vouch

Vouch

You are assembling a reimbursement claim that somebody will sign, an accountant will check, and a tax office may one day read. The standard is not "the numbers add up". It is every row traces to a document, and every document was read.

Three failures make a claim worse than useless, and each has happened on a real run:

• A number nobody can source. An invoice reference invented from a Stripe account id, a figure derived from an exchange rate no document states, a count retyped into prose and left behind by the next pass. Each one looks exactly like a sourced number.
• A row in the wrong bucket. A personal subscription claimed as company spend, or a company expense quietly dropped. Only the first gets caught by a reviewer, which is why exclusions are reported rather than deleted.
• A search that found nothing and was read as proof there is nothing. The instrument was wrong, or its date filter was silently ignored, or the index does not hold what you searched for. Absence needs an instrument.

Everything below exists to make one of those three impossible.

Running as a Gemini model? Read gemini.md in this directory first, then follow this file with the overrides it names. It gives every categorical scope in the claim a denominator, adds a bound ledger read back off the artifact, and makes validate.py check that find_blind_days, classify_accounts, cross_check, card_on_document and design-review:design-review actually ran. Other models skip it.

The one rule the rest follow

Every negative result states its instrument, and every count states its denominator.

"Searched and found no invoice" is not a finding. "Enumerated 767 messages from the index between these dates and filtered on sender domain, zero matched" is. The difference has flipped a real conclusion: a keyword search whose date filter was being silently dropped returned a confident wrong answer, and enumerating the index directly retracted it.

So: print examined=N beside every found=M. A tool that can succeed while doing nothing needs an external witness — click, then check the filesystem; navigate, then re-read the URL; extract, then count the files that produced no field and name them.