Hazmat Regulatory Classification

Hazmat regulatory classification turns structured product and Safety Data Sheet (SDS) data into a draft classification against the three main transport regimes: ADR (road), IATA Dangerous Goods Regulations (air), and IMDG (sea). The draft includes proposed UN numbers, proper shipping names, packing groups or equivalent risk bands where applicable, and the marking, labeling, and documentation instructions needed for the pack stage.

The model does not replace a certified dangerous goods (DG) specialist. It accelerates document assembly and cross-checks between SDS fields and modal rules so the specialist spends time on judgment, edge cases, and signature, not on first-pass lookup.

Typical inputs are SDS sections (especially identification, hazard identification, composition, physical properties, and transport information), SKU master data, pack configuration, and destination mode. Typical outputs are a classification worksheet, packing instruction hints, mark/label lists, and a documentation checklist for shipper’s declaration or equivalent modal paperwork.

Vendor ecosystems that already own master data or label workflows often sit alongside this pattern. Labelmaster is common for label content and regulatory reference. SAP EHS and Sphera often hold SDS, substance, and product compliance records that feed classification. The LLM layer sits on top of those systems as a drafting and consistency engine, not as a second master of record.

Classification quality tracks SDS completeness and internal consistency. The workflow should map SDS transport-relevant fields to modal rule checks: UN number candidates, proper shipping name language, class/division, packing group, limited/excepted quantity eligibility, and special provisions that change packing or documentation.

When the SDS already states transport information, the model should treat that as a strong prior and verify it against the current edition of the relevant code rather than inventing a parallel answer. When transport information is blank or outdated relative to hazard and composition sections, the model should surface the conflict and propose candidate classifications with explicit citations to the SDS fields it used.

Mode matters. A substance that moves as limited quantity by road may fail IATA packing or quantity limits. Multimodal shipments need one coherent story across ADR, IATA, and IMDG, with differences called out per leg rather than a single generic “hazmat” label. Pack-stage users need that modal split early enough to choose packaging, marks, and carrier acceptance paths before cartons close.

Empty classification is mandatory when the SDS is missing, unreadable, or contradictory in ways that would change UN number, class, packing group, or shipping name. Contradiction examples include transport section UN numbers that conflict with hazard class statements, composition that implies a different primary hazard than the stated class, or physical state/temperature controls that the transport section ignores. In those cases the system returns no signed-ready classification: status empty or equivalent, a structured list of blocking gaps, and a hold on marking, labeling, and documentation instructions until a human resolves the SDS or supplies authoritative source data.

Once a provisional classification exists, the pack stage needs executable instructions: which marks and labels apply, where they go on the package, which packing instruction or packing provision governs, and which documents must travel with the shipment.

The draft should separate package-level requirements from shipment-level paperwork. Package-level items include hazard labels, handling labels, UN marks, orientation arrows where required, and limited quantity marks. Shipment-level items include shipper’s declaration content, additional handling information, and references that customs or carrier acceptance teams will expect. Related pack controls such as Vision-Based Label and Contents Verification can later confirm that applied labels match the approved classification draft.

Documentation drafts should stay tied to the classification record ID and SDS revision. If either changes, marks and docs must regenerate. Autonomous document generation patterns used elsewhere in logistics, such as Autonomous Customs Documentation, are complementary when the same shipment also needs trade paperwork, but DG declarations remain under DG specialist control.

Damage and mishandling risk is adjacent, not a substitute for classification. Shipment Damage Risk Scoring may flag fragile or high-severity packs, yet it does not decide UN numbers or packing groups. Keep those decision paths separate so a risk score never silently overrides a regulatory class.

Ownership usually spans product compliance, packaging engineering, warehouse pack leads, and a certified DG specialist (or equivalent under local competent authority rules). The specialist remains accountable for the classification that ships.

Recommended control model:

1. Upstream systems supply SDS and product identity (often SAP EHS or Sphera).
2. The LLM produces a draft classification and instruction set with field-level provenance.
3. Rule checks enforce empty output on missing or contradictory SDS and on unsupported modes.
4. The DG specialist reviews, edits, and electronically signs.
5. Downstream label printers and WMS consume only signed packages of instructions (Labelmaster or plant labeling systems are typical consumers).

Audit trails should retain SDS revision, model prompt/version metadata if used, draft versus signed deltas, and the specialist identity and timestamp. Carriers and regulators will ask for the human-approved record, not the model transcript.

Training data and prompts must not invent UN numbers or packing groups. Prefer retrieval against controlled regulatory tables and company-approved substance masters. Where the model cannot ground a field, leave it blank and escalate rather than guessing.

The highest-severity failure is a confident wrong class that reaches the pack line. Mitigation is empty-or-hold behavior on weak input, mandatory specialist signature, and hard separation between draft and production label release.

Common operational failure modes:

• Stale SDS: product reformulation already live in manufacturing while transport section still reflects the old formula.
• Multi-component kits: outer pack treated as one substance when inner receptacles drive different classes or segregation rules.
• Mode switch late in the order: road classification reused for air without re-check.
• Over-reliance on SDS Section 14 when Section 2/3 contradict it.
• Vendor catalog mismatch: label SKUs printed from an unsigned draft.

Limits to state clearly to operators: the system drafts; it does not certify. It does not replace carrier acceptance checks. It does not authorize shipments when SDS data is absent or internally inconsistent. Quantity limits, segregation, and special provisions still require specialist confirmation against the current modal manuals and any company exceptions.

Use this sequence when piloting:

1. Inventory SDS coverage for the SKUs that actually ship as potential DG; measure missing and contradictory rates before automation.
2. Define empty-classification rules in writing (missing SDS, unreadable SDS, UN/class/name conflicts, unsupported mode).
3. Connect product compliance masters (SAP EHS, Sphera, or equivalent) as the SDS source of truth.
4. Map signed outputs to label and doc consumers (Labelmaster, WMS, print stations).
5. Require specialist sign-off on every production classification; block unsigned drafts from print and ship.
6. Sample audits: compare signed outcomes to specialist-only baselines on a fixed SKU panel each cycle.
7. Extend only after empty rates and correction rates are acceptable for your risk tolerance.

Success looks like shorter time from SDS-ready product to signed pack instructions, fewer last-minute holds caused by incomplete worksheets, and a clean audit path from SDS revision to applied marks. Classification authority stays with the certified DG specialist; the model’s job is faster, better-sourced drafts and disciplined silence when the data cannot support a class.