Electrotechnical catalogues live or die on datasheet truth. PartSentinel probes how every major LLM represents your ETIM classes, your numerical specs and your supersession chains — and flags the spec values models are inventing with high confidence.
GPT-4 hallucinates a torque value with 3% precision — picked from a similar but unrelated reference in a competitor's datasheet.
Claude infers an obsolescence date from a forum thread and recommends a replacement that's actually a tier below your spec.
Gemini cites an ETIM class that hasn't applied to your part since 2023 — from an archived datasheet still indexed.
Models confidently invent torque, current rating, IP class and dielectric strength values. We test every numeric spec against your published datasheet with a ±2% tolerance and flag the cases where models are simply making numbers up. Pinout coherence is checked separately for ICs and connectors.
When a model classifies your contactor as a relay, your datasheet leaks into the wrong information channel. PartSentinel tests ETIM class assignments per model and per language — and flags reclassification drift quarterly. We support ETIM 9.0 with the EC-class extensions.
Multi-step supersession chains (A → B → C → D) confuse models the most. They tend to anchor on the oldest reference and miss the newest replacement. We test the full chain per family and per manufacturer; you get the structured response template to push to upstream model providers.
When a procurement copilot answers 'is this part RoHS-compliant?', it must be right. We test models on REACH SVHC list inclusion, RoHS compliance status, and dual-use export-control classification — flagging cases where the answer is plausible but wrong.
Native parsers for the standard formats. Custom mapping assistant for proprietary PIM exports.
Class IDs, feature values, units, multi-language descriptors
Native parser, full classification + characteristics
MIME features, attachments, datasheet PDFs
OCR + structured extraction (Schaeffler, Phoenix Contact, Siemens, ABB)
Akeneo, Pimcore, Contentserv, Stibo, SAP MDG
Direct delta loads via signed S3 / GCS bucket
Electrotechnical PartSentinel engagements are bought by the people who own datasheet truth — not the people who run brand campaigns.
Per-spec, per-reference, per-model. Each numeric value (torque, current, IP class, dielectric strength) is compared to your published datasheet with a configurable tolerance band — default ±2%. We score 'within tolerance', 'outside but plausible' and 'invented' separately so you can read the failure mode.
Pinout coherence is checked for ICs, connectors and terminal blocks. Schematic-level coherence (signal routing) is on the roadmap for Q4 2026.
Yes. We test model recall against your published regulatory status — both list inclusion (REACH SVHC) and binary status (RoHS-compliant yes/no). Dual-use export-control classification is supported but requires a signed information-sharing agreement.
Yes. We have an OCR + structured extraction pipeline calibrated on Schaeffler, Phoenix Contact, Siemens and ABB datasheet templates. Custom layouts are processed with a 24-hour onboarding overhead.
We pick 50 representative references with you (across families and ETIM classes), audit them on 10 LLMs, and ship a counsel-ready report with the spec hallucinations, the supersession drift and the AI Act evidence pack.