 2026Interactive instrument
The Interpretability Map
The papers, instruments, and open problems of mechanistic interpretability as one navigable document. A force-directed dependency graph of the literature (edges are prerequisites, not citations), a chronology from 2020 to 2026, a seven-layer reading pathway, runnable TransformerLens and SAELens protocols, a model-access matrix, open problems written as experiments, guided tours, search, filters, and progress tracking. Every node cites its primary source; the whole graph is downloadable as JSON.
Model inspection · adaptationJuly 2026Interactive instrument
Transformer Explainer: GPT-2 Live in the Browser

In development

Planned instruments follow the papers that need them. None of the following exists yet, they are listed so collaborators can see where the tooling is heading, not to imply availability. If one of these is the tool you need, build it with us.

IN PREPARATION

The Biology of an Open Model

An interactive paper applying the attribution-graph method end to end to one open model: traced case studies, intervention checks, and honest failure cases, in the institute's paper system. The Circuit Traces instrument is its laboratory.

In development
PLANNED

Evaluation Awareness Testbed

Matched prompt pairs, per-layer probes, and steering controls for the E1–E3 experiments proposed in Model Entrenchment. Blocked on: running the experiments.

In development
PLANNED

Probe vs SAE Benchmark

Side-by-side comparison of what linear probes and sparse autoencoders each recover from the same activations, on the same tasks, with the same controls.

In development
PLANNED

Entrenchment Simulator

The Figure 1 coordinate system as a manipulable model: set integration depth, switching costs, and behavioural parameters; watch where a system lands and which interventions move it.

In development

Elsewhere · instruments we recommend

Not ours, and not rehosted: tools whose licenses or hosting keep them where they are, linked because they are excellent. LLM Visualization (Brendan Bycroft) walks a 3D nanoGPT through every matrix of a forward pass. Neuronpedia is the open interpretability platform whose API powers two of our instruments. Transformer Circuits is the publication standard this site's paper system learns from.