AI Analysis You Can Verify
For regulatory bodies reviewing AI-assisted submissions, and quality teams evaluating AI tools: we provide the documentation, lineage tracking, and audit trails needed to assess AI-powered analysis with confidence.
A regulator-defensible framework for AI-derived behavioral data
EMA/CVMP has indicated that AI-based pruritic behavior scoring may be appropriate in preclinical efficacy studies when used within a defined context of use and supported by a regulator-defensible validation approach.
OneKind’s validation architecture is designed around that standard, including component-level validation, temporal concordance analysis, and explicit treatment of uncertain outputs to support auditable, reproducible measurement.
This creates a framework for AI-derived behavioral data that is structured for regulatory review in controlled preclinical settings.
Documentation for Defensible AI
Built on the ORIGIN platform, our AI systems come with comprehensive documentation designed for regulatory scrutiny.[1] Every analysis includes these core components, with automated behavioral detection validated against published methodology.[2]
Model Documentation & Ground Truth
Complete documentation of model architecture, training methodology, and ground truth data. Training datasets are annotated by expert veterinary behaviorists under documented guidelines with measured inter-annotator agreement.
Version Control — Models & Datasets
Full versioning for models, configurations, training datasets, and inference datasets. Every model traces to the exact data it was trained on. Every analysis traces to the exact input it processed.
Lineage Tracking
Automated data-to-decision lineage. Every transformation, preprocessing step, and model inference is linked — from raw input to final endpoint output.
Reproducibility Evidence
Deterministic inference ensures identical inputs produce identical outputs. Any historical analysis can be replayed against its exact model state with bit-identical results.
Audit Trail & Change Control
Every action is logged with full attribution. Model updates follow formal change control with documented impact assessment and approval workflows.
Human Oversight & Review
Human review is integrated at defined stages. Reviewers can override AI outputs — all override decisions are recorded in the audit trail with full attribution and rationale.
Uncertainty Handling & Performance Monitoring
Low-confidence predictions and out-of-distribution inputs are flagged for human review. Deployed models are continuously monitored for performance degradation and data drift. Substandard video quality is detected automatically.
Transparency & Limitations
Published documentation of model scope, known limitations, supported species, and validated behavioral classes. No black-box claims — every capability boundary is stated.
We have participated in multiple rounds of regulatory feedback with the European Medicines Agency through a pharmaceutical sponsor, and our platform is built to meet both EMA and FDA-CVM standards.
Report outputs are structured to align with standard clinical study report formats, with additional evidence layers for model provenance, reproducibility, and audit documentation.
Auditability & Explainability in Practice
Concrete scenarios showing how the ORIGIN platform responds to the questions regulators ask most.
Auditability in Practice
Scenario: An FDA auditor asks, “How was this model’s output on Subject 47 calculated?”
Retrieve the exact input: The ORIGIN platform provides the raw video file, preprocessing parameters, and metadata from the original trial.
Trace the model version: See exactly which model weights, hyperparameters, and dependencies were used at that point in time.
Reconstruct the decision: Replay the exact computation path, showing transformations, confidence scores, and intermediate outputs.
Export the proof: Generate a cryptographically-signed audit report showing the complete lineage — ready for submission.
Explainability in Practice
Scenario: A CVMP reviewer asks, “What model version was active during the pruritus scoring window, and has it changed since?”
Identify the model state: The ORIGIN platform returns the exact model version, threshold configuration, and runtime environment active at study lock.
Confirm no changes occurred: The model registry shows no releases between study start and lock date — cryptographically verified.
Demonstrate reproducibility: Re-run a sample of inferences against the locked model state — output is bit-identical to the original.
Export the evidence package: Generate a signed report showing model lineage, change history, and reproduction results — ready for submission.
Regulatory Alignment
Aligned with SOC 2 Type II best practices. Our platform is designed with the following frameworks in mind.
Aligned with EMA AI reflection paper requirements, FDA-CVM expectations, and Health Canada VDD guidance.
Key Questions We Address
Is the platform aligned with VICH GL42?
Designed with VICH GL42 requirements in mind, our behavioral analysis pipeline supports statistical rigor, documentation, and reproducibility expectations for veterinary clinical studies.
Is the platform 21 CFR Part 11 compatible?
Built to support 21 CFR Part 11 requirements including audit trails, electronic signatures, version control, and access management for electronic records.
Where does our data reside?
Configurable deployment with Canadian sovereign infrastructure as the default. On-premises and private cloud options available for organizations with specific data residency requirements.
Can OneKind integrate with our existing EDC system?
Yes, supports export to standard EDC formats including CSV, SAS Transport, and CDISC-compliant structures for seamless integration with your existing workflows.
For Internal Quality Teams
We provide the information QA and compliance teams need to evaluate our platform and determine if it meets your documentation and validation requirements.
Technical Briefings
Detailed presentations on methodology and validation approach, with opportunities to ask specific questions about our technical implementation.
Documentation Samples
Examples of the documentation packages that accompany our analyses, so you can evaluate whether they meet your internal standards.
Qualification Q&A
Structured responses to your vendor qualification questionnaire, addressing compliance, validation, and documentation requirements.
Pilot Study Options
Opportunity to run a pilot study with your own data to validate that our approach meets your needs before full deployment.
Key Questions We Address
Questions About AI Documentation?
Whether you're reviewing a submission that includes our analysis or evaluating our tools for internal use, our team is ready to provide the documentation and technical detail you need.
References
- Plant JD. “Repeatability and reproducibility of numerical rating scales and visual analogue scales for canine pruritus severity scoring.” Veterinary Dermatology, 2007. View →
- Park I, Lee K, Bishayee K, Jeon HJ, Lee H, Lee U. “Machine-Learning Based Automatic and Real-time Detection of Mouse Scratching Behaviors.” Experimental Neurobiology, 2019. View →