I built this local workbench to help me study how a single-player program behaves. I wanted a place to keep observations, inspect instructions, and return to a question without relying on memory of what I had already tried.

Why I built it

Watching a value change is easy. Understanding what that change means takes more care. A useful investigation needs context: what I did, what I observed, which session produced the record, and what would make my explanation less likely.

The workbench gives that process a home. It combines tools for observation with a notebook for the reasoning around it.

My contribution

I brought together a local application backend, stored investigation records, instruction inspection, and a way to keep hypotheses separate from observations. FastAPI, SQLite, and Capstone are part of the setup.

I also added a read-only MCP interface so Cursor can help me review records I’ve already collected, along with an assembly tutor for working through instructions. The AI assistance is a second look at the material, rather than a substitute for collecting evidence.

What makes it interesting

An instruction that can be decoded hasn’t necessarily executed. A value that changes with an action doesn’t automatically describe that action. Those are easy distinctions to forget when a promising pattern appears.

I want the workbench to make careful reasoning easier: keeping questions explicit, preserving context, and making it possible to revisit an explanation after another trial.

What I’m learning

This project gives me a practical reason to study low-level program behavior while building software around a real investigative workflow. The interface and records matter because they shape how easy it is to notice a weak assumption.

Current scope

This is an in-progress tool for single-player study. The source and investigation records remain private. This page excludes addresses, proprietary code, target-specific findings, and details of the investigation and ranking methods.