I measured his app with his own code. He measured my claim with his own corpus.

This fifth installment in a series reveals how an audit log exposed accounting errors within two applications. This installment focuses on the aftermath of the report, as the maintainer promptly fixed the issue in three days and acknowledged errors in the original analysis. The app, Clawdmeter, displays live Claude Code usage with a pixel mascot.

Upon filing an issue, the maintainer verified the claim using his own code, reporting that the number of token events was 2.34 times higher than expected. The issue stemmed from the app summing usage objects in three places, leading to inflated numbers.

The maintainer's reimplementation of the problematic code, utilizing his own functions, allowed for an unbiased assessment of the issue. This approach removed the need for lengthy arguments about the correctness of the original code. The report's findings included two key points. Firstly, the repeated usage objects were not always identical, with a discrepancy of 3,799 groups carrying different output_tokens values.

The second finding highlighted that collapsing usage per file alone was insufficient, as some message IDs appeared in multiple transcript files. To resolve this, the fix had to be implemented at the account-wide aggregator level, ensuring all records were folded into their respective message IDs under a per-bucket maximum. The maintainer not only confirmed the bug but also ran nine test cases against the fix, with all nine cases passing, providing additional validation of the resolution.