ADR 0011: Analytical Engines Never Fetch¶

Status¶

Accepted

Date¶

2025-02-03

Context¶

The old hotspot modules mixed mathematical evaluation with database and provider access:

analytical logic was difficult to test without runtime wiring;
deterministic computation was hidden behind repository calls;
engines could not be reused safely because their correctness depended on live IO.

The service-layer standard separates orchestration from computation. That separation fails if analytical engines are allowed to fetch their own inputs.

Decision¶

Analytical engines may not perform IO and may not fetch their own data.

The practical rule is:

services, repositories and integrations gather inputs;
engines receive typed inputs and return typed outputs;
engines stay pure with no database, HTTP, queue, cache or settings fetches;
if a domain needs cross-domain data, the service layer introduces an explicit adapter before calling the engine.

Consequences¶

This preserves deterministic analytical behavior:

engine tests run without DB/runtime setup;
orchestration and retry behavior stay outside the math layer;
cross-domain boundaries remain explicit instead of leaking into engine code;
“just fetch what the engine needs” is no longer an acceptable shortcut.