> For the complete documentation index, see [llms.txt](https://docs.havl.io/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.havl.io/products/havl-verify.md). # havl verify **havl verify** is the human verification layer for AI-generated content, predictions and decisions. It enables AI systems to request independent human judgment before an output is accepted, published, stored or used by another system. Instead of treating every model response as correct by default, havl adds human verification where confidence, accuracy or judgment matters. **AI generates. Humans verify.** svg ### Human verification for AI outputs AI models can generate enormous volumes of useful information, but their outputs are probabilistic. A model can produce a convincing answer that is incorrect. It can misunderstand context, extract the wrong information, misclassify content or provide an answer with greater confidence than the underlying information justifies. For many applications, this uncertainty is acceptable. For others, it is not. havl verify provides an independent human layer between the model and the system that relies on its output. The AI continues to handle the scale. Humans become involved selectively when additional confidence is required. svg ### What can be verified? havl verify can be applied to almost any AI-generated output that can be evaluated by a human. Typical verification tasks include: * classifications and labels * summaries and generated text * extracted information * AI-generated answers * image and visual interpretations * structured data * model predictions * content moderation decisions * comparisons between model outputs * domain-specific results Clients define exactly what the verifier should evaluate and what a valid response should look like. A verifier might confirm whether an output is correct, choose between alternatives, correct a value, classify information, provide a structured response or escalate a result that cannot be reliably verified. ### Verification where it matters Not every AI output needs a human. havl verify allows verification to be applied selectively according to the confidence, risk and importance of an output. #### Sample A percentage of routine AI outputs can be sent for independent human verification. This creates continuous quality control without requiring humans to review every result. Sampling can help identify recurring errors, monitor model performance and detect changes in output quality over time. #### Escalate Human verification can be triggered when predefined conditions are met. For example, an output might be escalated when model confidence falls below a threshold, multiple models disagree, an unusual result is detected or a business rule identifies additional risk. Routine outputs continue automatically. Uncertain outputs go to humans. #### Require For sensitive or high-impact use cases, verification can become a mandatory checkpoint. The AI output cannot be accepted or passed to the next stage until the required human verification has been completed. This allows organizations to define specific areas where autonomous model output alone is not sufficient. svg ### Different levels of verification The required level of human involvement can vary with the importance of the output. A routine classification may only require one independent verifier. A more consequential result could require several humans to review the same information independently. Clients can define the required number of verifiers and the level of agreement necessary for a result to be considered verified. This enables different verification models, from simple confirmation to stronger forms of human consensus. **One verifier for routine checks.** **Multiple verifiers when greater confidence is required.** **Qualified experts when specialist knowledge matters.** The objective is not to apply maximum verification everywhere, but to apply the appropriate level of human judgment to each situation. ### Human consensus Some questions do not have a result that should depend on a single person's judgment. havl verify can distribute the same task to multiple independent verifiers and combine their responses according to predefined consensus requirements. For example, a task could require agreement from two out of three verifiers before the result is accepted. If consensus cannot be reached, the task can be escalated for additional review. This creates another layer of resilience against individual mistakes and subjective judgments. ### Expert verification Some AI outputs require more than general human judgment. Medical information, legal documents, financial analysis, engineering data or other specialized domains may require reviewers with relevant expertise. These tasks can be routed to **Certified Experts** where available. Certified Experts represent an additional verification layer for situations where professional or specialist knowledge is necessary to meaningfully evaluate an AI-generated result. This allows the level of human expertise to scale with the complexity of the AI output. svg ### Independent human judgment The verifier does not need access to the client's entire application. Each verification task contains the information necessary to make the required decision: the AI output, relevant context, the verification question and the expected response format. This keeps the task focused. The verifier evaluates the information independently and returns a structured response that can be used by the requesting system. Depending on the task, the result may be: **Verified** — the AI output is accepted. **Corrected** — the verifier provides the appropriate result. **Rejected** — the output does not meet the defined requirements. **Escalated** — additional judgment or expertise is required. ### Humans for uncertainty The goal of havl verify is not to put a human behind every AI response. That would remove one of the fundamental advantages of AI: scale. Instead, havl focuses human attention on the outputs where it creates the most value. AI can handle millions of routine operations. Humans can review samples, resolve uncertainty, verify important results and intervene when models reach their limits. As models improve, the number of outputs requiring verification may decrease. But for the outputs that matter most, independent human judgment remains valuable. havl makes that judgment available as a layer. **AI generates. Humans verify.** --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation by asking a question. Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters: ``` GET https://docs.havl.io/products/havl-verify.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.