DIVERGENCE · three real records, side by side

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.

D1 — the hard case

Stablecoin, weekend settlement

12 methods

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 →
C1 — plain domestic invoice

INR to INR, no conversion at all

1 method

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 →
C2 — the scalability proof

An ordinary cross-border wire, no crypto

1 method

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.