Industries / Industrial equipment and services / Industrial services / Process safety consulting
Layers of protection
Layers of protection work decides how many independent safeguards a hazard needs, and shows that each one really is independent.
5
Milestones achieved, from 1978 to 2009.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1978
Keep less, leak less
Trevor Kletz of ICI argued in "What You Don't Have, Can't Leak" that the first safeguard is to take the hazard away: hold less hazardous material, use a safer substitute, run under milder conditions, or make the plant simpler. The lecture was given to the Society of Chemical Industry in December 1977 and printed in Chemistry & Industry in May 1978. Engineers at AcuTech Group, writing in AIChE's magazine CEP in 2023, call it the first clear and concise discussion of inherently safer plants and say the 1974 Flixborough explosion inspired it.
Credit
Trevor Kletz, ICI
1979
Bow-ties in ICI's notes
A bow-tie diagram puts one hazardous event in the middle, its causes on the left with the barriers that stop them, and its consequences on the right with the barriers that limit them. David Gill of ICI's Mond Division wrote an internal paper on the method, and it formed part of ICI's 1979 Hazard Analysis course notes. Shell staff first saw it on a course Gill co-taught in the Netherlands in 1991, and in the early 1990s the Royal Dutch/Shell Group made it the company standard for analysing risk.
Credit
David Gill, ICI; Royal Dutch/Shell
2001
A book for counting layers
Layer of protection analysis (LOPA) takes one accident scenario and estimates how often its cause occurs. It then credits each independent protection layer with an order-of-magnitude chance of failing, and checks the result against a risk target. Companies built their own versions during the 1990s, papers appeared around 1997, and a CCPS conference in October 1997 recommended a book. CCPS, the Center for Chemical Process Safety, published "Layer of Protection Analysis: Simplified Process Risk Assessment" in 2001.
Credit
Center for Chemical Process Safety (CCPS)
2008
New Jersey asks for alternatives
On 5 May 2008 New Jersey adopted a rule under its Toxic Catastrophe Prevention Act. Every covered site had to evaluate inherently safer technology: less hazardous material, substitutes, milder conditions, and designs that cut equipment failure and human error. Plants had to evaluate these options, not adopt them. All 87 covered sites submitted reviews. The state's environment department reviewed 19 of the 42 reviews from sites newly covered (water, food, petroleum and energy) and found all 19 deficient, most often for missing alternatives and for options called infeasible without support. These figures are from the department's own testimony to Congress in June 2009.
LessonA duty to look for safer designs works only when someone checks that every alternative was considered.
Credit
New Jersey Department of Environmental Protection
2009
LOPA studies checked by regulators
After the Buncefield fire, the UK Health and Safety Laboratory reviewed seven LOPA studies on tank overfill that fuel storage sites had sent to HSE, most of them written by consultants. The data varied widely in quality and some was inappropriate. Dependencies between causes and safeguards were handled inconsistently, and some studies split a cause into parts and treated the parts as independent, which could give unrealistically low frequencies. Most did no sensitivity check.
LessonA LOPA is only as good as its numbers and its independence claims, so a reviewer has to test both.
Credit
Colin Chambers, Jill Wilday and Shane Turner, Health and Safety Laboratory, for HSE
Watch for
Bow-ties drafted from tables
Three open-weight language models (Mistral, Qwen-2.5 and LLaMA-3) drafted bow-tie diagrams from failure mode and effects analysis (FMEA) tables. With strict prompts they often scored above 0.8 for nodes and links, using F1, a 0-to-1 measure of agreement with diagrams drawn by experts. On narrative text all the models struggled, inventing nodes and assigning wrong roles.
Achieved whenanother group reaches similar scores on a different set of FMEA tables, or a plant reports using the drafts in routine work.
Claimed by
Priyank Venkatesh, Jules Oudmans and Florian Zurfluh, PHM Society European Conference, July 2026
LOPA studies open to agents
Kenexis says its Open-PHA software stores HAZOP and LOPA studies as open JSON text, that an AI agent can read a whole study, reason about its causes, consequences and safeguards, and write changes back, and that agents can import studies made in older software.
Achieved whena user outside Kenexis publishes a migration or review done this way, with a count of errors against the original study.
Claimed by
Kenexis, brochure dated July 2026
Open
01
Flag safeguards that share parts
A checker that reads LOPA worksheets and flags every credited protection layer that is not independent of the initiating cause or of another credited layer: an alarm on the same controller as the control loop that failed, a trip that uses the same transmitter as the alarm, or one operator response counted twice.
The briefCredit · unclaimed02
Bow-ties from accident reports
A tool that reads a published investigation report and drafts its bow-tie: the top event, the causes, the barriers that were in place, and which of them failed and why. The draft is scored against the investigators' findings.
The briefCredit · unclaimed
Open 01 · the brief
Flag safeguards that share parts
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A checker that reads LOPA worksheets and flags every credited protection layer that is not independent of the initiating cause or of another credited layer: an alarm on the same controller as the control loop that failed, a trip that uses the same transmitter as the alarm, or one operator response counted twice.
Why it is still open
HSE's review found inconsistent handling of these dependencies, and causes split into parts that were assumed independent. LOPA software calculates the risk but leaves independence to the team: GE Vernova's APM, for one, puts it to the analyst as a yes-or-no question, whether the protection layer is "Independent of the initiating cause of the hazardous scenario". Kenexis's 2026 brochure says only that an agent could read the studies.
Built with what exists
Kenexis Open-PHA Desktop, free on Windows, Mac and Linux, which stores studies as JSON; Python rules for shared sensors, logic solvers and final elements; a model such as Claude or Gemini to read the free-text safeguard descriptions, citing the worksheet row and rule behind each flag (Note: "Findings that cite notes").
First step
Rewrite the dependency problems HSE describes in its seven case reviews as ten test scenarios in Open-PHA Desktop.
Done looks like
A public set of 100 LOPA scenarios written by two process safety engineers, 40 with one seeded dependency error and 60 clean. The checker flags at least 36 of the 40 and wrongly flags no more than 6 of the 60, and the set and the scores are published.
Open 02 · the brief
Bow-ties from accident reports
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that reads a published investigation report and drafts its bow-tie: the top event, the causes, the barriers that were in place, and which of them failed and why. The draft is scored against the investigators' findings.
Why it is still open
The authors of the 2026 bow-tie study found that all their models struggled on narrative text, and their working method covers FMEA tables only.
Built with what exists
US Chemical Safety Board final reports, which are public PDFs; a bow-tie schema in JSON; Claude, Gemini or open models such as Qwen; the study's node and link F1 scoring.
First step
Draw the reference bow-tie for the CSB's Formosa Plastics Illiopolis report by hand, run one model on the report, and score the result.
Done looks like
20 CSB reports, each with a reference bow-tie drawn by two engineers from the report's stated causes. The tool averages a node F1 of at least 0.7, and in at least 18 of the 20 it names no barrier that the report does not mention.