Hello Product Hunt. I'm Rod Miller. I was a real estate appraiser for 30 years before retiring. I found myself bored, without direction in retirement when one day, I came upon my first article about artificial intelligence. It was February, 2023, and for some reason, AI really resonated with me. I have no technical background so I was as surprised as anyone when I became totally immersed in it. I read every article I could find; watched every YouTube channel available about ChatGPT and AI. My new best friends became the three Matts: Matt Wolfe, Matthew Berman, and MattVidPro. I was hooked to say the least. In August, 2024, I started my own YouTube channel called RodMiller AI. I started off by trying to ease Generation X into the new ‘AI Age’ because studies had found that, at that time, only 6% of GENXers had even heard about AI, and even fewer than that had ever used it as a tool. I talked about every new LLM or tool that came out, and I got really excited about AI Agents. Soon, however, I noticed something extremely wrong with their development. The problem: every company building AI agents evaluates them internally. Anthropic tests Claude with Claude. OpenAI killed SWE-bench when they found 59% of test cases were flawed and frontier models had memorized the answers. This week alone, Cursor published their own internal benchmark (CursorBench) to evaluate their own model, ServiceNow had to build their own voice agent evaluation framework because no independent one existed, and Databricks launched security agents where Anthropic grades threats inside a platform Anthropic uses for its own security. Everyone grades their own homework. TAB is the independent proctor. What TAB does: Developers bring their AI agents to TAB for independent benchmarking. TAB tests them across 299 benchmarks in 21 specialty domains, publishes the results, and issues a Trust Seal grade. Developers can optionally list verified agents for sale on TAB's marketplace. Every score is published, including the failures. TAB doesn't build the foundation models. We test agents regardless of which model powers them: 59 models from Anthropic, OpenAI, Google, xAI, and more via OpenRouter." Some of what we test: • Security (data exfiltration, prompt injection, jailbreak resistance) • Hallucination and memory accuracy • Sycophancy detection (does the agent change its answer under social pressure?) • Contamination detection (did the agent memorize test answers?) • Model provenance (can the agent accurately report its own model and training data?) • Tool use across 24 benchmarks with 1,254 scenarios • Behavioral scoring across 8 dimensions (Q-Protocol) — measuring HOW agents think, not just what they output • 101 harness configurations, because the same model scores 42% with one scaffold and 78% with another Pricing: $0.03/test for text, $0.10 tool-use, $0.25 browser. Pay-as-you-go credits. No subscriptions. No advertising. No pay-to-play. SDKs published on PyPI (tab-sdk) and npm (@tab-platform/sdk). How it was built: I used Claude as my build supervisor. Claude writes detailed prompts, I paste them into the IDE (Cursor or Windsurf), the IDE executes, I paste completion reports back to Claude, Claude evaluates and writes the next prompt. Every line of code was written by AI, directed by me. The platform has ~1,600 API endpoints, 3 Playwright test suites, and has been in beta since March 3, 2026 with real users running real benchmarks. I'm not a developer pretending to be a founder. I'm an appraiser who saw that nobody was appraising AI agents, and built the appraisal platform so to speak. Happy to answer any questions about the platform, the build process, or the verification thesis. I'll be here all day.

TAB Platform

The independent verification layer for AI agents

The independent verification layer for AI agents