ADAS models validated faster and documented completely.
The Challenge
General Motors develops ADAS perception models across more than a dozen vehicle platforms, each with regional variants tuned for different road conditions, regulatory environments, and sensor configurations spanning Chevrolet, GMC, Buick, and Cadillac lines in 30+ markets. ISO 26262 functional safety requirements and UNECE WP.29 type approval regulations both require detailed documentation of the AI development lifecycle: training data provenance, validation datasets, performance benchmarks, and post-market monitoring plans.
The safety engineering team was maintaining this documentation manually, with a dedicated team of technical writers translating pipeline artifacts into regulatory-format documents after each development cycle. The process introduced a lag of six to eight weeks between model validation completion and regulatory submission readiness, and version mismatches between engineering records and regulatory documents had caused two submission rejections in the prior model year, a costly delay when vehicle programs are measured in months. As GM accelerated its Super Cruise and Ultra Cruise rollouts across additional markets, the manual documentation burden scaled with every new geographic variant.
The Solution
The platform was integrated into General Motors' ADAS model CI/CD pipeline to capture training data lineage, benchmark results, bias detection outputs, and validation metadata automatically as each model is built. The model card framework was configured to produce UNECE WP.29-aligned documentation directly from pipeline metadata, eliminating the manual translation step. Regulatory submissions now draw from the same system of record that engineers use to track model performance.
Field monitoring was connected to the over-the-air telemetry pipeline across the active Super Cruise and Ultra Cruise fleet, enabling continuous drift detection on perception model behavior at global scale. When edge-case detection rates drift outside validated performance bounds, the safety engineering team receives a structured alert with the affected geographic market, operating conditions, and edge case categories, giving them the specific information needed to trigger a targeted revalidation rather than a full development cycle.
Data Lineage Across Multi-Source Training Sets
General Motors' ADAS perception models are trained on data from internal proving grounds, third-party data vendors, synthetic simulation pipelines, and crowd-sourced edge case collections across 30+ markets. Regulatory documentation requires the organization to demonstrate that training data sources are known, quality-controlled, and traceable. The data catalog captures lineage from raw sensor data through preprocessing, augmentation, and dataset versioning, producing a complete provenance record that satisfies both internal safety review and external regulatory audit without manual reconstruction.
Performance Benchmarking Across Operating Domains
ISO 26262 validation requires demonstrating model performance across the full operational design domain, different lighting conditions, weather environments, road surface types, and geographic contexts. The experiment tracking module stores benchmark results against each condition category, enabling validators to run pass/fail checks against predefined performance thresholds rather than manually reviewing results tables. When a new model variant is promoted, the benchmark comparison is generated automatically against the previously approved baseline.
Post-Market Monitoring as a Regulatory Deliverable
UNECE WP.29 requires manufacturers to maintain a post-market monitoring plan and demonstrate its operation. The drift detection configuration: thresholds, monitoring frequency, escalation paths, and response procedures is exported as a structured monitoring plan document that maps directly onto the regulation's requirements. When monitoring detects drift in the field, the alert record and response action are automatically appended to the monitoring log, creating a living regulatory artifact rather than a document produced at audit time.
Results
Automated documentation generation eliminated the 6–8 week manual translation lag between engineering completion and regulatory readiness.
Single system of record ensures engineering artifacts and regulatory documents are always synchronized.
All Super Cruise and Ultra Cruise perception models across Chevrolet, GMC, Buick, and Cadillac lines governed from one platform.
Over-the-air telemetry feeds continuous performance surveillance across the active Super Cruise and Ultra Cruise fleet globally.
See what Thndr AI can do for your team
Talk to our team about your specific AI governance challenges.
