Skip to content

Benchmarks and Evidence

CodeCortex treats benchmark results as evidence with a scope, revision, methodology, and limitations. Numbers are not generalized beyond what was actually measured.

Current committed evidence

The Evidence Fusion hardening report records:

Measurement Recorded result
Hardening test suite 711 passed, 28 skipped, 0 failed
Repository coverage 91.74%
Cold real-index import Python 0.93 ms / TypeScript 0.99 ms median
Warm exact definition lookup 0.19–0.23 ms median
Warm exact reference lookup 0.29–0.33 ms median
Unicode caret navigation 0.18–0.32 ms median
Freshness scan, 600 documents 4.25 ms median

These values are hardening-specific measurements on committed fixtures and indexes. They are not promises for arbitrary repositories or machines.

Evidence-quality benchmark

The same report records three bounded fixture cases:

Case Baseline strategy Evidence strategy
Duplicate symbols graph heuristic precision 0.50 precision index 1.00
Dependency version source-only 0.00 dependency intelligence 1.00
Mechanical migration lexical scan 0.50 structural search 1.00

The purpose is to demonstrate which evidence layer resolves each controlled ambiguity. It is not a claim that every real-world task improves by the same amount.

Reproduce local benchmark runs

cortex benchmark --path . --cases benchmarks/cases.json --output benchmarks/results.json

Production benchmark tooling uses revision-pinned public repositories:

python scripts/run_production_benchmark.py --provision

See benchmarks/README.md and benchmarks/production/README.md in the repository for the exact harness and output fields.

Real-agent measurements

For an instrumented coding-agent command:

python scripts/run_agent_matrix.py --command "./my-instrumented-agent"

Missing provider token counts, cost, files-read telemetry, or other unobserved metrics remain missing. CodeCortex does not infer or invent them.

Known unavailable evidence

The live dependency-documentation provider remains credential-gated. The hardening report explicitly records its live smoke path as skipped when no credential is configured. Do not interpret that skip as a successful live validation.

Source of record

For current evidence and caveats, read: