What counts as a wrong answer.
The full rules for what qualifies — with real examples, so there's no ambiguity about what's worth submitting.
01/ The Default
If we can verify it, we pay it.
We've designed Verifiable AI to default toward verifying submissions, not finding reasons to deny them.
If a reasonable reviewer can verify your reported issue against an authoritative source, we pay $100 by Visa Prepaid Gift Card. We don't engineer reasons to say no. We don't have a formal appeal process because we don't need one.
Three types of issues are in scope: factual errors, non-existent references, and reasoning errors. The rest of this page describes each type with real examples of what we accept and reject. If your submission falls cleanly into one of the in-scope categories and is verifiable, we'll pay it. If it's borderline, we still tend to favor the submitter — but we want you to know what borderline looks like before you spend time writing it up.
Program scope at a glance.
Submissions must be based on a single-prompt, text-only interaction using Pro or Deep mode. Standard mode outputs, multi-turn conversations, and prompts involving uploaded files (images, PDFs, spreadsheets, etc.) are not in scope. Open to MiroMind Pro subscribers in the United States (excluding FL, NY, RI, AZ, CO) and Singapore. Per-participant cap of $500/month and $3,000 over the program. Subject to the Official Rules.
02/ Type 1
Factual errors.
Definition
A statement of fact in MiroMind's response that can be disproven by an authoritative source. The statement must be presented as fact (not opinion, prediction, or hypothetical), and the source must be of a type listed in Section 06 — What counts as authoritative.
This is the broadest category and covers most things people intuitively think of as "wrong." Examples below clarify the boundary.
What we pay for
We don't pay for
A few specifics worth knowing
Partial accuracy counts. If MiroMind's answer is technically true but omits material exceptions or qualifications that change the practical meaning, that's a factual error. You need to specifically identify what was omitted and why it matters.
Outdated information counts — but only for facts that have been stably true for at least 6 months. Same-week news items don't qualify.
Domain-specific errors count. If you're a specialist and you spot something wrong that only specialists would catch, that's exactly what this program is for. You'll need to cite an authoritative source from your field.
03/ Type 2
Non-existent references.
Definition
A specific named source — paper, case, statute, dataset, person, or document — that MiroMind cited and that doesn't exist or doesn't say what's claimed. This is one of the highest-confidence error types: either the cited thing exists in the form described, or it doesn't.
This category exists because LLM-generated citations are unusually prone to fabrication, and the harm in professional contexts (law, medicine, science, finance) is high.
What we pay for
We don't pay for
How we verify
For each non-existent reference report, we check the cited source in at least two authoritative databases (e.g., for a legal case: Westlaw and the official court database; for a paper: the journal directly and Google Scholar). If neither finds it in the form cited, the issue is verified.
Tip for submitters
Tell us specifically what's wrong — "this paper doesn't exist in this journal" vs. "the author is wrong" vs. "the page numbers are wrong." Specificity speeds up verification.
04/ Type 3
Reasoning errors.
Definition
A multi-step argument or calculation in MiroMind's response with a clear logical or computational break that leads to a wrong final answer. The break must be specific and identifiable — not just "I disagree with the conclusion."
This is the hardest category to evaluate, because reasoning is partly a matter of method. We focus on cases where the chain of inference contains an identifiable error, not just a different valid path to a conclusion.
What we pay for
We don't pay for
Reasoning errors vs. prompt ambiguity
The single most common dispute in this category is: did MiroMind make a reasoning error, or did your prompt under-specify what you wanted? Our default is to look at the chain of inference itself. If the inference contains an internal break (a step that doesn't follow from the previous steps), it's a reasoning error regardless of the prompt. If the inference is internally consistent but based on assumptions you didn't specify, it's a prompt issue.
Tip for submitters
The clearer you can identify the broken step, the faster we can verify.
05/ Out of scope
What we don't pay for
For completeness, here's the full list of things that don't qualify under this program. None of these are about hiding behind rules — they're about preserving the meaning of "verified issue" in a way that can be checked consistently.
Subjective opinions or aesthetic preferences
"This recommendation is bad," "the writing style is awkward," "the analysis isn't deep enough." These are judgments, not verifiable facts.
Creative writing outputs
Poems, fiction, marketing copy, story drafts. Creative work isn't evaluated for factual accuracy in the same sense and isn't in scope here.
Predictions about future events
"Will the Fed raise rates in Q3?" — a prediction is not verifiable as wrong until much later, and even then attribution to model error vs. model uncertainty is not clean.
Errors caused by ambiguous or misleading prompts
If your prompt left material ambiguity and MiroMind picked one reasonable interpretation that turned out wrong, that's a prompt issue, not a model error. Clearer prompts get cleaner answers.
Errors from Standard mode
Only issues from Pro mode and Deep mode are eligible for verification. Standard mode uses a smaller model with intentionally higher error tolerance for everyday questions where speed matters more than rigor. Holding it to the same accuracy standard as Pro and Deep would misrepresent what each mode is built for.
Multi-turn conversations
We currently review only single-prompt errors — one prompt, one response. Errors that emerge over a back-and-forth conversation are too sensitive to context that came before, and we haven't optimized our system for that scenario yet. We may expand scope here later.
Errors involving uploaded files or attachments
When MiroMind processes uploaded files (PDFs, images, spreadsheets), the answer depends both on how the file is interpreted and on the model's reasoning over that interpretation. Reasonable interpretations can vary, and third-party tooling is involved in the pipeline — making it hard to attribute errors cleanly to MiroMind. Not in scope right now.
API-generated outputs
Only outputs from MiroMind's web or mobile app are in scope. Outputs generated through the API run in different contexts, with different settings, and don't have shareable conversation links — making consistent verification difficult. We may expand scope here later.
Code generation issues
Bugs in code, off-by-one errors, security vulnerabilities in generated snippets — none of these are in scope right now. MiroMind is positioned as a verifiable research and reasoning system, not a code generator.
Safety policy refusals
If MiroMind declines to answer because of safety policy, that isn't a wrong answer for purposes of this program.
Already-known errors
We pay only the first reporter of each issue. If your error was previously reported and is in our Known Issues list, we won't pay it again — but you're recognized as having spotted a real one.
Errors we can't reproduce
We try at least 10 reproductions across different conditions. If we can't reproduce the error, the claim is marked inconclusive and you have 14 days to provide additional context (a screen recording, related prompts) for re-review.
06/ Sources
What counts as authoritative.
For us to verify a factual error or non-existent reference, you'll typically need a source from one of the following categories. The stricter the field's evidence standard, the stricter the source we need.
Domain
Sources
Medicine
WHO, NIH, FDA, EMA, peer-reviewed journals (NEJM, Lancet, JAMA, BMJ, etc.), Cochrane reviews, official clinical practice guidelines.
Law
Official court databases, primary statutes and regulations (legislation.gov.uk, eur-lex, cornell.edu/cfr, etc.), Westlaw / LexisNexis case databases, official agency rulings.
Finance
Regulatory bodies (SEC, MAS, FCA, ESMA), exchange filings (EDGAR), IFRS Foundation, FASB, CFA Institute, central bank publications.
Science
Peer-reviewed journals with DOI, official scientific bodies (NASA, ESA, IPCC, NOAA), institutional preprint servers (arXiv, bioRxiv) for fast-moving fields, established textbooks.
Industry standards
ISO, IEEE, W3C, IETF (RFCs), national standards bodies (ANSI, BSI, JIS, GB), manufacturer official documentation for product specifications.
Geography & official data
National statistical offices, UN Statistics Division, World Bank, IMF, OECD, Eurostat. Government primary sources beat third-party aggregators.
Not authoritative
Wikipedia, blogs, forums, social media posts, news articles (without primary-source backing), AI-generated summaries. These may help find sources but don't themselves count as authoritative for verification.
If you don't have a "perfect" source
Submit anyway. Our reviewers will independently verify, and we often find authoritative sources you didn't have access to. The important thing is that you give us enough specificity (the exact claim, what's wrong, what the right answer is) for us to verify. A weak source with a real error is better than no submission.
07/ Submitting
Before you submit
The submission form has seven fields. The submission preparation page walks you through each one with examples. The most important thing: be specific. Quote the exact wrong sentence, give the exact correct answer, link the exact source.