Nutrient DWS grounds the facts. ProofDesk decides when the evidence is sufficient to act.
the problem: two documents can be individually extracted with high confidence and still describe a transaction that should never be approved. extraction accuracy is not execution authority.
what we built: nutient dws extracts grounded evidence with confidence, page, and bounding box. proofdesk cross-checks facts across documents, applies a calibrated authority gate, and routes uncertain cases to humans. every decision is audit-trailed with hash chains and merkle proofs. human feedback improves future calibration.
the scenario: a $42,500 vendor onboarding. four documents. the agent reads them all correctly. but together they describe a transaction that should not happen.
insurance claim. $67,500 property damage. contractor estimate matches claimed amount. but water damage exceeds the $50k sublimit and equipment breakdown is excluded.
human decisions become calibration data — future thresholds improve
the convergence loop
each time a human resolves an exception, proofdesk captures: what confidence level, did they accept or reject, what field. this becomes calibration data. the online calibrator updates. future decisions improve. the trajectory: human review falls over time while measured false-authorization risk stays bounded.
confidence algorithms
conformal-style risk calibration
angelopoulos et al., iclr 2024 — selective prediction research
13 PDF fixtures across six document types (procurement, kyc, invoice, trade, mortgage, medical). each field extracted with confidence and source grounding.
historical calibration research
heterogeneous classification datasets used for threshold tuning and calibration experiments. these are not Nutrient document benchmarks — they are tabular/text classification datasets used to develop the confidence algorithms.
coldhearted fraud
19,872 transaction records
contractner
3,241 contract text examples
fatura
1,400 invoice records
cuad
509 contract clauses
invoicebenchmark
200 invoice examples
key result: at 1% false-sign rate, calibrated thresholds achieve 59.8% auto-sign coverage — 2.7x over baseline logistic regression. this was measured on the classification datasets above, not on the live Nutrient document pipeline.
evidence vs authority
the core thesis: correct extraction does not equal sufficient authority. two fields can both be extracted with high confidence and still describe an unsafe transaction.
tamper detection
post-approval. any byte-level modification changes the SHA-256 hash. the gate verifies the artifact before any irreversible action.