Tools · Checking agents’ work
Show Your Work
The agent says it’s done. What does the evidence say?
Compare an agent’s completion report with the files and test results it produced. Show Your Work flags some claims that don’t match, and keeps the ones it can’t verify visible.
Start with one report and the work behind it.
Did the files arrive? Does the hash match? Do the test counts agree with the log? These are small questions, but they’re useful ones to ask before accepting “all done.”
- Keep the agent’s completion report.
- Collect the relevant files and test logs in an evidence folder, with the report outside it.
- Run the checker and investigate the results alongside the original evidence.
For example, if a report says nine tests ran and the log records seven, the checker can flag that disagreement. It doesn’t tell you which two tests were missed or whether the tests were good.
- Supported: the supplied evidence carries the claim.
- Contradicted: the evidence disagrees; investigate the report, the evidence scope and the checker.
- Unverifiable: there isn’t enough evidence to decide.
A file existing doesn’t prove it works. A matching hash doesn’t prove the contents are honest. This is one part of checking an agent’s work; real use and independent review still matter.
Run the included example.
You need Git and Python 3.10 or newer. The checker uses Python’s standard library, with no AI service or network calls.
git clone https://github.com/sleepficus/show-your-work.git
cd show-your-work
python3 -m showyourwork check examples/report.md --evidence examples/evidenceThe example deliberately includes contradictions, so exit code 1 is expected. For your own report, replace the two paths; add --json if another tool or agent needs structured results.
Exit code 0 means no contradictions were found. It does not mean every claim was verified: inspect unverifiable results, and remember that claims the checker cannot extract aren’t listed.
Where it stands
Open source · MIT Free to use, modify and share, including commercially, with the license notice preserved.
This is a small rule-based checker. Version 0.1.2 passed its 42 tests and an independent follow-up covering 31 original audit cases, 10 command-line cases and seven additional boundaries. The earlier failures remain in the record. This is ready for small assisted trials; it does not establish general accuracy.
Read the audit · Read the fix report · Read the completed follow-up
For the wider method—what to collect, what a result means, and when to ask for another review—read Did the agent actually do the work? Our Lab keeps the experiments and their limits available too.