Twelve answers. One answer. One answer.
The same pipeline, unchanged code, run against three different real receipts. The system that says "twelve" here is credible specifically because it also says "one" there — a tool that always finds a dispute isn't measuring disagreement, it's manufacturing it.
Stablecoin, weekend settlement
USDC received on a Sunday no bank published a rate for. The market never stopped trading. Twelve defensible figures, ₹47,868.76 apart — and the "No rule found" chip on the valuation-method question, because none of the text this record was given prescribes how to pick one.
Open the record →INR to INR, no conversion at all
No currency question, no crypto, nothing to dispute. The same pipeline, the same schema, the same page template — and it correctly renders exactly one figure with zero spread, because that is what this receipt actually has.
Open the record →An ordinary cross-border wire, no crypto
A normal weekday bank wire, an SBI rate that was actually published that day. No crypto anywhere in this case. Same pipeline, no code change, no prompt change — proof this isn't a crypto-only tool wearing a tax-law costume.
Open the record →This is the whole thesis on one screen. A system
that reports a dispute on every input isn't trustworthy when it reports
one on the input that actually has one. These three records are real,
generated from saved runs by the same deterministic composer
(node7_disclosure.py) — none of the three pages was
hand-edited after being generated.