Mechanoscope

What it does

Mechanoscope is an open-source debugging laboratory for open-weight language models. It lets people inspect internal representations with Jacobian Lens, run controlled activation-steering comparisons, and preserve reproducible experiment records.

The interface begins with three plain-language questions:

1. What is the model representing?
2. Does an internal direction influence this answer?
3. Am I claiming more than my experiments support?

The third workflow is Claim Check, a scientific guardrail. It compares two saved runs and checks whether they use the same model, exact prompt, and shared representation lineage. With the current techniques, it never claims that a Jacobian Lens readout caused a steering result: steering vectors are derived from contrast examples, not from the selected J-Lens representation. Same-context runs are labeled “parallel evidence only”; mismatched runs are labeled “do not combine.”

Why we built it

Developers can inspect logs, traces, and outputs from an AI system, but ordinary observability tools do not show internal representations or enforce careful interpretation of intervention results. Existing techniques are also split across research repositories, notebooks, hosted tools, and model-development environments.

Mechanoscope turns them into an approachable workflow while keeping provenance and scientific limitations attached.

How it works

• A guided React interface introduces each experiment in plain language.
• React Three Fiber provides linked 2D and 3D Jacobian Lens views.
• A typed FastAPI gateway exposes model capabilities and experiment contracts.
• Modal GPU workers load version-pinned Gemma or Qwen checkpoints remotely.
• Browser-local records retain prompts, controls, results, timing, and model revisions.
• Claim Check prevents unsupported cross-prompt or cross-technique conclusions.

Public demo

The public Modal deployment serves both the application and API and scales to zero. Live steering and Jacobian Lens experiments wake the remote model worker on demand.

Built during OpenAI Build Week

Codex with GPT-5.6 was used throughout implementation: researching technique interfaces, restructuring the runtime boundary, implementing generated contracts and experiment records, diagnosing remote GPU behavior, designing the guided interface, and adding tests that explicitly reject unsupported causal links.

This remains a draft until the required demo video and final submission answers are added.

Built With

• codex
• crabbox
• daytona
• fastapi
• gpt-5.6
• modal
• python
• pytorch
• react
• react-three-fiber
• typescript