Would you have used this output?
Below are extracts from the barrier analysis an AI workflow produced for this incident, from the CSB report alone. Some are sound. Some are not. Decide for each one whether you would put it in a client deliverable, then see what the source actually says — and what the rest of the room decided.
A guard fitted for food safety became the explosion chamber
Fourteen people were killed: twelve Imperial Sugar employees and two contractors. Eight died at the refinery; six died afterwards at the Joseph M. Still Burn Center, the last of them six months after the explosion. Two of them had gone back into the building to try to reach their colleagues. Thirty-six more were treated for serious burns and injuries.
The CSB investigation report does not publish their names.
- Fatalities
- 1412 employees, 2 contractors
- Injured
- 36serious burns and injuries
- Date and time
- 7 Feb 2008approx. 19:15 local
- Dust violence
- 139bar·m/s KSt · class ST2
- Explosible above
- 95g/m3 minimum concentration
- Documented since
- 1961internal memos on the dust hazard
What happened
An overheated bearing in the enclosed steel belt conveyor beneath silos 1 and 2 most likely ignited a primary dust explosion. The pressure wave travelled up through the pantleg rooms into the packing buildings, lifting the sugar dust that had accumulated on floors, beams and elevated surfaces. That dust fed a chain of secondary explosions through both packing buildings, the palletizer room and on into the refinery and bulk loading areas. Concrete floors heaved, brick walls collapsed across the exits, the sprinkler pipes ruptured and the emergency lighting failed.
Why it still matters
Nothing here was unknown. The company’s own material safety data sheets stated the explosible concentration of sugar dust. Internal memoranda had described the hazard since 1961, and a refinery engineer had written down the mechanism of secondary explosions in 1967. An insurance audit ten months before the explosion recorded the housekeeping and said nothing about dust. The information was complete and correct at every level. What failed was the translation of information into action — and the modification that created the explosible space was never assessed as a change.
Three silos, one tunnel, and a modification nobody assessed
The Port Wentworth refinery dated from 1916. Granulated sugar left the silos on a steel belt conveyor running in a tunnel beneath silos 1 and 2, and reached the packing buildings through spaces that were open to one another. Click a position to see its data and its part in the event.
Select a position
The drawing is an original reconstruction from the descriptions and dimensions in CSB Report 2008-05-I-GA. It is not a copy of a figure in the report.
Sugar is a fuel, and a data sheet had already said so
For the process engineer
Fuel: a combustible solid, finely divided
Dispersion and concentration above the MEC
Confinement: the reason the enclosure mattered
Ignition: an ordinary hot surface
Primary to secondary: where the fourteen deaths came from
The five conditions, and which ones were present
Measured values
What the safety data sheet said
“Explosion: NFPA Class 2 Group G. Airborne sugar dust accumulation ignition temperature is 370 °C. At airborne concentrations of 0.045 gm/L or higher, sugar dust accumulations are explosive.” This sheet was on file at Imperial Sugar.
What happened anyway
The information was correct, held internally, and specific enough to act on. It had been reinforced by internal memoranda since 1961, by a 1967 memorandum that described the mechanism of secondary explosions, by dust explosions in 1961 and 1968, by a severe burn injury at another Imperial site in 1998, and by small fires in the packing buildings through 2006 and 2007. A 2006 corporate memorandum recorded that no site had a formal housekeeping policy; the draft policy attached to it did not mention sugar dust.
Forty-seven years, then ninety seconds
Most event timelines cover the last hours. This one has to cover decades, because the decisive opportunities were not on the evening of 7 February 2008 — they were in 1961, 1967, 2006 and 2007. Step through it and open “what could still have been done” at each point.
What could still have been done at this point
Twelve barriers, and the one that was never asked for
This bow-tie was built by a process safety engineer from CSB Report 2008-05-I-GA. It is the reference standard on this page: the AI outputs in the LIVE section are measured against it, not the other way round. Click a barrier for what it should have done, what it did, and what follows for other plants.
Prevention barriers
Mitigation barriers
Select a barrier
Status codes follow the rhysafe barrier taxonomy. Colour is never the only carrier of meaning: each status also has a symbol and a word.
Six levels, and every one of them has a finding
An AcciMap reads from the sharp end upward. If a level comes out empty, the analysis is usually not finished. None of these levels is empty, which is the point: a plant does not arrive at forty-seven years of documented, unaddressed hazard through one bad decision.
6Execution — the shift4
- Operators knew that conveyor bearings sometimes became very hot, and had no route by which that knowledge became a safety-critical maintenance item.
- Spillage from conveyor blockages was cleared as a housekeeping task, not logged as a loss of containment of a combustible material.
- Alerting on the night depended on radios, mobile phones and people warning each other face to face.
- Workers were told to familiarise themselves with evacuation routes; no work-location-specific evacuation training was given and no drills were held.
5Plant and technology5
- The 2007 stainless steel enclosure reduced the free volume around the belt to roughly one tenth of the tunnel, with no extraction and no explosion venting.
- Twelve-inch openings connected the pantleg rooms to the conveyor tunnel, giving the pressure wave a direct path into the packing buildings.
- The 1916 structure interconnected the silos, packing buildings, palletizer room and refinery, with timber roof and ceiling elements.
- Sprinkler pipework and hydrants shared the space with the hazard and were lost in the event.
- No hazardous area classification existed, and no requirement for electrical equipment rated for dust locations.
4Site management4
- Normalisation of deviance: years of small fires and dust explosions without catastrophic outcome were absorbed as the normal condition of the plant.
- An internal inspection six weeks before the explosion recorded that many tons of spilled sugar had to be routinely removed from the floors.
- Monthly safety training covered fourteen topics; a review of more than ten thousand pages of training material found no combustible dust content since 2005.
- The emergency procedure specified an intercom that was not used in the buildings it covered.
3Corporate4
- A 2006 corporate memorandum recorded that no site had a formal sanitation or housekeeping policy; the attached draft policy did not mention sugar dust.
- A severe burn injury from a sugar dust explosion at the Sugar Land site in 1998 did not produce a group-wide dust programme.
- Internal correspondence naming the hazard, and the mechanism of secondary explosions, existed from 1961 and 1967 and was never converted into engineering requirements.
- The company's own material safety data sheets carried the explosible concentration and the dust cloud ignition temperature.
2Industry, insurers and auditors4
- The April 2007 Zurich Services audit examined building construction, occupancies, storage and general housekeeping, and made no mention of spilled sugar or dust accumulation, although CSB photographs show significant accumulations were present.
- Zurich trained its own auditors on combustible dust in September 2007 and did not share the training material with Imperial Sugar or its other clients.
- Food sector audit regimes addressed product contamination; the same dust that was a contamination concern was not assessed as an explosion hazard.
- The CSB directed recommendations to AIB International, the American Bakers Association, RIMS and Zurich Services — an unusual spread that reflects how much of the gap sat outside the operating company.
1Regulation4
- No comprehensive combustible dust standard covered general industry in the United States in February 2008.
- A grain handling standard had existed since 1983 and had reduced grain dust deaths and injuries by around 60 per cent, demonstrating that the regulatory instrument worked.
- The CSB's 2006 Combustible Dust Hazard Study had already recommended that OSHA issue a general industry standard.
- OSHA began a National Emphasis Programme in October 2007, including sugar facilities; it announced rulemaking for a combustible dust standard in April 2009, fourteen months after fourteen people died.
Check it yourself
This is the section the QR code exists for. The bow-tie in section 04 was built with an AI-assisted workflow and then verified element by element against the CSB report by a process safety engineer. Everything needed to repeat that, and to disagree with it, is here: the method, the run data, the traceability, the figures that do not agree, and the limits.
Four moves — and an engineer at both ends
The first human in the loop
One document was admitted as evidence: CSB Investigation Report 2008-05-I-GA. Deciding what counts as an authoritative source is engineering judgement, and it is made before the model is opened. The analysis cannot be better than the evidence it is given, and a person chooses that.
The model may only look inside the source
Retrieval-augmented generation: the question pulls the relevant passages out of the report, and the answer is built from those passages, with the page it used attached. Every statement becomes traceable, so invention has nowhere to hide — which is not the same as saying it does not happen.
Method in the prompt, not in the weights
The prompt carries the bow-tie discipline: threats, top event, prevention and mitigation layers, barrier status, consequences. “Teach” here means instruct, not train. No weights were changed, no fine-tuning, no proprietary data. The engineering knowledge stays readable, editable and reusable on the next case.
The weak link is now the safeguard
For decades the human was the error to be designed out. Here the engineer is the last barrier: every element checked against the report, every citation opened, then signed off. A barrier is only as good as its competence — which is why retrieval, prompting and model limits are now process safety skills.
Run data
Pre-registered predictions
P1–P6 were written and dated before the second run, P7–P9 before the third, following the protocol used for the three Buncefield runs. Predictions are not corrected afterwards. Four of nine were confirmed; the five that were wrong are published as wrong — that is what makes the rest of it worth anything.
| # | Prediction | Outcome |
|---|---|---|
| P1 | The model lists NFPA requirements and CSB recommendations as barriers, rather than measures that were actually in place. | CONFIRMED |
| P2 | The model does not recognise that the 2007 food safety enclosure created the explosible atmosphere; it classifies it as a containment barrier or omits it. | falsified |
| P3 | The model hardens the CSB’s “most likely” on the bearing ignition into an established fact. | falsified |
| P4 | The barrier count does not converge with the Buncefield runs; the number remains a function of the prompt and the corpus, not of the incident. | CONFIRMED |
| P5 | The model fabricates at least one verbatim citation to a passage that is not in the report. | falsified |
| P6 | The model places “the explosion” as the top event rather than the loss of control that precedes it, collapsing primary and secondary explosions into one. | falsified |
| P7 | The emergency notification and alarm barrier reappears in the third run with a status that is not constant across the three. | CONFIRMED |
| P8 | The total barrier count falls outside the 14–17 range observed in the first two runs. | CONFIRMED |
| P9 | At least one barrier appears that is absent from both earlier runs. | falsified |
Traceability
Each element of the barrier analysis in section 04, against the passage of the CSB report that supports it. Where a status rests on the engineer's reading rather than on an explicit statement in the report, it says so.
| Element | As stated here | Basis | Source |
|---|---|---|---|
| Top event | Primary dust explosion inside the enclosed steel belt conveyor | CSB states the first dust explosion initiated in the enclosed steel belt conveyor below the silos. | report |
| Ignition | Overheated bearing, most likely | CSB: an overheated bearing in the steel belt conveyor most likely ignited a primary dust explosion. | report |
| Enclosure volume | approx. 850 ft³ inside a tunnel volume stated as approx. 8,700 ft³ | CSB states the ratio as one tenth and links it to accumulation above the MEC. | report |
| Explosibility | K_St 139 bar·m/s, class ST2; MEC 95 g/m³ powdered, 115 g/m³ granulated, 45 g/m³ per safety data sheet | Table 4 of the CSB investigation gives 95 and 115; the company safety data sheet quoted in the report gives 0.045 g/l. Granulated sugar is what spilled into the enclosure. | report |
| Dust cloud ignition temperature | 370 °C | Imperial Sugar material safety data sheets, quoted in the CSB report. | report |
| Housekeeping status | FAILED | CSB cause 2: inadequate housekeeping resulted in significant accumulations. Internal inspection six weeks prior recorded many tons of spilled sugar. | report |
| Management of change status | ABSENT | The report describes the purpose and the absence of extraction and venting; it does not use the term management of change. Classifying the absence as a barrier is the engineer's reading. | engineer |
| Compartmentation status | UNDERSIZED, not FAILED | The report describes the construction and the propagation. The distinction between a barrier that failed and one that performed to an inadequate design basis is the engineer's. | engineer |
| Emergency response status | FAILED | CSB cause 7: emergency evacuation plans were inadequate. Training, drills, intercom, lighting and egress findings are all stated in the report. | report |
| Regulatory barrier status | ABSENT | The report sets out the absence of a general industry combustible dust standard and the 2006 CSB recommendation to OSHA. Treating it as a barrier layer in the bow-tie is the engineer's framing. | engineer |
Figures that do not agree
Published sources on this incident do not fully reconcile. They are reported as found rather than averaged or quietly harmonised, because the discrepancies are themselves instructive — and because a reader who checks deserves to find the same thing.
| Figure | The discrepancy | How it is handled here |
|---|---|---|
| Injured | 36 in the title of the CSB report; 38 on the CSB case page for the same investigation. | This sheet uses 36 with the report as source, and states the difference. |
| Time of the explosion | Approximately 19:15 in the CSB report; 19:00 in several secondary accounts. | The report is used. |
| Tunnel volume | Stated as approximately 8,700 ft³, while the stated dimensions of about 130 × 12 × 7.5 ft give roughly 11,700 ft³. | Both are quoted. The stated volume is most likely net of equipment; the ratio the CSB draws — one tenth — is what the argument rests on. |
| Report number | The WOS 2026 abstract for this paper cites “2008-05-1-GA” with the digit one. | The correct number is 2008-05-I-GA, with the letter I for Investigation. Corrected here; the deposited abstract is left as submitted. |
| Minimum explosible concentration | Table 4 of the report gives 95 g/m3 for powdered sugar and 115 g/m3 for granulated sugar as received; the company safety data sheet quoted in the report gives 0.045 g/l, that is 45 g/m3. | All three are given with their provenance. Granulated sugar is what spilled into the enclosure, so 115 g/m3 is the figure relevant to that scene; the gauge in section 03 is drawn at the lower test value. |
| Date of the earlier dust collector explosion | The report dates it as less than two weeks before the February incident in one section and as 10 days before in another. | Both are quoted. They do not contradict each other — ten days is less than two weeks — and the pair is a fair test of any reader who checks by sampling a single passage. |
Limits of the method
- The source report already contains the analysis. A model working from it may reproduce the investigators' conclusions rather than derive them, and the output will look identical either way.
- Investigation reports carry hindsight bias by construction, and it propagates into anything generated from them. A live plant does not present its evidence pre-sorted into causes.
- The workflow is untested on sparse, contradictory or commercially sensitive evidence — which is the normal condition of consulting work.
- Three runs on one case give counts, not statistics. Across the three runs the barrier count was 14, 17 and 13, four barriers received more than one status, and one moved from the prevention to the mitigation side. Reported as counts, never as percentages. Part of the spread is granularity — run 3 merged measures that run 2 kept separate — and saying so is part of reporting it honestly.
- The mapping of the CSB's stated causes onto NFPA 654 clauses in this sheet is the engineer's, not the model's.
- Nothing here was validated by anyone outside rhysafe. Disagreement with the barrier statuses in section 04 is a legitimate and useful outcome of reading this page.
AI tools used
- NotebookLM (Google) — Retrieval-augmented, document-grounded interrogation of the primary source.
- Claude (Anthropic) — Structured analytical generation applying a bow-tie methodology framework.
No fine-tuning, no proprietary data and no site access were used at any stage.
Eight questions
Every answer is derivable from this sheet. Reveal the answer only after you have decided.
Take this to your own plant on Monday
Eight questions that can be answered with a document or a walk, not with an opinion. If a question cannot be answered by close of business, that is the finding.
Sources and further reading
Primary source
- U.S. Chemical Safety and Hazard Investigation Board, Investigation Report — Sugar Dust Explosion and Fire, Imperial Sugar Company, Port Wentworth, Georgia, February 7, 2008, Report No. 2008-05-I-GA, September 2009. Full report (PDF)
- U.S. Chemical Safety and Hazard Investigation Board, investigation case page, including the eleven recommendations 2008-05-I-GA-1 to -11 and their status. csb.gov
Context
- U.S. Chemical Safety and Hazard Investigation Board, Combustible Dust Hazard Study, Report No. 2006-H-1, November 2006 — the study that recommended OSHA issue a general industry combustible dust standard.
- NFPA 654, Standard for the Prevention of Fire and Dust Explosions from the Manufacturing, Processing, and Handling of Combustible Particulate Solids.
- NFPA 61, Standard for the Prevention of Fires and Dust Explosions in Agricultural and Food Processing Facilities.
- NFPA 499, Recommended Practice for the Classification of Combustible Dusts and of Hazardous (Classified) Locations for Electrical Installations in Chemical Process Areas.
- NFPA 484 is not cited in this investigation report — it covers combustible metals. Noted because it is a common misattribution.
European framework
- Directive 1999/92/EC (ATEX workplace) — zoning of explosive atmospheres and the explosion protection document.
- Directive 2014/34/EU (ATEX equipment) — equipment and protective systems for use in potentially explosive atmospheres.
- EN 14491 (dust explosion venting protective systems) and EN 14460 (explosion resistant equipment) for the venting and containment measures absent in this case.
Video
- The CSB produced a safety video on this investigation. CSB material is a work of the U.S. federal government and in the public domain. CSB video room
About this sheet
Produced by rhysafe AG as training material, from the public investigation report. It is not a work of the U.S. Chemical Safety and Hazard Investigation Board, carries no endorsement from it, and any error in interpretation is rhysafe's. The barrier statuses, the causal levels and the classification of the 2007 enclosure as a management-of-change failure are rhysafe's analysis and are open to disagreement. The technical drawing in section 01 is an original reconstruction from the dimensions and descriptions in the report; no figure or photograph from the report is reproduced. Figures accessed September 2026.
Citing this sheet
Scotton, M. S. and Rizzi, M., Learning Sheet 02 — Sugar dust explosion and fire, Imperial Sugar Company, Port Wentworth, 2008. rhysafe AG, Basel, 2026. Prepared alongside “Integrating AI Tools into Process Safety Consulting Practice”, WOS 2026, Limassol, Paper 39.
rhysafe.ch · martina.scotton@rhysafe.ch rhysafe.ch↗