Chemicals catalogues are regulated to the molecule. PartSentinel measures, per CAS number and per formulation, what models claim to know — and flags the cases where a hallucinated boiling point, a wrong hazard code or a partial composition leak could become a regulatory or safety incident.
GPT-4 invents a hazard code combination for one of your formulations — risk: a downstream user follows it and creates a real safety incident.
Perplexity recommends a substitute coating from a competitor based on REACH registration — without checking that the substitute is restricted in your customer's market.
Claude pulls an outdated MSDS section from an archived product page — a section that was withdrawn after a 2024 EINECS update.
When a procurement copilot answers 'is this substance on the SVHC list?', it must be right. We test models against the current ECHA SVHC list (refreshed monthly) and flag plausible-but-wrong answers. Per-substance, per-model, per-language.
Models routinely propose EINECS numbers that don't exist or that belong to a different substance. We diff every model-returned EINECS against the official ECHA registry and flag mismatches. Particularly critical for substances with similar CAS prefixes.
Boiling point, flash point, melting point, hazard codes (GHS, CLP), pictograms — all are routinely hallucinated. We test against your published MSDS with configurable tolerance bands for numeric values and exact match for hazard codes. Findings are categorised by severity and routed to regulatory affairs.
Even partial formulation fragments — a precursor, a stabiliser, a trace additive — can compromise a moat. PartSentinel probes for compositional fragments your competitors shouldn't know. Findings are logged confidentially and never enter our public benchmarks.
Native parsers for the standard formats. Custom mapping for proprietary R&D PIM exports.
SVHC list, CSR, IUCLID dossier extracts
Hazard codes, pictograms, signal words, multi-language descriptors
OCR + structured extraction (BASF, Dow, Solvay, Arkema templates)
Native parser, MIME features, datasheet attachments
Akeneo, Pimcore, Siemens Opcenter, Dassault BIOVIA
Mapping assistant for proprietary trade formats
Chemicals engagements are typically driven by regulatory affairs in conjunction with data governance — and signed off by general counsel.
Yes. We refresh against the ECHA registry monthly and store the version manifest with each audit. Models are tested against the SVHC list version that was published when the audit ran — so findings remain reproducible.
Yes — and that's a primary use case. Confidential compositions are treated as private ground-truth: they are used to detect partial leaks but never published, never shared, and never enter our public benchmarks. EU-resident, named-engineer access, full audit log.
Our OCR + structured extraction pipeline is calibrated on BASF, Dow, Solvay and Arkema templates. Custom layouts are processed with a 24-hour onboarding overhead. We tag every extracted property with its layout confidence.
Yes. We test pictogram recall (presence / absence per substance) and signal-word coherence. Multi-language descriptors are tested in the languages where you sell.
We pick 50 representative substances with you (across hazard classes and confidentiality tiers), audit them on 10 LLMs, and ship a counsel-ready report with the SVHC recall, the MSDS coherence and the AI Act evidence pack.