Core idea: connect internal model change to behavioral cause.

Flashlight is a documentation-first causal interpretability project. Its purpose is to move beyond correlational explanations and test whether specific internal changes actually drive observable behavioral differences in a model.

The emphasis is intervention-based validation. It is not enough to find an internal signal that co-occurs with a behavior. The research question is whether changing that internal signal reliably changes the behavior in the predicted direction.

Interpretability work often stops at descriptive patterns:

• this neuron activates here
• this feature cluster appears there
• this representation correlates with a response style

Flashlight is interested in the next step: causal evidence. If an internal feature matters, intervention should reveal that relationship more directly.

• dummy brains: controlled model surrogates used to make analysis loops cheaper and more testable
• analyzer: the component that extracts candidate internal signals and hypotheses
• intervention loop: the validation process that perturbs internal state and measures behavioral effects

This repository documents the conceptual system, research direction, and validation philosophy. Detailed instrumentation, training procedures, and experimental implementation are intentionally kept abstract.

• STATUS.md: current project state
• docs/overview.md: plain-language overview
• docs/methodology.md: high-level methodology
• docs/roadmap.md: staged research plan