AI transparency disclosure

Last updated: June 22, 2026. Effective: July 22, 2026.

1. Overview

Strand Health Inc., doing business as Everfur ("Company," "we," "us," or "our"), is committed to transparency regarding our use of artificial intelligence ("AI") and machine learning ("ML") technologies. This AI Transparency Disclosure ("Disclosure") describes how AI and ML are used within the Everfur Health Intelligence application, the Everfur Clinical Intelligence platform, the Care Team Feed, the Everfur Marketplace, and all related services (collectively, the "Services"), including the types of data processed, the nature and limitations of AI generated outputs, our AI governance framework, and your rights and choices.

This Disclosure supplements, but does not replace, our Privacy Policy, Terms of Service, and Veterinary Disclaimer. In the event of a conflict between this Disclosure and the Terms of Service, the Terms of Service shall control.

2. AI Technologies Used

Our Services employ multiple AI and ML technologies, including but not limited to:

Natural Language Processing (NLP). Used to interpret free text symptom descriptions, clinical notes, user queries, and conversational inputs, and to generate human readable health assessments, educational recommendations, and evidence grounded responses. NLP models include large language models, encoder models, named entity recognition systems, relation extraction pipelines, and text classification systems.

Computer Vision. Used to analyze photographs of pet skin conditions, eye conditions, dental conditions, ear conditions, wound presentations, mass presentations, and other visual clinical presentations for screening level educational assessments. Computer vision models include convolutional neural networks, vision transformers, and multi class image classifiers.

Audio Analysis. Used to analyze recordings of coughs, respiratory sounds, heart sounds, and other vocalizations to provide preliminary educational assessments of potential respiratory and cardiopulmonary patterns.

Knowledge Graph Reasoning. Used to traverse our proprietary veterinary knowledge graph, which encodes relationships between diseases, symptoms, medications, breeds, diagnostic tests, laboratory values, and other clinical entities derived from peer reviewed veterinary literature and structured clinical databases. Knowledge graph reasoning enables multi hop inference, evidence chaining, and structured clinical logic.

Drug Interaction and Dosing Models. Used to calculate species specific and weight adjusted medication dosing references and to identify potential multi drug interactions, contraindications, pharmacogenomic alerts, and dosing adjustments for educational purposes.

Differential Diagnosis Ranking. Used to generate ranked lists of potential conditions based on presented symptoms, signalment (species, breed, age, sex), clinical context, and prior history, for educational reference.

Metabolomics Analysis Pipelines. Used to process and interpret fur metabolomics data, including biomarker concentration analysis, multivariate pattern recognition, risk scoring, reference population comparison, and health profile generation.

Product Recommendation Systems. Used to suggest products from the Everfur Marketplace based on pet profile information (species, breed, age, weight, known conditions), health history, breed specific nutritional and wellness needs, and published veterinary guidelines. Product recommendations are educational and informational and do not constitute veterinary medical advice, prescriptions, or professional endorsements.

Content Moderation. Used to detect and flag potentially harmful, misleading, policy violating, or spam content in the Care Team Feed, product reviews, and other user generated content. Content moderation models include text classifiers, image safety classifiers, and rule based filtering systems.

Walk and Activity Pattern Analysis. Used to analyze walk tracking data to generate activity summaries, trend visualizations, and informational wellness insights, including distance trends, frequency patterns, pace changes, and comparative activity metrics.

Personalization and Recommendation. Used to personalize content delivery in the Care Team Feed, prioritize relevant health alerts, and tailor the user experience based on pet profiles and usage patterns.

Search and Retrieval. Used to power search functionality across clinical knowledge, product catalogs, and help content using semantic search, embedding based retrieval, and relevance ranking.

3. Training Data Sources

Our AI models are trained on the following categories of data:

Peer Reviewed Veterinary Literature. Over 50,000 publications from 47 veterinary journals, accessed under valid content licenses or through text and data mining (TDM) agreements. Source publications include journals published by Wiley, Elsevier, and other veterinary publishers.

Structured Clinical Reference Data. Including drug formularies, pharmacokinetic databases, pharmacodynamic databases, breed specific disease prevalence data, clinical practice guidelines, laboratory reference ranges, and treatment protocols from recognized veterinary authorities.

Proprietary Knowledge Graph. A structured database encoding clinical relationships, disease ontologies, drug properties, symptom entity relationships, diagnostic test associations, treatment linkages, and evidence hierarchies, curated by our scientific and veterinary advisory team.

Product and Nutritional Data. Publicly available product information, AAFCO nutritional profiles, ingredient databases, nutritional reference data, and product safety databases used to inform product recommendation models.

De Identified User Data (Opt In Only). Data from users who have provided explicit, informed, opt in consent through a clearly presented mechanism within the Services. De identification procedures are applied before any such data enters a training pipeline.

We do not use user submitted content (including pet photographs, clinical notes, symptom descriptions, chat interactions, Care Team Feed posts, product reviews, walk tracking data, or audio recordings) to train our AI models unless the user has provided explicit, informed, opt in consent as described in Section 7 and in our Privacy Policy.

4. How AI Outputs Are Generated

4.1 Input Processing

When you use an AI powered feature, your input (text, image, audio, or structured data) is transmitted to our secure servers via encrypted connection (TLS 1.2 or higher). The input is preprocessed (including tokenization, normalization, image resizing, and audio feature extraction as applicable) and then processed by one or more AI models to generate the relevant output.

4.2 Output Generation

AI outputs are generated through computational analysis of your input against our training data, knowledge graph, and model parameters. Outputs may include educational health assessments, differential diagnosis references, drug dosing references, interaction alerts, product recommendations, walk activity insights, content moderation decisions, and other informational content. Multiple models may be invoked in sequence (a "pipeline") to produce a single output. For example, a health chat response may involve named entity recognition, knowledge graph traversal, evidence retrieval, and natural language generation.

4.3 Confidence Scores

Where applicable, AI outputs include confidence scores, probability rankings, or certainty indicators. These scores represent statistical likelihood based on the model's training data distributions and the specific inputs provided. Confidence scores are NOT diagnostic certainty ratings, clinical probabilities, or evidence strength indicators. A high confidence score does not guarantee that a particular condition is present, and a low confidence score does not guarantee that a condition is absent. Confidence scores should be interpreted as relative, not absolute.

4.4 Product Recommendations

When the Services generate product recommendations, those recommendations are based on: (a) information in your pet's health profile (species, breed, age, weight, known conditions); (b) published veterinary nutritional and wellness guidelines; (c) product ingredient and formulation data; and (d) general breed specific health considerations. Product recommendations are algorithmically generated and are not individually reviewed by a veterinarian before presentation. They are educational suggestions and do not account for your pet's full clinical picture, specific dietary sensitivities, individual allergies, current medication regimen, or specific treatment plan. Always consult your veterinarian before making dietary or supplement changes.

4.5 Citation Generation

AI generated responses may include citations to published veterinary literature. Citations are generated by retrieval augmented generation (RAG) pipelines that search our licensed content databases for relevant source material. Citations may contain errors in attribution, reference details, or interpretation. The accuracy of citations depends on the quality of the retrieval and the model's ability to correctly associate claims with sources. Users should verify citations against the original publication.

5. Limitations and Risks

You should be aware of the following limitations of our AI systems:

Accuracy. AI models are not infallible and will produce incorrect outputs. Error rates vary by task, input quality, condition rarity, and model maturity. We do not guarantee any specific accuracy rate, sensitivity, specificity, positive predictive value, or negative predictive value.

Hallucination. Language models may generate plausible sounding but factually incorrect information ("hallucinations"), including fabricated citations, incorrect drug dosages, invented disease associations, or other false statements presented with apparent confidence.

Bias. AI models may reflect biases present in the training data, including over representation of certain breeds, species, conditions, geographic populations, or clinical presentations and under representation of rare conditions, uncommon breeds, or non Western veterinary practices.

Currency. AI models reflect veterinary knowledge at the time of their last training data update and may not reflect the most current research, drug approvals, safety communications, recalls, practice guideline revisions, or evolving clinical consensus.

Context Limitations. AI models process only the information explicitly provided in the current session and cannot perform physical examinations, palpation, auscultation, or any hands on assessment. They cannot order tests, access records not provided, or observe the animal directly.

Edge Cases. AI models may perform less reliably on unusual, atypical, complex, multi systemic, or rare cases that are poorly represented in training data. Out of distribution inputs may produce unreliable outputs.

Product Recommendation Limitations. Product recommendation models are trained on general nutritional and wellness data and cannot account for individual animal sensitivities, allergies, drug interactions, specific medical conditions, individual taste preferences, or product quality variations between batches.

Walk Data Limitations. Walk tracking AI insights are based on GPS data, which is inherently imprecise and affected by environmental factors. Activity pattern analysis is informational and not a substitute for veterinary assessment of exercise appropriateness.

Image Quality Dependency. Computer vision model performance is significantly affected by image quality, including lighting, focus, resolution, angle, distance, presence of fur covering the area of interest, and whether the relevant pathology is visible in the image.

Audio Quality Dependency. Audio analysis model performance is affected by recording quality, ambient noise level, distance from the animal, device microphone characteristics, and whether the target sound (cough, wheeze, murmur) is captured in the recording.

6. Human Oversight

For consumer users (Everfur Health Intelligence). AI outputs are informational and educational and should be shared with your licensed veterinarian for professional interpretation and clinical correlation. AI outputs should never be used as the sole basis for medical decisions.

For professional users (Everfur Clinical Intelligence). AI outputs are clinical decision support tools designed to assist, not replace, your independent professional clinical judgment. All outputs should be reviewed, validated, and contextualized by a licensed veterinary professional before being applied to patient care. Professional users are responsible for exercising their own clinical judgment and are not absolved of professional liability by reliance on AI outputs.

For Marketplace product recommendations. Always verify product suitability with your veterinarian, particularly for animals with known health conditions, dietary restrictions, allergies, or medication regimens.

7. User Data and Model Training

7.1 Default (No Training Without Consent)

By default, we do not use your individually identifiable clinical data, photographs, audio recordings, chat transcripts, Care Team Feed posts, product reviews, walk tracking data, or any other user generated content to train, fine tune, validate, benchmark, or otherwise improve our AI models. When you use AI features, your inputs are processed to generate the requested output (inference) and are not incorporated into model training datasets.

7.2 Opt In Training Consent

You may voluntarily opt in to allow us to use your de identified data for model training and improvement. Opt in consent is: (i) separate from acceptance of our Terms of Service or Privacy Policy; (ii) informed, with a clear explanation of what data will be used and how; (iii) granular, allowing you to consent to specific data types (images, text, audio) independently; (iv) revocable at any time through your account settings, though revocation does not require us to retrain or delete models that have already been trained on previously consented data prior to revocation; and (v) not a condition of using any feature of the Services.

7.3 De Identification

Before any user data is used for model training (with consent), we apply de identification procedures designed to remove or obscure all personal identifiers, including pet names, owner names, email addresses, phone numbers, specific dates, addresses, clinic names, veterinarian names, and any other information that could reasonably be used to identify an individual or household.

8. Automated Decision Making

Our Services do not make autonomous decisions that produce legal effects or similarly significant effects concerning any individual. All AI outputs are advisory, informational, and educational in nature and require human review before any action is taken. We do not use AI to make decisions about account eligibility, subscription pricing, service access, e-commerce availability, content visibility (beyond automated content moderation for policy compliance), insurance eligibility, employment, credit, housing, or any other determination that would affect your legal rights, financial standing, or significant life opportunities.

To the extent that any applicable law (including the CCPA/CPRA, GDPR, EU AI Act, Colorado AI Act, or other applicable legislation) grants you rights with respect to automated decision making or AI profiling, you may exercise those rights by contacting us at privacy@everfur.com.

9. AI Governance and Ethics

9.1 Governance Structure

We maintain an internal AI governance framework that includes designated personnel responsible for AI safety, model evaluation, and ethical review. Our veterinary advisory team reviews model outputs and performance metrics.

9.2 Ethics Principles

Our AI development is guided by the following principles: (a) transparency about AI capabilities and limitations; (b) safety as a primary design constraint, particularly regarding clinical content; (c) fairness and attention to bias across breeds, species, and populations; (d) user control over data and AI interactions; (e) accountability for model performance and failures; and (f) continuous improvement based on monitoring and feedback.

9.3 Bias Monitoring

We monitor our AI models for bias across key dimensions including breed, species, condition type, input quality, and geographic representation. When bias is identified, we prioritize remediation through training data rebalancing, model recalibration, or output adjustment.

10. AI Safety Measures

We implement the following safety measures in our AI systems:

Out of Distribution Detection. Mechanisms that flag inputs falling outside the model's reliable operating range, in which case the system may decline to provide an output, provide a low confidence warning, or recommend professional consultation.

Safety Gates. Rules based and model based systems that prevent the generation of outputs for high risk scenarios (such as emergency triage, controlled substance dosing beyond published ranges, euthanasia guidance, and irreversible treatment decisions) without explicit safety disclaimers and professional consultation recommendations. Safety gates also prevent the generation of content that could constitute the unauthorized practice of veterinary medicine.

Model Monitoring. Continuous monitoring for model drift, performance degradation, anomalous output patterns, increased error rates, and distributional shifts in input data.

Human Review Pipelines. Processes for qualified veterinary and data science professionals to review model performance, audit sample outputs, and flag concerning patterns.

Content Moderation. AI and human moderation systems for Care Team Feed content and product reviews to prevent the spread of harmful health misinformation, dangerous treatment recommendations, and policy violating content.

Feedback Mechanisms. In app tools for users to report inaccurate, harmful, or concerning AI outputs, which are reviewed by our team and used to prioritize model improvement, safety gate refinement, and output quality enhancement.

Incident Response. We maintain procedures for responding to identified AI safety incidents, including rapid model rollback capability, output correction mechanisms, and user notification processes when a significant error or safety issue is identified that may have affected user outputs.

Red Teaming and Adversarial Testing. We conduct periodic adversarial testing of our AI systems to identify failure modes, boundary conditions, and potential misuse vectors, and we incorporate findings into model and safety gate improvements.

11. Model Documentation and Versioning

11.1 Model Documentation

We maintain internal model documentation ("model cards") for each production AI model that includes: the model's intended use and scope, training data sources and composition, known limitations and failure modes, performance metrics on representative evaluation datasets, and identified bias considerations.

11.2 Versioning

AI models are versioned internally. Model updates may occur without user notification unless the update constitutes a material change in functionality or scope. Material changes will be communicated through this Disclosure or through in app notifications.

11.3 No Performance Guarantees

We do not guarantee any specific performance level, accuracy rate, sensitivity, specificity, F1 score, AUC, or other performance metric for any AI model. Performance may vary by input type, condition, breed, image quality, and other factors.

12. Third Party AI Components

Our AI systems may incorporate third party AI models, libraries, APIs, or services in addition to our proprietary models. Third party AI components are used in accordance with their respective terms of service and data processing agreements. We evaluate third party AI components for safety, accuracy, and privacy compliance before integration. Third party AI components are subject to the same safety gates and output monitoring as our proprietary models.

13. AI Fairness and Testing

13.1 Fairness Evaluation

We evaluate our AI models for fairness across key dimensions, including but not limited to: breed representation (ensuring that common and rare breeds are adequately served); species balance (ensuring that both canine and feline models receive appropriate development attention); condition type coverage (ensuring that both common and rare conditions are represented in training data); input quality robustness (ensuring that models perform reasonably across varying input quality levels); and geographic representation (ensuring that training data reflects diverse veterinary practice patterns where applicable). We do not guarantee equal performance across all dimensions, breeds, conditions, or input types, but we systematically identify and prioritize remediation of significant performance disparities.

13.2 Testing Methodology

Before deployment, AI models undergo a multi stage testing process that includes: unit testing of individual model components; integration testing of model pipelines; held out test set evaluation against predefined performance thresholds; clinical review of sample outputs by qualified veterinary professionals; adversarial testing to identify failure modes and boundary conditions; regression testing to ensure that model updates do not degrade performance on previously well served use cases; and staged rollout with monitoring before full production deployment. Post deployment testing includes continuous monitoring of key performance indicators, automated anomaly detection, periodic sample audits, and user feedback analysis.

13.3 Model Performance Transparency

We do not publicly disclose specific performance metrics (accuracy, sensitivity, specificity, F1 scores, AUC values) for our production AI models because: (a) performance varies significantly by condition, breed, input quality, and other factors, and single metrics can be misleading; (b) aggregate metrics do not represent the performance a specific user will experience; (c) detailed performance data constitutes proprietary trade secret information; and (d) published metrics could be used to calibrate adversarial attacks. However, we are committed to the general principle that our AI models should provide educationally valuable outputs for the conditions, breeds, and input types within their documented scope, and we continuously invest in improving model performance.

14. AI Supply Chain Transparency

Our AI technology stack includes both proprietary components developed in house and third party components sourced from established AI/ML providers and open source communities. We maintain a registry of AI components used in production, including their source, version, license terms, and data processing characteristics. We do not use AI components from sources that are subject to U.S. sanctions or export restrictions. We periodically review our AI supply chain for security vulnerabilities, licensing compliance, and alignment with our safety and ethics standards.

15. Data Processing and Privacy

For detailed information about how we collect, use, and protect your data in connection with AI processing, please refer to our Privacy Policy at everfur.com/privacy. Key privacy protections relevant to AI include: encryption of data in transit and at rest, access controls limiting who can view AI processing data, retention limits on inference data, opt in only model training, and de identification procedures for any data used in training.

16. Your Rights and Choices

You have the following rights and choices with respect to AI processing of your data:

Opt Out of Model Training. By default, your data is not used for model training. If you have opted in, you may opt out at any time through your account settings.

Access and Deletion. You may request access to or deletion of data processed by our AI systems by contacting us at privacy@everfur.com.

Explanation. You may request a general explanation of how our AI systems generate outputs, including the types of data used, the general methodology, and the known limitations. We may not be able to provide explanations of specific individual outputs due to the probabilistic nature of AI systems.

Product Recommendation Opt Out. You may disable AI generated product recommendations through your account settings without affecting your ability to browse and purchase products in the Marketplace.

Feedback and Correction. You may report inaccurate or concerning AI outputs through the in app feedback mechanism. We review all feedback and use it to improve our systems.

Objection. If applicable law grants you the right to object to AI processing or automated profiling, you may exercise that right by contacting us at privacy@everfur.com.

17. Regulatory Framework

As of the effective date of this Disclosure, the regulatory landscape for AI in consumer applications and veterinary decision support is evolving. We monitor applicable laws and regulations, including the EU AI Act, proposed U.S. federal AI legislation, state level AI governance laws (including the Colorado AI Act), and any veterinary specific AI regulatory guidance. We will update this Disclosure and our AI practices as applicable regulatory requirements evolve.

18. Updates to This Disclosure

We may update this Disclosure from time to time to reflect changes in our AI practices, model capabilities, regulatory requirements, or industry standards. Material changes will be communicated through the Services, via email, or through in app notification. The effective date at the top indicates when it was last updated.

19. Contact Information

Strand Health Inc.
d/b/a Everfur
1002 Dean Street, Suite 101, Brooklyn, NY 11238
Email: privacy@everfur.com