Compass is the independent trading research platform I’m building. The work spans data capture, research engines, evaluation, an operator workspace called Helmsman, and the infrastructure that keeps those pieces running.
One repository is also named compass. That repository contains the engine. The wider product story is about how the engine works with the rest of the platform.
A useful research question might be whether a price that starts at $100 tends to reach $101 under a particular set of conditions. Writing down that idea is easy. Establishing what the system knew, how it interpreted the evidence, and what happened afterward takes considerably more engineering.
Three different responsibilities
The platform grew around three distinct jobs.
The engine reads evidence. It combines captured inputs, independent model assessments, and deterministic calculations into a recorded result.
Helmsman organises the work. It gives research ideas, monitors, decisions, and audits a place to live. It connects the human workflow to the underlying systems.
Structure Lab makes definitions testable. It lets us turn a visual idea about price behaviour into explicit configuration that can be inspected and evaluated.
Those responsibilities share data, but they do not need identical state machines. An observation can exist without a human committing to an idea. A research definition can exist before it deserves an active monitor.
The decision that changed the shape
In July, the design moved from tuning one engine toward supporting named engine variants over a common data layer.
That mattered because disagreement became something the architecture could represent. Different interpretations could have separate identities and versions while reusing captured evidence. Comparing them no longer needed to mean copying the whole system.
This also introduced a more useful question than “which configuration looks best?” We could ask which parts differed: the input, the model’s interpretation, or the deterministic rules applied afterward.
The infrastructure is part of the story
A research platform needs repeatable runs, durable records, understandable failures, and controlled releases. A successful run should leave enough evidence for someone to inspect it later. A failed source should remain visible as a failed source.
That is why the build includes orchestration, revision history, grading, and operator-facing explanations. Each helps turn a collection of scripts into a system we can reason about.
This series follows those decisions, including experiments that did not justify further investment. Compass remains privately operated; request a walkthrough from the project page if you want to see the broader system.