AI inspection table extraction
Extract inspection readings without changing row identity, decimal values, missing measurements or the units attached to a table.
Synthetic evaluation input
record,reading_mm,note I-01,3.05,measured I-02,,not measured I-03,3.50,measured No acceptance limit or minimum thickness is supplied.
Original Inzonex training example, version 1.0. No customer document or real equipment data. No model has been scored on this page.
Expected fields
| Field | Reference answer |
|---|---|
| I-01_mm | 3.05 |
| I-02_mm | null |
| I-03_mm | 3.5 |
| unit | "mm" |
| acceptance_limit_mm | null |
Check a structured response
The comparison runs in your browser against this answer key. It checks exact fields and values, not the accuracy of an explanation or a model's general ability.
Recommended workflow
Recover the table structure first, then validate values against row IDs and the measurement unit. Preserve missing readings as null, not zero.
Table extraction can fail without producing obviously broken text. A value can move into the next row, a decimal separator can disappear, or a dash can become zero. The resulting spreadsheet may look orderly while misrepresenting the inspection record.
The synthetic packet contains a missing measurement and two close decimal values. The check requires the original row IDs and the missing-value distinction. No minimum-thickness threshold is supplied, so the model must not classify equipment condition.
For a scanned report, repeat the evaluation with representative page images and check headers carried across page breaks. This text packet is the semantic reference stage; it is not evidence of OCR performance on a real scan.
Failure checks
- Keep 3.05 distinct from 3.50.
- Return null for the missing reading.
- Do not invent an acceptance limit or fitness-for-service decision.
What this exercise does not prove
Not a fitness-for-service calculation, inspection recommendation or OCR benchmark. No equipment acceptance decision follows from this packet.
For an actual evaluation, keep a separate held-out set, record tool/model version and settings, and log raw outputs, corrections, elapsed time and actual charges. Do not compare tools tested on different inputs as though they ran the same benchmark.
What is checked in AI inspection table extraction?
Extract inspection readings without changing row identity, decimal values, missing measurements or the units attached to a table.
What does AI inspection table extraction not prove?
Not a fitness-for-service calculation, inspection recommendation or OCR benchmark. No equipment acceptance decision follows from this packet.
Related tasks
Match this workflow to your data requirements
Methodology and reuse
These packets and answer keys are original Inzonex educational material, licensed CC BY 4.0. Attribute Inzonex and link to this task page when reusing the packet. The licence does not cover third-party material linked from this site.
NIST AI 600-1: background on generative AI evaluation and risk. This exercise is not NIST-certified.