[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-1718":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":16,"subscribersCount":16,"size":16,"stars1d":16,"stars7d":16,"stars30d":16,"stars90d":16,"forks30d":16,"starsTrendScore":16,"compositeScore":17,"rankGlobal":10,"rankLanguage":10,"license":18,"archived":19,"fork":19,"defaultBranch":20,"hasWiki":21,"hasPages":19,"topics":22,"createdAt":10,"pushedAt":10,"updatedAt":25,"readmeContent":26,"aiSummary":27,"trendingCount":16,"starSnapshotCount":16,"syncStatus":28,"lastSyncTime":29,"discoverSource":30},1718,"AI-Truth","Kunkun2116\u002FAI-Truth","Kunkun2116","See when AI is making things up.Make AI more honest — and show its receipts.","",null,"JavaScript",44,3,8,6,0,41.81,"MIT License",false,"main",true,[23,24],"chrome-extension","prompt-engineering","2026-06-12 04:00:11","# AI Truth\n\n**See when AI is making things up.Make AI more honest — and show its receipts.**\n\nEvidence labels for Claude, ChatGPT, and Gemini.  \nShow what was **searched in this conversation**, what comes from **memory**, and what is **inference** — instead of presenting everything in the same confident tone.\n\n\u003Cp align=\"center\">\n  \u003Cimg src=\"https:\u002F\u002Fraw.githubusercontent.com\u002FKunkun2116\u002FAI-Truth\u002Fmain\u002Fdocs\u002Fai_truth_homepage_mockup.png\" alt=\"AI Truth — evidence labels for AI responses\" width=\"720\">\n\u003C\u002Fp>\n\n### Demo\n\nhttps:\u002F\u002Fgithub.com\u002Fuser-attachments\u002Fassets\u002Ff06293c7-12ce-4cde-b0c4-b5b708c0c122\n\n---\n\n## ✨ Why this exists\n\nLLMs often mix **facts**, **memory recall**, **analysis**, and **guesses** into one seamless answer.  \nThe problem is not only hallucination — it is that everything can *sound equally certain*.\n\nAI Truth makes that difference visible.  \nIt combines a structured credibility prompt with a Chrome extension that surfaces evidence labels directly in the chat UI.\n\nRead the [Design Journey](.\u002Fdocs\u002Fdesign-journey.md) — the reasoning, the mistakes, and the iterations behind the framework.\n\n## 📦 What this repo includes\n\n- **A credibility prompt framework** for Claude, ChatGPT, and Gemini\n- **Two prompt versions** for different use cases:\n  - **Compact** for always-on personalization \u002F custom instructions\n  - **Full** for tighter control in new chats, research, and fact-checking\n- **Inline evidence labels** such as `[S1]`, `[M2+R2]`, `[S3+R2+F]`, `[U+C]`\n- **Risk prefixes** for high-risk domains like `⚠Legal`, `⚠Finance`, and `⚠Medical`\n- **A Chrome extension** that can:\n  - visualize labels in AI responses\n  - copy the framework prompt in one click\n  - inject prompt versions into supported chat interfaces\n\n## 🏷️ What the framework does\n\n### 1) Pre-output review\n\nBefore answering, the model runs a **hidden 6-point check** on key claims:\n\n1. factual accuracy\n2. unsupported conclusions\n3. time-sensitivity\n4. reasoning gaps\n5. missing premises\n6. completeness\n\nClaims that fail should be revised, downgraded, or marked uncertain before the final answer is shown.\n\n### 2) Evidence labeling\n\nThe final answer labels claims by evidence family, so the user can see what kind of support each statement actually has.\n\n| Family | Tags | Meaning |\n|--------|------|---------|\n| **S** (Searched) | S1, S2, S3 | Actually searched in this conversation: multi-source verified → single strong source → weak \u002F secondary source |\n| **M** (Memory) | M1, M2, M3 | Stable consensus → possibly outdated → time-sensitive, should search first |\n| **U** (User) | U | User-provided, not externally verified |\n| **R** (Reasoning) | R1, R2, R3 | Mechanically verifiable → framework-dependent → open synthesis |\n| **C** (Creative) | C | Generative ideation or design |\n| **F** (Fragile) | F | Insufficient, conflicting, or missing support |\n\nExamples: `[S1]` `[M2+R2]` `[S3+R2+F]` `[U+C]`\n\nHigh-risk content should also use domain prefixes such as:  \n`⚠Legal` `⚠Finance` `⚠Tax` `⚠Medical` `⚠Safety` `⚠Compliance` `⚠Engineering`\n\n## 🧠 Which prompt should you use?\n\nTwo versions are included:\n\n| | **Compact** | **Full** |\n|---|---|---|\n| **Length** | ~400 tokens | ~2500 tokens |\n| **Best for** | Personalization \u002F custom instructions \u002F always-on use | New chats \u002F research \u002F fact-checking \u002F tighter labeling |\n| **Behavior** | Lighter, faster, good default coverage | More explicit rules, stronger boundary control |\n| **Tradeoff** | More drift in long chats, more edge cases left to model judgment | Better consistency, but too long for most personalization settings |\n\n**Recommendation:** Start with **Compact** for everyday use. Use **Full** when you want tighter boundaries, stronger labeling, and less drift.\n\nPrompt files:\n\n- **Compact** → [`model_instructions\u002Fprompts-v7-compact.md`](model_instructions\u002Fprompts-v7-compact.md)\n- **Full** → [`model_instructions\u002Fprompts-v7-full.md`](model_instructions\u002Fprompts-v7-full.md)\n\nSuggested setup:\n\n- **Claude** → Settings → Customize Claude\n- **ChatGPT** → Settings → Personalization → Custom Instructions\n- **Gemini** → style \u002F preferences area\n\n## 🧩 Chrome extension\n\n### Install\n\n1. Open `chrome:\u002F\u002Fextensions\u002F`\n2. Enable **Developer mode**\n3. Click **Load unpacked**\n4. Select `extensions\u002Fchrome`\n5. Open Claude, ChatGPT, or Gemini\n\nSupported platforms:\n\n- **Claude**\n- **ChatGPT**\n- **Gemini**\n\n### Features\n\n#### 🎨 Label visualization\n\nWhen the AI outputs labels like `[S1]`, `[M2+R3]`, or `[S3+R2+F]`, the extension highlights paragraphs by credibility level.\n\nTwo display modes are available:\n\n- **Simple** — clean badge labels (`Verified`, `Caution`, `Ref`, `Alert`)\n- **Audit** — stronger color treatment with hoverable label pills and explanations\n\n#### 📋 Copy prompt\n\nCopy the current framework prompt from the popup in one click.\n\n#### ⚡ Prompt injector\n\nA **Credibility prompt** selector is injected into supported chat interfaces so you can pick and insert a prompt version without leaving the page.\n\n## 🔍 Why this is different\n\nThis project does not just ask the model to “be more careful.”  \nIt changes **what becomes visible to the user**.\n\n- **Unlike approaches that show how the model reasons, this project shows what each conclusion stands on.**\n- **Unlike confidence-style methods that output probability-like scores, it uses categorical labels that do not pretend to be calibrated.**\n- **It separates source type from reasoning type.**\n- **It adds a UI layer, so the framework is visible in real chat workflows rather than hidden in a prompt.**\n\n## 🗂️ Project structure\n\n```text\n├── docs\u002F                        # Design docs and assets\n│   └── design-journey.md\n├── extensions\u002F\n│   └── chrome\u002F                  # Chrome extension for visualizing labels\n│       ├── manifest.json\n│       └── src\u002F                 # content scripts, popup, utils, bundled prompts\n├── model_instructions\u002F          # Versioned prompt files\n│   ├── prompts-v7-compact.md\n│   └── prompts-v7-full.md\n├── README.md\n└── limitations.md               # Known failure modes and tradeoffs\n```\n\n## ⚠️ Known limitations\n\nSee [limitations.md](limitations.md) for the full list.\n\nImportant ones include:\n\n- self-checking still inherits the model’s own blind spots\n- long conversations can weaken prompt adherence\n- this framework has only been tested with frontier models; weaker models may mislabel claims and reduce reliability\n- platform DOM changes may break UI injection\n- labels improve auditability, not guaranteed truth\n\n## 🤝 Contributing\n\nContributions are welcome, especially in these areas:\n\n- selector robustness across supported platforms\n- label parsing and visualization quality\n- prompt versioning and evaluation\n- localization and bilingual documentation\n- docs, examples, and demo assets\n\n## 📄 License\n\nMIT\n","AI Truth 是一个旨在提高人工智能诚实度的Chrome扩展程序，它能够显示AI在对话中生成的内容是基于搜索、记忆还是推理。核心功能包括为Claude、ChatGPT和Gemini等模型提供证据标签，明确区分信息来源，并通过结构化的可信度提示框架进行预输出审查，确保关键声明经过六点检查后才呈现给用户。此外，该工具还支持一键复制框架提示及直接在聊天界面中展示证据标签等功能。适用于需要对AI生成内容的真实性进行验证的各种场景，如个人研究、事实核查以及高风险领域的应用（法律、金融、医疗等）。",2,"2026-06-11 02:45:38","CREATED_QUERY"]