01 Pick a document
02 The document
03 Two independent reads pre-computed
Model read A
Model read B
04 Consensus diff runs live
| Field | Read A | Read B | Status |
|---|
05 Seven deterministic gates runs live
06 Why this shape
A single model read fails silently: an ambiguous date, a hallucinated field, a faithfully-copied typo all come back looking equally confident. Running the same document through two independent reads and diffing the results turns silent failure into a visible disagreement — cheap to compute, impossible to ignore.
But agreement is not correctness (try the third document). So no model output ships on consensus alone: a layer of deterministic gates — checksums, format rules, date arithmetic, required-field checks — makes the final call in plain code. Models propose; gates dispose. This demo mirrors a pattern running in a document-extraction engine in production for a client, rebuilt here with fully synthetic data and pre-computed model outputs so it can run anywhere, for free, with no API keys.