Retrieval is the quiet bottleneck in agentic coding. Most public SWE-bench submissions show strong single-file editing ability but fall apart when the fix requires touching files not explicitly mentioned in the issue, or when the repository exceeds the model's context window. The retrieval step — deciding which files to load into context before reasoning — is where a large share of end-to-end failures originate.

This axiom isolates that step. The solver delivers a retrieval function that takes (repository snapshot, natural-language issue) and returns at most 20 file paths believed to be relevant to the fix. The repository is guaranteed to be at least 500,000 lines of code. The retrieval must operate in a single pass — no agentic exploration, no test-time search, no iterative query refinement. The solver is permitted to use an open-weight model of their choosing, with the constraint that weights and retrieval code are fully published.

Verification is an F1 measurement against a labeled ground truth. The held-out evaluation set contains 2,000 issues across at least 50 repositories, each labeled with the true set of files that were modified in the historical fix. Precision and recall are computed per issue on the returned file set, then averaged. Axiom passes when mean recall ≥0.85 and mean precision ≥0.60.

The single-pass constraint is important. It forces a real retrieval solution rather than a disguised agent, which is what production systems actually need.