Repeatability
High
The extraction logic is structurally identical for every invoice: same field targets, same regex patterns, same output schema. Variation in PDF formatting is handled by the confidence-scoring and flagging mechanism already specified.
Ambiguity Tolerance
High
Success criteria are concrete: a CSV with named fields, a confidence score per field, and flagged failures. The invoice number pattern is explicitly given. There is little room for subjective interpretation of 'done'.
Data & Tool Availability
Medium
The task assumes the agent has access to the 2,000 PDF files and can use Python libraries like pdfplumber or PyMuPDF for text extraction — neither of which is guaranteed without setup. If files are provided and the environment is configured, this is straightforward; if not, there's a dependency gap.
Error Cost
Low
Output is a CSV that a human can audit before any downstream use. Extraction errors are surfaced by the confidence score and failure flags, making mistakes visible and correctable rather than silently damaging.
Human Judgment Required
Low
Regex pattern design, PDF parsing logic, and CSV formatting are all deterministic engineering tasks. Edge cases like ambiguous date formats or missing fields are handled by flagging, not by subjective human calls.