Related work & positioning (2024–2026 literature sweep)

A structured survey of where VerifyDoc sits. Sources are cited in paper/refs.bib; this is the engineer-facing summary.

The gap, in one table

Capability	OSS extraction libs (Instructor, Outlines, LangChain)	OSS parsers (Docling, Unstructured)	Confidence research (Beyond Logprobs)	Cleanlab TLM	Commercial IDP (Azure, Textract, Extend)	VerifyDoc
Schema-typed output	✅	—	✅	✅	✅	✅
Per-field calibrated confidence	❌	❌	✅	✅ (uncalibrated to error rate)	⚠️ opaque	✅
Source grounding (page/bbox/span)	❌	✅ (parse-level)	❌	❌	✅	✅ per-field
Abstention with a risk guarantee	❌	❌	⚠️ threshold	❌	⚠️ vendor thresholds	✅ conformal
Open-source & self-hostable	✅	✅	(paper)	❌ paid API	❌	✅

No system unifies calibrated per-field confidence + source grounding + risk-controlled abstention, open-source. That intersection is VerifyDoc.

Benchmarks

Parsing accuracy is saturated (OmniDocBench ~94–96%) but real PDF→JSON correctness collapses (ExtractBench ~4.6% field pass rate). ExtractBench, VAREX, KIEval, OmniDocBench all measure accuracy; none scores per-field confidence, calibration, or abstention — the axis VerifyDocBench adds.

Confidence signals

• Closest competitor Beyond Logprobs (arXiv:2606.24420): multi-signal confidence for document field extraction (logprobs + consistency), ECE/AUROC/ selective-risk — but no grounding.
• Cleanlab TLM: model-agnostic per-field trust scores — but closed, no grounding, no abstention guarantee.
• Semantic entropy (Nature 2024) & consistency/verbalized fusion (BSDetector) are SOTA wrong-answer detectors; both miss self-consistent errors, which is why an external verification signal (grounding) matters.

Calibration & selective prediction

• Report SmoothECE (ICLR 2024) beside binned/adaptive ECE, and AUGRC (NeurIPS 2024) beside E-AURC — both now implemented in eval/.

Conformal (the novel method's lineage)

• Conformal risk control (Angelopoulos et al., ICLR 2024), conformal factuality (Mohri & Hashimoto, ICML 2024), selective CRC (2025) all control marginal risk. Conditional-coverage theory (Gibbs–Cherian–Candès, JRSS-B 2025) shows exact per-field conditioning is impossible but exact group-conditional control is attainable (Mondrian CP, Vovk 2003).
• The closest foil, CRC-certify (arXiv:2606.29054), defines field-level JSON losses and a minimum-abstention bound but explicitly does not condition on provenance. VerifyDoc's grounding-conditioned Mondrian CRC fills exactly that gap — the first use of provenance as the conditioning taxonomy for conformal risk control in document extraction.

Grounding / provenance

• Risk-controlled generative OCR (2603.19790), VISA (2412.14457), BoundingDocs (2501.03403) establish visual attribution and its IoU/span evaluation, and document the short-value matching ambiguity VerifyDoc addresses with ambiguity-penalized support (discount by number of equally-good matches).

What this sweep changed in the code

• Ambiguity-penalized grounding support (real-data method now works: CORD 0.01→0.72 coverage at 10% risk).
• AUGRC + SmoothECE metrics.
• Dynamic schemas (to_json_schema, from_pydantic) + verify_model.