BRENNEN

Industries / Industrial equipment and services / Industrial services / Process safety consulting

Incident investigation

Incident investigation finds out why an accident happened, beyond the last error made, so that it does not happen again.

5

Milestones achieved, from 1973 to 2020.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1973

    A tree of management questions

    W. G. Johnson of Aerojet Nuclear Company, the prime operating contractor at the US government's reactor test site in Idaho, published MORT, the Management Oversight and Risk Tree, in February 1973. MORT is a logic diagram, similar to a fault tree, that leads an investigator through what management should have had in place. A 1975 manual applied it to accident investigation, and the 1992 user's manual lists over 1,500 elements of an ideal safety program.

    osti.govosti.govosti.gov

    Credit

    W. G. Johnson, Aerojet Nuclear; System Safety Development Center

  2. 1982

    Europe pools its accident reports

    The EU's first Seveso Directive set up MARS, the Major Accident Reporting System, in 1982. Member countries report major accidents and near misses involving dangerous substances to it, so that lessons can be shared. Once it went online it was renamed eMARS. The European Commission's Joint Research Centre ran it until the end of 2025, and the European Environment Agency has run it since 1 January 2026.

    industry.eea.europa.eu

    Credit

    European Commission

  3. 1997

    Causes mapped up to government

    Jens Rasmussen set out the AcciMap in "Risk management in a dynamic society: a modelling problem" (Safety Science, 1997). An AcciMap is a diagram that sorts an accident's causes into levels: physical events and front-line work at the bottom, the company in the middle, and regulators and government at the top. Its purpose is to keep blame from stopping at the last worker involved. The CSB used an AcciMap for the causal analysis in its 2016 Williams Geismar case study.

    doi.orgen.wikipedia.orgcsb.gov

    Credit

    Jens Rasmussen. 00052-0)

  4. 1998

    A board that only investigates

    The US Chemical Safety Board was authorized by the Clean Air Act Amendments of 1990 and began work in January 1998. It was modelled on the National Transportation Safety Board, and Congress kept its investigations separate from the rulemaking and enforcement of EPA and OSHA. It issues no fines, only recommendations. On 19 November 2026 its site listed 1,061 recommendations, 932 of them closed and 129 open.

    csb.gov

    Credit

    US Congress; US Chemical Safety Board

  5. 2020

    Reporting rule by court order

    The 1990 law told the board to require owners to report accidental releases. In Air Alliance Houston v. CSB (4 February 2019), a federal court ordered the board to issue that rule within 12 months. The rule took effect on 23 March 2020 and gives owners eight hours to report a release that falls under the board's jurisdiction. The board noted that the data a distributed control system records about a release may be overwritten within 24 to 48 hours.

    LessonA duty written into law does nothing until someone writes the rule that carries it out.

    govinfo.gov

    Credit

    US Chemical Safety Board; US District Court for the District of Columbia

Watch for

  1. Board's funding in doubt

    The White House's FY 2026 budget of May 2025 proposed closing the CSB. Congress instead gave it $14 million, in a bill signed on 23 January 2026. The board's own FY 2027 budget request says "The President's Budget proposes $0 for the CSB's FY 2027 budget."

    Achieved whena FY 2027 appropriation is enacted that either funds the board or ends it.

    kslaw.comcsb.gov

    Claimed by

    the White House, in its budgets of May 2025 and for FY 2027

  2. AI agents read CSB reports

    Wen, Parsaee and Sajid built two GPT agents that pull time-ordered warning signs from 100 CSB reports. They report precision of 0.88 and recall of 0.84 across 90 cases, measured against their own expert baselines.

    Achieved whenanother group repeats the test on reports outside the set and publishes its scores.

    doi.org

    Claimed by

    the authors, in the journal AI, November 2025

Open

  1. 01

    A double-coded cause set

    200 public accident reports from the CSB and eMARS, each coded by two trained investigators against one published scheme of causes, with their agreement reported as Cohen's kappa, a standard statistic for agreement between two coders. The set is released openly so that AI coders can be scored against it.

    The brief
    Credit · unclaimed
  2. 02

    What the first report missed

    A set of accidents that each have both an early report (the company's own or the first official account) and a later independent investigation. A tool is given only the early report and scored on how many of the causes the later investigation added it can raise.

    The brief
    Credit · unclaimed

Open 01 · the brief

A double-coded cause set

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

200 public accident reports from the CSB and eMARS, each coded by two trained investigators against one published scheme of causes, with their agreement reported as Cohen's kappa, a standard statistic for agreement between two coders. The set is released openly so that AI coders can be scored against it.

Why it is still open

The CSB study above scored its agents against the authors' own baselines, and a 2025 study that tried AI on AcciMap analysis of a single aviation accident found it weak at drawing the map. No double-coded process safety set with published agreement figures was found in November 2026.

doi.orgrpsonline.com.sg

Built with what exists

Public CSB report PDFs and eMARS reports; Label Studio, an open-source labelling tool; current large language models as the coders under test.

  1. First step

    Have two volunteers code 20 CSB reports and work out their kappa.

  2. Done looks like

    200 reports coded by two people, kappa published for each category, and scores published for at least three AI tools.

Open 02 · the brief

What the first report missed

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A set of accidents that each have both an early report (the company's own or the first official account) and a later independent investigation. A tool is given only the early report and scored on how many of the causes the later investigation added it can raise.

Why it is still open

The Baker Panel found that BP had no effective root cause analysis for "systemic causal factors," so its fixes could treat "immediate or superficial causes." No published test on pairs of process safety reports was found.

csb.gov

Built with what exists

BP's own Texas City report of 9 December 2005 and the CSB's report on the same accident; other pairs from the CSB and HSE; the AcciMap levels as the scoring frame; language model agents.

  1. First step

    List the CSB investigations for which the company's own report is public.

  2. Done looks like

    15 pairs, with the tool raising at least 60% of the company-level and regulator-level causes the later investigation added, and two investigators agreeing on the scoring.