Configuration coverage, measured deterministically.
ConfigReach discovers runtime configuration inputs, values, branches and important combinations, then maps them to test evidence without an LLM, GPU, API key or hosted service.
50,000 controlled cases. Balanced ground truth. One production scanner.
The committed curated benchmark contains 25,000 positive and 25,000 negative scenarios across supported programming languages and configuration formats. Every case has a deterministic expected label, belongs to an explicit version-controlled scenario family, and is scored through configreach.engine.scan.
Precision
100.0000% — 25,000 TP and 0 FP on the committed controlled benchmark.
Recall
100.0000% — 25,000 TP and 0 FN on the committed controlled benchmark.
F1 / Accuracy
100.0000% / 100.0000% with TP 25,000, FP 0, TN 25,000, FN 0.
Strong regression evidence, clearly scoped.
This is a controlled curated benchmark with explicit, version-controlled scenario families and deterministic ground-truth labels. It is not described as independently human-labelled and is not a claim of universal real-world accuracy.
Positive cases
Real supported configuration accesses and declarations across languages and structured configuration formats.
Adversarial negatives
Comments, inert strings, custom look-alike APIs, dynamic names, quoting traps and structured value-only cases.
Reproducible
The dataset, manifest, aggregate result and row-level predictions are all committed and published.
Run the released package locally.
The v0.9.5 release line includes the curated 50K benchmarked implementation and enterprise security hardening.
python -m pip install --upgrade configreach configreach scan . configreach coverage .
Repository package metadata: v0.9.5. Supported Python versions: 3.10–3.14.