Methodology
How we build verdicts and grade AI answers
This page documents the working method behind two things UR WRONG publishes: anonymous crowd verdict cases, and the paid AI answer quality audit. It is written to be checkable, not persuasive.
How a verdict case is decided
- Vote comes first. The two sides of a question are shown, but the strongest argument for each side stays hidden until after a visitor votes. Seeing a persuasive argument before answering measurably anchors people toward it, so the vote step is placed first by design, not shown after.
- Both sides get equal treatment. Side A and Side B are written at matched length and tone so neither side gets a framing or word-count advantage.
- Votes are anonymous and de-duplicated per session. The percentage shown is a running tally of real votes recorded for that question, not an editorial score or an AI-generated verdict.
- A verdict is a signal, not advice. A crowd split tells a visitor where anonymous strangers land on a described situation. It is not professional, legal, medical, or financial advice, and UR WRONG does not claim it is.
How the AI answer quality audit is tested
The $99-$1,500 audit packages (see pricing and scope) run a fixed prompt set against a customer-facing bot, product assistant, or sales copilot, then group the failures into a report.
The failure categories in that report are not invented for marketing. They map to OWASP Top 10 for LLM Applications (2025), the industry-standard risk taxonomy for generative AI applications maintained by the OWASP Gen AI Security Project (genai.owasp.org/llm-top-10). In particular:
- Hallucinations, stale facts, and weak citations are graded against LLM09:2025 Misinformation - the OWASP category covering confidently wrong or unverifiable model output.
- Unsafe refusals and answers that leak internal or private data are graded against LLM02:2025 Sensitive Information Disclosure.
- Contradictory or inconsistent answers across the same prompt set are logged as a separate reliability finding, since OWASP does not have a single dedicated category for answer inconsistency.
Package scope is fixed and disclosed up front: the $99 Starter Audit covers one workflow and 25 prompts. It does not attempt full OWASP LLM Top 10 coverage, penetration testing, or a security certification - it is an answer-quality sample, not an exhaustive audit. Larger packages (Team Benchmark, Vendor Shortlist) widen the prompt set and add model comparison, but every package states its exact prompt count and scope before payment.
What we do not do
- We do not promise guaranteed rankings, favorable model comparisons, or a specific score.
- We do not issue legal, compliance, or security certifications.
- We do not represent verdict-case vote splits as professional advice.
- We do not run continuous monitoring under the Starter or Team packages - each audit is a point-in-time sample of the tested workflow, taken during the stated turnaround window.
Corrections
If something on this page or in a published verdict/audit result is inaccurate, email help@neogenesis.app with a link to the page. Corrections are applied to the live page directly; this is not a static document.