// the reproducible leaderboard for cleaning up AI-generated code

Which cleanup tool actually removes dead code from AI-generated repos without breaking anything?

MuckLedger runs established, developer-trusted static-analysis and cleanup tools like ruff and knip against real open-source repositories and publishes only what can be proven: a passing test suite, zero deleted live call sites, and a run log you can read yourself.

view leaderboard →read the methodology

the promise

reproducible	Every run pins a tool version and a base commit SHA. Anyone can re-run it with the open-source harness.
tests pass	The target repo's own test suite is run after cleanup. A failing suite excludes the run outright.
no live callers deleted	Every symbol the tool removed is checked for remaining references. One broken call site excludes the run.
real metrics	LOC counts, file counts and test counts are measured by the harness, never typed in by hand.
approved by hand	Harness output lands unpublished. Nothing appears on the leaderboard until a human reviews the evidence.

what a run records

run
  tool          ruff @ 0.16.7
  repository    ticketctl @ 5301eb0
  tests_passed  true
  no_live_callers_deleted  true
metrics
  findings_total     14
  findings_resolved  14   (100%)
  loc_before       1138
  loc_removed         2   (0.18%)
  files_changed       5
evidence
  diff   evidence/<run>/cleanup.diff
  log    evidence/<run>/tool.log
  log    evidence/<run>/tests.log
  log    evidence/<run>/vulture.log   (detection only)

ranking rule

if not (tests_passed and no_live_callers_deleted):
    exclude          # not scored, not ranked

if findings_total == 0:
    clean_run        # counted separately, never averaged in as 0%
else:
    resolution_rate = findings_resolved / findings_total * 100

score = avg(resolution_rate) across every approved run for the tool
rank by score desc   # tie-break: avg loc_removed / loc_before

Have a tool you want benchmarked? Submit it.