Agents ingesting contracts, invoices and filings need nested JSON matching a caller-supplied schema, but 2026 extraction benchmarks report frontier models achieving five-percent field-level pass rates on complex nested documents, because repeated arrays must be row-aligned and an absent field is indistinguishable from an invented one. A naive single prompt with schema-constrained decoding guarantees syntactically valid JSON, not correct JSON: it silently fabricates plausible values for missing fields and emits no per-field trust signal, so callers ingest hallucinations. Evaluate a vision-language extractor combining page-level retrieval, per-field self-consistency across repeated re-reads and layout crops, logprob-and-agreement confidence fusion, mandatory source-span grounding with page and bounding-box coordinates, an explicit not_present sentinel, and arithmetic cross-checks on totals. Measure field-level precision, recall and F1, hallucination rate on deliberately absent fields, span-grounding accuracy and risk-coverage AUROC of the confidence, against a hand-labeled gold set of nested documents with injected omissions, beating a naive base-model prompt, constrained-decoding-only extraction, and a commercial OCR pipeline. Deliver as the prototype a minimal runnable Python MCP server (stdio) exposing the priced AI tool extract_schema_with_abstention that invokes a language or ML model to produce its output, with a typed input/output schema, an x402-style pay-per-call metering stub that records a per-call price in USDT and emits a settlement receipt, and one smoke test that exercises the tool end to end.

Direct user-specified research topic: Agents ingesting contracts, invoices and filings need nested JSON matching a caller-supplied schema, but 2026 extraction benchmarks report frontier models achieving five-percent field-level pass rates on complex nested documents, because repeated arrays must be row-aligned and an absent field is indistinguishable from an invented one. A naive single prompt with schema-constrained decoding guarantees syntactically valid JSON, not correct JSON: it silently fabricates plausible values for missing fields and emits no per-field trust signal, so callers ingest hallucinations. Evaluate a vision-language extractor combining page-level retrieval, per-field self-consistency across repeated re-reads and layout crops, logprob-and-agreement confidence fusion, mandatory source-span grounding with page and bounding-box coordinates, an explicit not_present sentinel, and arithmetic cross-checks on totals. Measure field-level precision, recall and F1, hallucination rate on deliberately absent fields, span-grounding accuracy and risk-coverage AUROC of the confidence, against a hand-labeled gold set of nested documents with injected omissions, beating a naive base-model prompt, constrained-decoding-only extraction, and a commercial OCR pipeline. Deliver as the prototype a minimal runnable Python MCP server (stdio) exposing the priced AI tool extract_schema_with_abstention that invokes a language or ML model to produce its output, with a typed input/output schema, an x402-style pay-per-call metering stub that records a per-call price in USDT and emits a settlement receipt, and one smoke test that exercises the tool end to end.. Investigate this topic end-to-end: survey the state of the art, identify a concrete tractable research question within it, design and run an experiment, and report results.