Industries / Industrial equipment and services / Industrial services / Non-destructive testing
Reading and reporting results
This is the work after the scan: deciding what a signal or image means, proving how often a method finds a flaw of a given size, certifying the people who make the calls, and keeping honest records.
7
Milestones achieved, from 1968 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1968
Employers certify their own
In 1968 ASNT, the American Society for Nondestructive Testing, published the first edition of Recommended Practice No. SNT-TC-1A, which set out three levels of inspector, Level I, II and III. Under it, each employer writes its own "written practice" and decides whether its staff meet it; ASNT calls this employer-based qualification. In 1977 ASNT added its own central exam for Level III.
Credit
ASNT
1974
Detection put in numbers
NASA planned to design the Space Shuttle around the crack size that inspection could be counted on to find. For that, Ward Rummel and colleagues at Martin Marietta inspected 118 aluminum specimens holding 328 fatigue cracks with X-rays, penetrant, ultrasound, eddy current and two other methods, then broke the specimens open to measure every crack. From the hits and misses they set a threshold crack size for each method at 95% probability and 95% confidence.
LessonThe authors wrote that inspection results can serve as design limits only with "rigorous inspection qualification and application programs".
Credit
Ward D. Rummel, Paul H. Todd Jr., Sandor A. Frecska and Richard A. Rathke, Martin Marietta, for NASA (report CR-2369)
2002
Cracks erased from records
In August 2002 it emerged that Tokyo Electric Power Co. (TEPCO) had removed signs of stress corrosion cracking in reactor core shrouds and recirculation pipes from its inspection records and had never given them to regulators. An internal probe found that TEPCO managers had told inspectors from its contractor, General Electric International, to falsify specific reports at its Fukushima and Kashiwazaki-Kariwa plants. TEPCO's top management resigned.
LessonTEPCO's recovery included an independent assessment body and an open program for logging nonconformances, so that findings could not be dropped quietly.
Credit
a failure; described by TEPCO's S. Kawamura in a 2006 IAEA paper and reported by the Japan Times
2004
Left off the list
On 9 August 2004 a pipe burst in the steam and water system of Kansai Electric Power's Mihama-3 reactor in Japan, and the steam killed five workers and seriously injured six. Fast-flowing water just past a flow-measuring orifice had worn the carbon steel wall thin from inside, a process called erosion corrosion. Inspection guidelines called for tracking the remaining life of pipe at such spots, but the inspection list drawn up by Mitsubishi Heavy Industries left this section out, and it had not been inspected.
LessonAn inspection program is only as complete as its list of locations; the same gap had been found and fixed at four other reactors, but not at Mihama-3.
Credit
a failure; investigated by a committee under Japan's Nuclear and Industrial Safety Agency (final report 30 March 2005), as summarized by the Citizens' Nuclear Information Center
2009
One handbook for proof
The US Department of Defense handbook MIL-HDBK-1823A, dated 7 April 2009, sets out how to run a probability of detection (POD) study: how to make specimens with flaws of known size, collect hit-or-miss calls or signal readings, and fit a curve. Its key figure is a90/95, the flaw size found 90% of the time, stated at 95% confidence, which the handbook says "has become a de facto design criterion". It comes with POD software, mh1823, that runs in the free R statistics system.
Credit
US Department of Defense, with comments handled at Wright-Patterson Air Force Base
2012
One standard for certification
EN ISO 9712, published in June 2012, replaced both the European standard EN 473 and the earlier international ISO 9712 as the rule for qualifying and certifying inspectors. Under such schemes a certification body, not the employer, issues the certificate, in ASNT's description. The UK's PCN scheme, run by the British Institute of Non-Destructive Testing (BINDT), says it was the first fully accredited personnel certification scheme under the 2012 standard.
Credit
the CEN and ISO technical committees for non-destructive testing
2025
AI in a qualified inspection
On 3 June 2025, during an outage at the Ringhals nuclear plant, Sweden's second-largest power station, AI screened the data from a qualified ultrasonic inspection of reactor vessel head penetrations, a first by the account of ASNT's journal. A qualified inspection is one whose procedure, equipment and people have passed a formal test on specimens with known flaws. The system, from Trueflaw of Finland and EPRI, had been qualified by the Swedish Qualification Centre (SQC). It flags possible flaws, and human analysts make the final call.
Credit
Trueflaw and EPRI; inspection by Wesdyne; qualified by SQC
Watch for
SNT-TC-1A to be rewritten
ASNT says it is reworking SNT-TC-1A, which it says has guided employers for nearly 60 years. It wants "a firm line between employer-based qualification and independent certification", qualification based more on shown performance, and added guidance and templates. It says the name stays, nothing changes yet, and training hours and the three levels are still undecided.
Achieved whenASNT publishes a revised edition with effective dates and employers or certification bodies adopt it.
Claimed by
ASNT, on a page last updated 10 August 2026, with expert input due 1 October 2026
Review work cut 95%
Vattenfall says the Trueflaw and EPRI system cut the ultrasonic data review workload at Ringhals 3 by up to 95%, and saved about a week of schedule when it was used again at Ringhals 4.
Achieved whenan outside body, such as a regulator or qualification center, publishes review hours and missed or extra calls for AI-assisted and fully manual reviews of the same data.
Claimed by
Vattenfall's Ola Johansson, quoted in an article by EPRI's Michael Ruszkowski in Nuclear News, October 2026
Open
01
A ledger of missed flaws
A public table of failures, across industries, in which the part had been inspected without cutting before it failed and the inspection missed the flaw, built from official investigation reports. Each row gives the method used, the flaw's size at failure and, where the report says, at the last inspection, and a reason from a fixed list: area not covered, wrong method or procedure, signal seen but misjudged, equipment, access, or records.
The briefCredit · unclaimed02
Open exam for AI readers
An open exam of 600 items in which a person or an AI model reads real inspection data, such as ultrasonic scans, eddy current signals, X-ray images and penetrant or visual photos, and answers fixed questions: is there a flaw, what kind, where, and does it pass the stated acceptance rule. Two certified Level 3 inspectors set each answer, and certified Level 2 inspectors take the same exam to give a human baseline.
The briefCredit · unclaimed
Open 01 · the brief
A ledger of missed flaws
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table of failures, across industries, in which the part had been inspected without cutting before it failed and the inspection missed the flaw, built from official investigation reports. Each row gives the method used, the flaw's size at failure and, where the report says, at the last inspection, and a reason from a fixed list: area not covered, wrong method or procedure, signal seen but misjudged, equipment, access, or records.
Why it is still open
Studies so far cover one trade. The UK's Health and Safety Executive (HSE) reviewed 47 incidents in which weld quality contributed and found too little inspection coverage among the most common causes (2024). Reporting on a 2018 jet engine fan blade failure noted that inspections twice missed a crack that had begun years earlier, but it named cases without counting them. Searches in November 2026 found no table across industries coded by method and reason.
Built with what exists
Public reports from bodies such as the US National Transportation Safety Board and Chemical Safety Board, the UK's Air and Rail Accidents Investigation Branches, Australia's ATSB, Canada's TSB and the HSE; an AI agent that reads each report and proposes a coding, quoting the passage behind each field, checked by two people (see the Note "Seven more flagged", a draft, on a Claude-based reviewer that raised issues people had missed in drug-making records); Python for agreement statistics.
First step
Code ten well-documented cases by hand, among them the 2018 fan blade failure and the 2004 Mihama-3 pipe, and write down the coding rules.
Done looks like
At least 150 failures coded, each by two people, with their agreement on the reason code (Cohen's kappa, a measure of agreement beyond chance) of at least 0.7; the AI agent's coding scored against theirs; the table released under an open license.
Open 02 · the brief
Open exam for AI readers
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open exam of 600 items in which a person or an AI model reads real inspection data, such as ultrasonic scans, eddy current signals, X-ray images and penetrant or visual photos, and answers fixed questions: is there a flaw, what kind, where, and does it pass the stated acceptance rule. Two certified Level 3 inspectors set each answer, and certified Level 2 inspectors take the same exam to give a human baseline.
Why it is still open
A July 2025 pilot asked nine language models to describe five contour maps from bridge inspections and judged their write-ups, with no answer key from certified inspectors and no human baseline. In August 2026 Mehrdad Shafiei Dizaji and Hoda Azari described ChatNDE, a model that writes captions for NDE images and answers questions about them; they scored its captions by word overlap with experts' written descriptions, call its accuracy low, and report no human baseline. A Chinese set of machine-written questions and answers on radiography theory has been on Hugging Face since April 2025, as text only. A March 2026 study ran 60 multimodal models through 1,171 certification exam questions, but for software testing. Certification exams such as ASNT's use closed question banks. Searches of the web, arXiv and Crossref in November 2026 found no open exam on reading inspection data.
Built with what exists
Practice data donated by training schools and inspection firms with permission to publish; public sets such as GDXray, whose weld X-ray images came from Germany's Federal Institute for Materials Research and Testing (BAM); Label Studio for answer keys; vision-language models, and Google's Gemini Decision 1, which returns a probability for each fixed answer instead of writing text (see the Note "A thousand questions, one pass", a draft); Python for scoring.
First step
Ask two inspection training schools for 50 practice items each, with answers and permission to publish, and have two Level 3 inspectors check the answers separately.
Done looks like
600 items across at least five methods released with answer keys; scores for at least six AI models and 20 certified Level 2 inspectors; and, for each, the share of flaws that should have been rejected but were passed.