Each check ties a figure to the rows it came from, lists every difference with its size and its most likely cause, and refuses to bridge anything. They augment the person who owns the numbers rather than replace them: the check does the tying-out the same way every month, and the call stays with you. Built as Claude Code skills, tested on a synthetic pack with planted differences before they ship, and free under the MIT licence. Run one on your own month-end pack or client report in about ten minutes.
For the pack that goes to a CFO, a board or a lender. Every figure on the page tied to the ledger, the model, or the roster it came from.
Every figure on a report page tied to its source rows. Each gap tested for timing, a duplicated id, rounding scale, and page arithmetic. Run it in your browser with nothing uploaded, or install the Claude Code skill from GitHub.
Run it in your browser ›Budget against actual in, first-draft commentary out, every number traced to a row. Next.
A revenue or P&L bridge from one period to the next, each step tied to its rows. Next.
For the report an agency sends a brand, where the platform's revenue and the store's orders disagree and someone closes the gap by hand.
An ad platform's revenue by campaign tied to the store's orders carrying its UTMs. Each gap tested for attribution-window timing, a double-exported order, and platform rounding. The same page, with the campaign report loaded; the Claude Code skill is on GitHub.
Run it in your browser ›An A/B or holdout result read against the rule written down before the test ran. Planned.
The weekly report rebuilt from source with its counts printed and compared to last week's. Planned.
Drop the report and the source files into the page and read the one-page note. Or open the printed checks and the rules file, change the tolerance, the period, or the mapping, and run it again. Both readers are tested on the same planted differences before a check ships. The browser version uploads nothing; the Claude Code skills at github.com/bhoff105/kinetric-checks use only the Python standard library. Neither ever changes a number.
A check shows where two numbers disagree and how big the gap is. It does not know which date basis your board expects, which attribution window your brand should report, or whether a duplicated journal was meant to be reversed. Those are rules someone set once, usually by habit. Finding each one, writing it down, and testing it against your own history is the work Kinetric does for a fee, and sometimes the honest answer is that the rule you have is fine.
We'll tell you what your own history can say about it, and which offering fits. If none does, we'll say that too.
Start a conversation