Last Updated on June 10, 2025
33.1 Model Families Used
33.1 Model Families Used
Chariot employs a multi-model architecture that may utilize various large language models (LLMs), computer vision models, and hybrid reasoning systems. Depending on your usage context, subscription tier, feature invoked, and real-time load conditions, the models used to generate AI output may include, but are not limited to:
33.1.1 Language Model Providers
OpenAI
GPT-4o
GPT-4 (legacy versions)
GPT-3.5
Anthropic
Claude 2.x
Claude 3 Opus, Sonnet, and Haiku families
Other Providers (as integrated in the future)
Meta (e.g., LLaMA)
Mistral
Google DeepMind (e.g., Gemini)
Open-source or private LLM deployments
33.1.2 Vision and Hybrid Model Layers
Proprietary vision classifiers
Fine-tuned visual language models (VLMs)
Optical character recognition (OCR) and document parsing engines
External APIs (e.g., VIN decoders, valuation APIs, PDF text extraction layers)
33.1.3 Model Selection Logic
The model used for any specific output is determined dynamically based on:
User’s subscription tier or report purchase type
Input type (e.g., photo, VIN, PDF, chat)
Model availability or load-balancing needs
Feature requirements (e.g., vision vs. text-only)
Users are not guaranteed access to any specific model unless explicitly marketed or displayed at time of use.
33.1.4 Model Output Disclaimers
Each model family has its own strengths, limitations, and behavioral characteristics. Accordingly:
Outputs may vary in format, tone, precision, or verbosity
Model hallucinations, factual errors, and contradictions may occur
No output should be interpreted as certified professional advice or final fact
By using Chariot, you acknowledge that:
Model assignment is handled internally and may change without notice
Chariot may modify, upgrade, or deprecate models at its sole discretion
Chariot disclaims liability for variance in response behavior across vendors or versions
33.2 No Output Uniformity Guaranteed
Answers provided by one AI model may differ from another—even for identical prompts. Chariot does not guarantee output consistency across models.
33.3 Model Selection May Vary by Plan
Chariot does not guarantee access to any specific model (e.g., GPT-4o, Claude 3 Opus) regardless of your subscription tier or report purchase. Model selection is determined dynamically based on internal cost thresholds, vendor availability, technical load, and usage caps.
While higher-tier users may receive access to newer or more capable models at times, this is not guaranteed. You may receive outputs from a range of models, including base-tier systems (e.g., GPT-3.5, Claude Haiku), depending on current system conditions and economic feasibility.
You acknowledge that:
Model access is not a promised benefit of any plan
Output quality may vary from session to session
Chariot makes ongoing adjustments to its AI routing logic to balance cost, speed, and availability
This policy protects Chariot’s ability to sustainably deliver AI features while adapting to backend model pricing, vendor terms, and compute constraints.
33.4 Output Style Differences
Some models are more verbose, some more direct. Users should expect stylistic, interpretive, and tone-based variation in responses.
33.5 Underlying Knowledge Cutoffs
Each model has a training data cutoff (e.g., April 2023 vs October 2023), which may affect accuracy or relevance of responses depending on prompt context.
33.6 Same Prompt ≠ Same Result
Even the same model may return different responses to the same prompt across sessions due to temperature, randomness, and contextual state.
33.7 No Output Determinism Promised
Responses are generated probabilistically. Chariot cannot force exact repetition of results and does not offer deterministic mode guarantees.
33.8 Prompt Phrasing Sensitivity
Minor rewording can shift AI behavior. Chariot encourages users to rephrase or retry queries for clarification, not rely on one-shot perfection.
33.9 Session State & Memory
Chat sessions are stateless unless otherwise noted. Chariot does not persist AI memory beyond each conversation unless premium memory tiers are offered.
33.10 Chariot Responses May Drift
Image uploads processed by chariot may return different object labels, tone, or observations over time—even for the same file.
33.11 Model Switching for Load Management
Chariot may dynamically switch between models for availability, latency, or pricing optimization without notice to the user.
33.12 Scheduled Model Upgrades
When models are deprecated or replaced (e.g., GPT-4o superseding GPT-4), users will be notified of any relevant changes impacting subscriptions or reports.
33.13 Vendor Substitution Right
Chariot reserves the right to replace OpenAI or Anthropic with another vendor if a superior, compliant model becomes available.
33.14 Multi-Vendor Stacking
Some queries (e.g., report generation) may be composed using multiple models across vendors for different segments of the output.
33.15 No Guaranteed Vendor Continuity
No model provider (OpenAI, Claude, etc.) is guaranteed to remain part of the stack. If vendor terms or APIs change, model access may shift.
33.16 Non-Human Output
All responses are machine-generated and may contain errors, hallucinations, or inconsistencies. Users should verify critical information independently.
33.17 Non-Expert Content
AI-generated advice is not equivalent to professional legal, medical, or financial counsel. Users are responsible for applying content appropriately.
33.18 Interpretation Risk
User misinterpretation of AI content does not constitute product defect or grounds for refund unless content violates usage policies.
33.19 Model Limitations in Nuance
AI may oversimplify, generalize, or overlook context-specific nuances, especially in subjective or document-based questions.
33.20 Output Confidence Ratings (optional)
Future versions of Chariot may include confidence scores, model tags, or disclaimers to help users interpret source model differences.
33.21 Model Drift Disclaimer
Chariot makes no guarantee that model behavior will remain constant over time. Vendor fine-tuning may shift results.
33.22 Prompt Engineering Results May Vary
Output quality can be affected by how the user phrases questions. User experimentation is expected.
33.23 No Refunds for Model Differences
If a user prefers a certain model’s response style or accuracy, this is not grounds for refund. Tiered access reflects model cost differences.
33.24 AI Interpretive Risk Warning
Reports, contract scans, and visual analysis are interpretations, not facts. Accuracy is probabilistic and should be treated as guidance.
33.25 No Legal Reliance
Differences between model responses do not constitute legal inconsistency. Users must rely on human professionals for legal decisions.
33.26 Model Attribution
Where possible, Chariot will annotate which model generated which section of a report or response (e.g., “Generated with GPT-4o”).
33.27 Prompt Metadata Storage
To improve consistency tracking and abuse prevention, Chariot may log metadata including model used, token count, and session ID.
33.28 Cross-Session Output Comparison
Chariot may offer premium features that allow comparison of the same prompt across multiple models for advanced users.
33.29 Hallucination Reporting Tool
Users may flag suspect or inaccurate outputs. Internal logs will be reviewed for possible vendor feedback or user credit.
33.30 Survival of Clause
Chariot uses third-party SDKs (Software Development Kits) to power critical app functionality including authentication, payments, subscriptions, analytics, file uploads, and AI responses.These disclaimers apply permanently to all AI-generated responses regardless of version, plan, or delivery method.
Contact Us
If you have any questions or concerns about our Terms of Service or the handling of your personal information, please contact us at support@chariotreport.com