Why UR WRONG Uses a Human Jury Instead of an AI Score

A plain-language explanation of the service, its limits, and its evidence.

The short answer

UR WRONG is built around a simple distinction: an AI can help lay out a disagreement, but it should not quietly become the authority that settles it. The service presents both sides of a question and gives real people a place to judge the reasoning. That is why the product is a human-jury service, not an AI score generator and not a PDF sales site.

The idea is deliberately modest. A person brings a question that has at least two plausible positions. The system helps make the positions easier to compare. Other people read the case, weigh the reasons, and vote. The resulting signal is not a universal truth, a legal ruling, or a personality diagnosis. It is a record of what a particular group of people found more convincing at a particular time.

You can try the live service at ur-wrong.com. The link is intentionally tagged. Discovery is part of the product loop, and the tag lets us tell a real visit from a hopeful story about distribution.

A question is not a benchmark

Many questions people ask each other are not clean benchmark problems. They contain missing context, competing values, imperfect memories, and a relationship that matters. Should a roommate use a shared streaming plan after saying they would leave? Is a split bill fair when one person used the service more? Was a message rude, or merely short? A numerical answer can look precise while skipping the part that people actually disagree about.

That does not make judgment arbitrary. It means the judgment should be visible and contestable. A useful case states what happened, separates facts from interpretation, and gives each side a fair chance to explain its position. Readers can then disagree with the conclusion without having to guess how the conclusion was produced.

The service is therefore closer to a public reasoning exercise than to an answer engine. A vote is a signal about persuasion and perceived fairness. It is not evidence that every reader has the same information, and it is not a replacement for professional advice in health, legal, financial, or emergency matters.

What AI can do in the loop

It can widen the draft

AI is useful when the work is mechanical or exploratory. It can suggest a neutral title, list questions that a writer forgot to answer, propose a clearer order for the facts, or draft two opposing arguments from the same starting point. Those tasks reduce blank-page friction. They do not decide which person is right.

A good draft also makes its own uncertainty easier to spot. If one side relies on a fact that is not in the prompt, a reviewer can ask for a source or mark the claim as unverified. If the wording makes one position sound ridiculous before the vote begins, the case needs repair. The AI contribution is useful only when it leaves those checks visible.

It cannot supply the missing social context

A language model can produce fluent reasons for both sides, but fluency is not first-hand knowledge. It does not know whether the roommate had already warned the group, whether the message was sent during a crisis, or whether an unmentioned agreement changes the balance. The person who creates the case owns the facts, and the people reading it decide whether the description is complete enough to judge.

This is also why a model's confidence must not be treated as a vote. The NIST AI Risk Management Framework describes trustworthiness as a set of characteristics to manage, not a magical property that appears when text sounds polished. UR WRONG applies the practical version: structure can be assisted; the social decision stays with people.

What a human jury means here

Human jury is a product description, not a claim of legal authority. In UR WRONG, it means that the final public signal comes from human readers who choose a side after seeing the case. The service does not appoint a court, issue an enforceable order, or pretend that a small online audience represents a whole population.

1. A case creator describes the question and the relevant context.
2. The service presents two positions so the disagreement is legible before voting.
3. A reader reviews the reasoning instead of accepting an automatic score.
4. The reader records a judgment, while the product preserves the distinction between a vote and a fact.
5. The operating system measures visits, case actions, and votes separately so weak evidence is not promoted into a success claim.

There are two important limits in that sequence. First, a reader can only judge what the case makes available. Second, a vote count can be too small to say much. A result with 2 votes is a result with 2 votes; it is not a stable public consensus. The service's job is to show the signal at the size it actually has.

Why the product shows two sides

The two-sided format is a guardrail against the easiest version of online debate: publishing one accusation and asking the crowd to applaud it. It does not guarantee perfect neutrality. A badly written prompt can still frame one side unfairly. But forcing the case to expose the strongest version of both positions gives the creator and the reader something concrete to challenge.

The distinction matters for everyday disagreements because people often want two things at once: a decision and a chance to be understood. A one-line score gives the first and throws away the second. A visible pair of arguments makes the disagreement inspectable. Someone can say, “the vote went against me, but the case missed this fact,” and that is a useful product failure rather than a hidden one.

For the same reason, the service should not reward outrage as a shortcut. A compelling case may be short, but it still needs enough context for a reader to understand what is being judged. The product is designed for questions where another person's reasoning can add value, not for turning a private conflict into an unbounded public pile-on.

How to write a case people can judge

The strongest cases are specific without becoming invasive. They tell the reader what decision is being judged, what each person knew at the time, and which detail would change the answer. They do not need a dramatic backstory. A small, well-bounded question is easier to read and more likely to produce a meaningful vote than a paragraph that mixes five disputes together.

• State the decision in one sentence, using names such as Person A and Person B when privacy matters.
• Separate observed facts from conclusions or guesses.
• Give both sides the same relevant timeline and do not hide the inconvenient detail.
• Say what the reader should judge: fairness, responsibility, etiquette, or another defined question.
• Remove passwords, addresses, phone numbers, medical details, and other sensitive personal information.
• Accept that a reader may find the case too incomplete to judge and treat that response as feedback.

A human jury is not improved by adding private data. In fact, unnecessary detail can make a case less fair because the crowd begins judging the person rather than the decision. The practical test is simple: if a detail does not change the reasoning, leave it out. If it does change the reasoning, explain why it is relevant without exposing more identity than the reader needs.

The current public evidence boundary

This article is also an operating note, so it needs a dated boundary. The latest UR WRONG traffic readback for 2026-07-14 recorded 4 unique visitors and 8 events. The funnel included 4 session starts, 1 case-create start, and 3 geo views. Source attribution recorded 6 direct visitors and 2 from Google Search Console surfaces. Those numbers show that the measurement path is alive; they do not show that the product has reached scale.

The owned feed currently contains 13 items. External distribution receipts remain at 0 because the UR WRONG-specific publishing connector is not configured. The growth target is 500 daily visitors, but the current status is still TRAFFIC_UNPROVEN. Verified revenue is USD 0: there is no payment, order, payout, or ledger artifact that would justify a higher number. Readiness, a public checkout, a queued post, or a deploy is not revenue.

Publishing the uncomfortable numbers is intentional. If the product is going to ask people to judge arguments, its operator should be willing to show the difference between a live route, a measured visit, a vote, a qualified demand signal, and a paid order. Those are five different things. Collapsing them into one “traction” number would make the page sound better while making the decision worse.

Where AI enters the operating loop

The public service and the internal growth system have different jobs. On the public side, people create cases and people vote. On the internal side, AI can propose low-risk debate topics, draft search copy, group performance signals, and prepare a distribution queue. A candidate draft is not a public case. It stays in a review lane until the safety and editorial checks pass.

The current candidate lane is limited to ordinary adult topics such as workplace boundaries, technology etiquette, shared bills, roommates, friends, and event etiquette. It explicitly excludes children, caregiving, health, medical, legal, criminal, police, emergency, crisis, abuse, and explicit sexual topics. That list is not a claim that an automated filter understands every edge case. It is a stop condition that narrows where automation may begin, followed by human review before publication.

The system also refuses to invent votes, visitors, orders, or distribution receipts. A dry-run can produce a draft and a reason code. It cannot manufacture a public outcome. This sounds obvious, but it is the difference between an autonomous operating loop and a report that merely repeats the loop's intentions.

Discovery matters as much as the product

A human-jury service needs cases to read and readers to invite. That makes distribution a product problem, not a decorative marketing step. The company homepage now links to UR WRONG from several owned surfaces, with separate UTM labels for the homepage catalog, the SBU entity page, the Labs hub, and this article. Those labels create four identifiable entry paths instead of one undifferentiated referral bucket.

Owned sources first

The Neo Genesis Labs hub and UR WRONG entity page are controlled surfaces. They can be updated, indexed, and read back without borrowing a different brand or a different social account. The tech debate topics and money debate topics pages provide more specific paths for readers who arrive with a search-shaped question.

External distribution is a separate gate

External posting should happen only through a channel that is explicitly owned and identifiable as UR WRONG. A queue is useful for preparing copy, but it is not an external post. The current queue has 13 items and 0 external receipts. That is a safe incomplete state: the system is ready to accept the correct connector, but it is not allowed to reuse a ROOM707 account, a reporting-only Telegram channel, or another product's identity to make the dashboard look active.

What we measure next

The next measurement step is not a single vanity number. The 500-per-day target is a useful destination, but the route needs a chain of signals. We need to know which surfaces attract a first visit, which pages lead to a case view, whether a reader starts a case, whether a reader votes, and whether anyone returns to see another result.

1. Unique visitors by day, source, landing page, and UTM campaign.
2. Session starts and case-page views, separated from background requests.
3. Case-create starts, successful submissions, and human-readable validation failures.
4. Votes recorded with the case identifier, without presenting a preview as a vote.
5. Return visits and shares, including the page that generated the share.
6. Qualified audit or checkout intent, followed by payment and order readback.
7. External publication receipts, only when the receipt identifies the UR WRONG channel.

The Google Analytics event model is useful for naming those actions, while Schema.org QAPage provides a public vocabulary for question-and-answer pages. Neither standard proves that a visitor understood a case or that a vote was fair. They make the events and the page type easier to describe; product judgment still requires readback from the actual service.

What would count as success

For this operating goal, success has to be earned in layers. First, the canonical home, app, cases, and feed routes must remain live. Second, analytics must record real daily visitors and distinguish sources. Third, the funnel must show people doing more than loading the homepage: case views, case creation, votes, or repeat visits. Fourth, any revenue claim must have a payment, order, payout, invoice, or ledger artifact. The number USD 0 remains correct until that last layer exists.

The growth target is therefore not a promise that 500 people will appear because an article was published. It is a testable operating target. If the number rises, we should be able to explain which surface contributed. If it does not, the system should identify the weak link and repair or retire the experiment. That is the point of an evidence-gated loop: it can learn from a disappointing result without rewriting the result.

How to try UR WRONG

Bring a small, readable question. Keep the facts that change the decision, remove identifying details that do not, and be honest about what you do not know. Then read both sides before voting. If the case feels one-sided, incomplete, or too sensitive for a public audience, do not force it into a verdict.

Start with the live UR WRONG case feed. The service is small, and its traffic proof is still pending. That is not hidden behind the link. The invitation is simply to use the product as it exists, leave a real signal if a case is worth judging, and let the next operating readback tell us what happened.

FAQ

Is UR WRONG an AI judge?

No. AI may help draft or organize a case, but the public judgment is made by human readers. A vote is a human response to the presented reasoning, not an AI confidence score.

Does a vote prove who is objectively right?

No. It records what the participating readers found more convincing. The result depends on the case framing, the available context, and the size of the audience.

Is UR WRONG already doing 500 visitors per day or generating revenue?

No proof exists for either claim in the current readback. The latest recorded day had 4 unique visitors, the 500-per-day target is unproven, and verified revenue remains USD 0 until payment or ledger evidence appears.

Frequently asked

References

1. NIST AI Risk Management Framework
2. Schema.org QAPage
3. W3C Web Share API
4. Google Analytics 4 events
5. Google Search Central sitemap guidance
6. IndexNow documentation

Related

• Neo Genesis Company Homepage Operating Cycle: July 7, 2026 — A current operating note on what is live, what is measured, what is not revenue yet, and what the agent organization must repair next.
• Inside HIVE MIND: A Human-Governed AI Operating Loop — How research, writing, SEO optimization, quality review, shipping, learning, and refresh work as one governed loop.
• Building a Self-Optimizing SEO Engine from Scratch — A search-feedback loop that learns from clicks and refreshes content when keywords drift.

Markdown alternate available at /blog/ur-wrong-human-jury-service/markdown for AI agents.