Scoring IDP [R]

This is a "survey of current practices" type of research for a niche area, nothing earth-shattering. I’ve written a 2-part Medium series (free, no paywall) on scoring Intelligent Document Processing (IDP). It focuses on Key Information Extraction (KIE) from structured documents, especially forms. I lean heavy on metrics derived from the confusion matrix: different ways to map items such as blanks or misreads to the four possible values, how to count them, when to use each metric. Scoring IDP, Part 1: Confidence Is Not Correctness Scoring IDP, Part 2: Approaches to Evaluating Form Extraction Topics I include, but not limited to: recording raw outcomes before mapping them to the confusion matrix (correct, misread, missed, invented, correct blank); confusion matrix scoring for text fields and checkboxes; misread conventions and their effect on scores (FN-only, FP-only; FN & FP); how blanks, NULLs, unreported checkboxes, and unreadable values are counted; F1 vs. Recall vs. Accuracy, including how correctly identified blanks inflate Accuracy; Jaccard Index (IoU) as Accuracy without True Negatives, and its conversion to F1; Precision, Specificity, and False Positive Rate for LLM hallucination control; micro, macro, and weighted F1; choosing a headline score and stating the convention behind it; confidence vs. correctness and calibration (ECE, MCE); and Key Information Extraction (KIE) vs. OCR metrics (CER, WER). I believe these would fall under "machine learning scoring techniques" applied to Intelligent Document Processing (IDP)--how to transparently and fairly measure (score) rather than how to implement. It may also help folks understand how to read results coming from different benchmarks. I try not to recommend one opinionated way of doing things and instead cover the variations I’ve run into and the tradeoffs between them. I’m not endorsing any vendor; the series is about scoring itself.