Would you consider a non-LLM-based solution that provides validation scores based on ground truth, such as having a business analyst supply invoices and then validating the pipeline end-to-end?

A robust non‑LLM validation framework for invoice processing can be built around a rule‑based engine combined with statistical checks and human‑in‑the‑loop verification. The architecture starts with a data ingestion layer that pulls raw invoices from the source system. A pre‑processing module normalizes formats, extracts key fields (invoice number, date, amount), and applies OCR if needed. Validation rules are encoded in a policy engine (e.g., Drools) and cover business logic such as duplicate detection, date ranges, and amount consistency. For each invoice, the engine produces a validation score based on rule matches and confidence levels. Ground truth labels are supplied by business analysts who review a sample set; these labels feed into a continuous monitoring dashboard that tracks precision, recall, and false‑positive rates. The pipeline is instrumented with metrics (latency, throughput) and integrated into CI/CD so that any regression in validation logic triggers alerts. Complexity is linear in the number of invoices (O(n)), with rule evaluation cost depending on rule density. This design ensures that even without a large language model, the system can provide reliable, auditable validation scores and support end‑to‑end pipeline verification.