Start with a bounded referent
Define the system, geography, time horizon, actors, processes, outcomes, and what is outside the boundary. Do not start from a keyword and quietly change the object during research.
Systems Atlas
Method
A lightweight research and representation contract. It is not a universal methodology.
Separate formal and actual systems
Ask what policy says should happen, what software records, what actually happens, what workaround keeps the system operating, and what consequence follows from the gap.
Build representation from evidence
Identify actors, states, transitions, authority, dependencies, evidence relationships, constraints, outcomes, and unknowns before drawing the elegant map. A map compresses evidence; it does not replace it.
Keep epistemic status separate from lifecycle
OBSERVED, SUPPORTED, INFERRED, HYPOTHESIZED, UNKNOWN, and CONTEXT-CONDITIONAL describe evidence. ACTIVE, NARROWED, REJECTED, and RETIRED describe the proposition's lifecycle. Research-program stage is a third, separate dimension.
Separate causal identification from binding-constraint validation
A randomized or quasi-experimental design may identify a causal arrow. Calling the same variable a binding workflow constraint requires a stronger chain: prospective prediction, targeted relaxation, a successful mechanism check, and the predicted downstream response.
Maintain rival mechanisms
Important observations should permit more than one live explanation when evidence warrants it. Ask what different observation each mechanism predicts and what intervention can discriminate among them. NONE / UNKNOWN must remain possible.
Make prediction prospective
Record the mechanism, unit, setting, time horizon, predicted observable, intervention, mechanism check, rival, and falsification condition before the relevant outcome is known. Do not repair failed predictions retrospectively.
Use a small map grammar
Entities, states/activities, boundaries, authority, evidence, unknowns, required flows, inferred relations, feedback, and explicit labels are enough for most public maps. New visual primitives must earn a recurring semantic distinction.
Preserve system history and simplify
A current map and a historical map are different objects. Preserve the path of revision, but remove or compress concepts that add little prospective prediction, causal discrimination, intervention value, or explanatory gain.
Publication is a separate decision
The research repository may contain raw, historical, private, or premature material. The public website uses an explicit allowlist and a pinned, human-reviewed snapshot. Research updates never publish automatically.
Stop rule
Stop meta-work when the next uncertainty is better answered by data, interview, observation, experiment, field case, or implementation.
Public snapshot allowlist: 47 source artifacts. Private-by-default field discovery paths are explicitly excluded.