How recall works

Each step is a pure function with its own tests.

1. Tokenise. Lowercase, strip punctuation, drop stop words, matched on word boundaries, so everyone and milestone survive and removing one never corrupts Done.
2. Score. Sum weights across an inverted index: topic > router description > file name > content. No file is opened yet, and the cost grows with the question, not the corpus.
3. Select. Keep the top five with a confidence margin. When two sources are close, both come back, marked as ambiguous. When nothing matches, it says no match.
4. Read. Open only the winner.
5. Slice. Keep the ## section whose heading matches, or a window around the densest lines.
6. Hop. Follow at most one pointer out of that slice. The cap of one is enforced by a test.
7. Pack. Question, evidence and one instruction, capped at 9KB.

Why no model

Scoring takes well under a millisecond, so there’s no reason to spend tokens deciding where to look, and the result is the same every time for the same files. A model is only needed for what comes after: reading the evidence and answering.

The benchmarks, including the methodology mistakes found while producing them, are in bench/RESULTS.md.