04 / MEASURED EVIDENCE

Measured,
not claimed.

Fast, quiet, safe — before saying any of that, here are the conditions and the numbers. What AccordRail has measured is never mixed with what it hasn't.

REPRODUCIBLE
CONDITIONED
LOCAL
0FALSE POSITIVES220 commits of other people's code
88.1%SILENT COMMITSacross 47,799 commits
1.5KBASSIGNMENT CONTEXTvs 182KB of whole-repo reading

READ THE CONDITIONS

Numbers aren't
decoration.

Benchmarks vary with the target, its size, and cache state. The values on this page are not a universal speed guarantee — they are a record of the current implementation measured under stated conditions.

Zero false positives does not mean "catches every bug." It measures precision: AccordRail did not speak up about anything it could not be certain of.

What can't be measured
isn't promised yet.

LOCAL ENGINE / CURRENT BUILD

It scales,
and stays local.

5K FILES / CACHED70ms

Analysis of 5,000 files with a warm cache.

20K FILES / GRAPH182ms

Dependency graph construction at 20,000 files.

PEAK RESIDENT MEMORY24MB

Peak resident memory observed in the same run.

1,664 COMMITS / HISTORY137ms

History check across 1,664 commits.

Measured during 0.1.3 development. Varies with hardware, file layout, and disk state.

0BYTEEXTERNAL TRANSFER

PRIVACY BY ARCHITECTURE

Protected by structure,
not by promises.

Source, dependency graph, spec of record, history, results — all processed locally. There is no code path that sends your source to an external API for judgment.

  • 01No cloud analysis
  • 02No external model calls
  • 03No telemetry

QUESTIONS / STRAIGHT ANSWERS

Answered
up front.

Does it evaluate the AI itself?

No. It doesn't score a model's intelligence or output quality. It checks whether the code actually written matches the record a human approved.

Can I use it on an existing project?

Yes. The known shape at adoption time becomes the baseline, and only deviations added after that are reported.

Does it replace CI or a linter?

No. Compilers, tests, and linters handle their own kinds of correctness. AccordRail covers agreement with the design, the ownership, and the done-criteria.

Can it detect everything?

No. It limits itself to what a dependency graph and explicit rules can establish with certainty, and never stops work on weak inference.

05 / NEXT

Check the numbers in your own workspace.Get started ↗