{"slug": "the-answer-looked-right-that-was-the-problem", "title": "The Answer Looked Right. That Was the Problem.", "summary": "A developer built The Verifier, an evidence-first research assistant that checks public claims by searching for both supporting and contradicting evidence. The tool uses metadata such as publication dates to resolve conflicts between sources, as demonstrated when it determined that Brian Niccol is the current CEO of Starbucks despite an older source naming Laxman Narasimhan. The developer emphasized the importance of showing the 'messy middle' of research rather than hiding contradictions.", "body_md": "I started building **The Verifier** because of a slightly uncomfortable thought:\n\nWhat if an AI gives you a perfectly written answer based on yesterday's truth?\n\nNot a hallucination. Not an obviously fake source. Just an answer that *used to be correct*.\n\nThat felt more dangerous.\n\nA polished lie is easy to fear. An outdated fact wearing a confident suit? That one can quietly walk straight into reports, presentations, articles, and important decisions.\n\nSo I decided to build something that would not simply answer a claim.\n\nIt would argue with itself first.\n\nThe Verifier is an evidence-first research assistant for public claims.\n\nYou give it a claim such as:\n\nVerify that Brian Niccol is the CEO of Starbucks.\n\nInstead of immediately producing a confident paragraph, it investigates the claim from two opposing directions.\n\nOne research lane looks for evidence supporting the claim. Another actively searches for credible evidence contradicting it.\n\nThis matters because normal search behaviour has a bias: once we find one convincing source, we usually stop looking.\n\nThe Verifier does not get that luxury.\n\nIt has to search both sides, place the evidence next to each other, and show its work.\n\nThe Starbucks leadership example turned out to be a great demonstration.\n\nOne public source identifies Brian Niccol as Starbucks' current chairman and CEO.\n\nAnother Starbucks source names Laxman Narasimhan as CEO.\n\nBoth sources are real. Both come from Starbucks. Both look authoritative.\n\nSo which one should the system trust?\n\nThis is where the project became more than a fancy search interface.\n\nThe answer was hiding in the metadata.\n\nMost people do not get excited about fields such as datePublished, dateModified, or ISO timestamps.\n\nHonestly, neither did I.\n\nBut those tiny pieces of metadata became the detective of this story.\n\nThe Verifier extracts machine-readable dates from every source. It records which metadata field was used, preserves the original value, and normalizes the date to UTC.\n\nNow the system can compare evidence deterministically.\n\nIt does not choose a source because it sounds more confident. It can explain that one source represents an older leadership announcement while another contains newer evidence about the company's current leadership.\n\nThat difference matters.\n\nThe conclusion is not merely:\n\nBrian Niccol is CEO.\n\nIt becomes:\n\nThe claim is supported by newer public evidence. The conflicting source is legitimate, but outdated.\n\nSuddenly, the contradiction is not something to hide. It is part of the explanation.\n\nA lot of AI tools clean up the messy middle.\n\nYou provide a question. Something complicated happens. A neat answer appears.\n\nMagic.\n\nAlso slightly terrifying.\n\nI wanted The Verifier to show the messy middle.\n\nThe interface keeps the supporting and conflicting evidence in separate lanes. Each source displays its role, URL, relevant statement, and date information.\n\nThe user can inspect why the system reached its conclusion instead of being asked to trust a glowing confidence score.\n\nBecause “97% confident” sounds impressive, but it does not tell me whether the model checked last year's announcement against today's leadership page.\n\nEvidence does.\n\nI also wanted the experience to feel conversational.\n\nThe original idea was simple: click **Speak brief**, say the claim, and continue asking questions naturally.\n\nThe browser had other plans.\n\nAt one point, the button barely behaved like a button. Then voice mode required clicking it repeatedly. Even worse, every new interaction felt like the previous conversation had been erased from existence.\n\nThe Verifier had the memory of a goldfish with Wi-Fi.\n\nThat forced me to rethink the feature.\n\nVoice could not be a one-shot transcription gimmick. It needed to be a continuous conversation where the active claim, collected sources, and date resolution remained available for follow-up questions.\n\nI also added text as a permanent fallback because demos—and browsers—have a wonderful sense of dramatic timing.\n\nThis became my favourite part of the project.\n\nAfter completing the investigation, The Verifier does not automatically store the result.\n\nIt asks the user for approval.\n\nThat is not just a polite message in the interface. The restriction is enforced on the server.\n\nThe server creates a one-time approval token and refuses to persist the dossier until explicit approval is received. Export remains locked as well.\n\nIf the user chooses to keep investigating, nothing is saved.\n\nOnly after server-confirmed approval does the dossier become permanent and downloadable.\n\nThe system can investigate, compare, and explain.\n\nThe final decision still belongs to a human.\n\nThe first UI used fictional “Maya Chen” and “Northstar AI” evidence while the backend used the real Starbucks case.\n\nVery coherent. Absolutely no notes.\n\nThat mismatch had to go.\n\nThen came GitHub authentication, voice bugs, disappearing conversational context, provider configuration, Groq setup, and a Daytona API key that continued rejecting me like it had a personal grudge.\n\nSome parts worked immediately. Several parts worked only after I stopped assuming they worked. A few parts smiled during testing and broke the moment I opened the demo.\n\nThat is the real version of building for a hackathon.\n\nIt is not one cinematic montage with energetic music. It is a series of tiny mysteries, questionable browser tabs, and commits that begin with the word “fix.”\n\nBut each problem improved the project.\n\nThe fictional demo became a real public case. The voice button became a conversation. The date field became an explanation. The approval message became a server boundary.\n\nThe Verifier is not trying to make AI sound smarter.\n\nIt is trying to make AI easier to question.\n\nThe important output is not only the final conclusion. It is the path:\n\nThat path becomes an inspectable dossier that a researcher, journalist, policy team, or compliance reviewer can defend.\n\nThe internet changes. Leadership changes. Policies change. Yesterday's correct answer quietly becomes today's misinformation.\n\nThe Verifier is my attempt to catch that moment.\n\nNot by asking you to trust another polished AI answer.\n\nBy showing you the evidence, explaining the conflict, and waiting for you to decide.\n\n**Verify first. Save only when you are ready.**\n\n🚀 **Live demo:** [the-verifier.vercel.app](https://the-verifier.vercel.app)\n\n🔗 **Project:** [github.com/Reet24-del/the-verifier](https://github.com/Reet24-del/the-verifier)\n\nIf you build with AI, I would love to know: **what part of an answer do you wish AI tools showed more clearly?**", "url": "https://wpnews.pro/news/the-answer-looked-right-that-was-the-problem", "canonical_source": "https://dev.to/reet_singh_ee97da2986538e/the-answer-looked-right-that-was-the-problem-lhn", "published_at": "2026-08-30 16:28:44+00:00", "updated_at": "2026-08-30 16:53:07.067036+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-tools", "ai-products", "developer-tools"], "entities": ["The Verifier", "Brian Niccol", "Starbucks", "Laxman Narasimhan"], "alternates": {"html": "https://wpnews.pro/news/the-answer-looked-right-that-was-the-problem", "markdown": "https://wpnews.pro/news/the-answer-looked-right-that-was-the-problem.md", "text": "https://wpnews.pro/news/the-answer-looked-right-that-was-the-problem.txt", "jsonld": "https://wpnews.pro/news/the-answer-looked-right-that-was-the-problem.jsonld"}}