Turn the report pile into one trendable table.
UT surveys, PEC screens and walkdown checklists arrive as PDFs, scans and spreadsheets. The readings inside them are the history every corrosion rate and remaining-life figure depends on, and today someone re-keys them. This digitizer is designed to read each document, fill one record per reading, and say how sure it is about every field.
Fields it is sure about are accepted. Uncertain ones go to a short review queue that points back to the source cell. Nothing is silently guessed. What you see below is a simulated extraction over three synthetic documents; the pilot would run document models on your own reports and measure the result.
Inspection deliverables as they exist: UT thickness surveys, PEC CUI screening reports and CUI walkdown checklists, as PDF, scan or spreadsheet.
One row per reading with units normalized, source page and a confidence score, plus a review queue for the uncertain fields. CSV out to your IDMS or into the CUI Priority Workbench.
Fifty pages of your own reports. We hand-label ten as hidden truth and report precision and recall per field type, not a vendor's advertised number.
Documents and extractionThree sample documents, rendered from the same seeded data the extraction reveals.
Pick a document, run the extraction, then work the review queue. Every highlight on the left is one row in the table on the right.
These pages are HTML mock-ups drawn from the seeded data, not real reports. The panel is designed to show your PDF page with a bounding box on each extracted field in the pilot.
Extraction settings
Confidence is per field. At or above the auto-accept level a field is taken as read; between that and 0.60 it goes to review; below 0.60 it is treated as missing. Audit sampling pulls a random share of auto-accepted fields into the queue as a check on the model.
| Field | Extracted | Normalized | Conf. | Status |
|---|
Review queue
Accept the proposed value or type a correction. Corrections rewrite the table, the tiles and the CSV, and are logged against the source page.
Domain checks the confidence score cannot make
Where extraction goes wrongSix documented failure modes, each planted in one of the sample documents.
Generic OCR gets the characters right and the structure wrong. Each card names the failure, what the digitizer does about it, and where to see it above.
Failure modes as documented by Reducto (merged cells silently re-assigning values, tables continuing on the next page returning as two unrelated tables, multi-column interleaving, scan artifacts), imagetotable (checkboxes ignored by OCR, readings hallucinated from embedded photos, handwriting) and Microsoft (merged cells and rows split across pages carry lower confidence). Unit mixing is a recognized risk with no citable statistic; the documents in this domain mix mm and inches routinely.
Confidence per field, a review queue, a random audit, and sanity checks from the trade.
The routing pattern is the one the hyperscalers document for their own document-AI services. The sanity checks come from what inspection engineers say actually goes wrong with thickness data.
0.95 to accept, 0.60 to call it missing, per field type.
Microsoft's guidance: a field confidence of 0.95 means the prediction is right about 19 times in 20; aim for 0.80 or better and close to 1.0 for sensitive records. AWS's reference routing sends key fields below 0.60 to a person and words below 0.85 or 0.90. Secondary guides put production thresholds at 0.85 to 0.95 depending on field criticality. Readings, units, CML ids and dates are the critical fields here; free-text remarks are not.
Check a slice of what the model was sure about.
AWS's example routes a random 5 % of inferences to a reviewer regardless of confidence. It catches the confident mistake, which a threshold never will, and gives you a running accuracy figure on your own documents rather than a vendor benchmark.
Readings still go onto paper, then get keyed in.
Sensor Networks (an installed-sensor vendor) writes in its 2019 white paper on CML data collection that even with programmable data loggers, writing thickness readings on paper is still the norm, that keyed-in readings suffer keying errors and misalignment, that technicians lose their place in the route and readings land on the wrong TML, and that some sites may take over a million manual UT readings a year. The 38DL PLUS gauge itself exports CSV; where that file exists we ingest it and skip OCR entirely.
Most of the budget, and a lot of false alarms.
That is why the checks panel above recomputes the rate since the previous survey for every reading and flags anything over 50 mpy before it reaches your trend.
One row per reading, whatever the document looked like.
The schema is modelled first and the documents are mapped onto it, so a PEC grid cell, a UT reading and a walkdown finding land in the same table with the same provenance columns. The CSV button above writes exactly these columns.
| Column | Example | Where it comes from in the samples | Needed |
|---|---|---|---|
| site | Sample site | Report header; can be hashed on your side | Required |
| unit | U-310 crude preheat | Report header | Required |
| line_or_equipment | P-310-014-6"-CS-150 | Header of all three documents | Required |
| cml_id | CML-03; A4/180° for a PEC cell; S2 for a walkdown segment | CML column, grid row and column, segment column | Required |
| location | Elbow E1, extrados | Location column, grid position, rack bay | Required |
| method | UT / PEC / Visual | Document type | Required |
| nominal | 7.11 mm | UT nominal column; PEC reference wall | Required for UT and PEC |
| t_min | 3.56 mm | UT t_min column | Helpful |
| reading | 6.42; 71.4; "Open lap" | Each of the four UT readings, each PEC cell, the walkdown damage type | Required |
| units | mm; % of nominal; text | Instrument line, PEC legend; original preserved, canonical column added | Required |
| date | 2026-06-14 | Report header; previous-survey date kept as its own reading row | Required |
| instrument | 38DL PLUS S/N 1174-08; PEC-089 probe, 87 mm footprint | Header; for PEC the footprint and lift-off travel with the reading | Helpful |
| technician | Initials and level | Sign-off block | Helpful |
| source_document | UT-310-0614.pdf | File name; original kept immutable | Required |
| page | 2 | Page the value was read from; the pilot is designed to add a bounding box | Required |
| confidence | 0.97, or 1.00 (corrected) after review | Model score, replaced by the reviewer's decision when one is made | Required |
Fifty pages of your reports, ten of them hand-labelled, one written scorecard.
Public benchmarks put the best table parser around 90 % on hard tables and the hyperscalers between 65 and 83 % (vendor-run figures). The only number that matters is the one on your templates, so that is the one the pilot produces.
- STEP 1You · about an hourSend 50 pages
A mix of UT survey, PEC report and walkdown pages, anonymized or not, under NDA. Native Excel or gauge exports welcome; they skip OCR entirely.
- STEP 2Clear Anvil · one weekLabel 10, extract 50
Ten pages are hand-keyed into the target schema and held back as hidden truth. Document models would run on all fifty and return per-field confidence and page citations.
- STEP 3Clear Anvil · a few daysWritten scorecard
Precision and recall per field type against the ten hidden pages, the review minutes per page at the chosen thresholds, and the failure modes your templates actually trigger.
- STEP 4Both · your callBacklog run, or stop
If the numbers hold, one unit's history is digitized with your reviewer in the loop and delivered as CSV into whatever IDMS your client runs.
Sources. Technical Toolboxes, API 570 module (up to four TML readings per CML; lowest drives rate and remaining life) · Olympus 38DL PLUS manual (datalogger file types, CSV export) · Sensor Networks, Automating and optimizing the CML data collection process (vendor white paper) · Digital Refining, Optimising information from thickness data · Eddyfi, PEC technology (footprint = 0.65 × lift-off + zero-lift-off footprint; averaged thickness) · Eddyfi Lyft specifications · IRISNDT PEC datasheet (3.42 in footprint at 1.5 in insulation with PEC-089; 50 % overlap; ±10 % sizing) · API RP 583 public excerpt (jacketing, caulking, PEC AWT display) · MFE, CUI inspection guide (walkdown items) · Microsoft Learn, Document Intelligence accuracy and confidence · AWS, Textract human-review activation conditions (thresholds, 5 % sampling) · Reducto, RD-TableBench (vendor-run) · Reducto, table extraction on scanned PDFs · imagetotable, inspection report extraction guide (vendor blog) · Triotech, IDP comparison (secondary; production accuracy and thresholds) · Verdantix, Smart Innovators: Asset Integrity and Risk Management Software, June 2025 · InspectionsTrack, the hidden data problem in NDT inspections (vendor blog). Verdantix already names Imrandd (OCR of historical records) and Cenosco (GenAI summaries across decades of reports) as vendors in this space; InspectionsTrack lists spreadsheets, PDFs and shared folders as where service-company deliverables still live. This module's difference is the CUI-specific schema feeding the Priority Workbench, not OCR itself.
Send one export.
Get a written scorecard back.
Reply to the thread or write to the address below. No meetings needed unless you want one.
jamil@clearanvil.com