Code excerpts

KRONOS's source is private. These three files are reproduced here verbatim, unedited, from the private repository so that this case study isn't entirely claims about code you can't read.

They were chosen because each is self-contained, carries no product-specific logic, and shows a different facet of how the codebase is written.

File	From	What it shows
fusion.py	backend/aop/rag/fusion.py	Weighted Reciprocal Rank Fusion — the algorithm that merges dense (Qdrant) and sparse (BM25) retrieval into one ranking. 57 lines, pure, no I/O.
graders.py	backend/aop/evals/graders.py	The deterministic eval graders. No LLM judging, so an agent's score is exactly reproducible for a given output.
llm_adapter_base.py	backend/aop/llm/base.py	The abstract LLMAdapter every provider subclasses — the seam that makes swapping Ollama for Anthropic a config change (ADR-0002).

Every file here is mypy --strict clean and ruff-formatted as part of the full backend check — see docs/evidence for the run output.

What to look for, if you're skimming:

• Typed all the way down. No bare Any in a public signature; TYPE_CHECKING imports to keep runtime imports light; zip(..., strict=True) so a length mismatch is a crash, not a silent truncation.
• Docstrings that state contracts, not restate names. health() documents that it never raises for an unreachable endpoint — it reports DOWN with a fix path. That's the kind of thing a caller actually needs to know.
• Errors are typed and honest. ProviderHTTPError carries the provider, status, and body. An unconfigured adapter reports itself unconfigured rather than falling back to a mock — the same "no mocks" rule that governs the tests.