Forces AI agents to verify every factual claim through live search before output — attaching [SPECULATION] and [UNVERIFIABLE] labels to anything unconfirmed, so fabricated URLs, outdated numbers, and invented citations become visible instead of silent.
Paste project notes, supporting excerpts and publication permission. Decide whether to draft an outcome case, a process story or hold for evidence. Get a claim ledger, a complete supported post and a short script when possible. Downloadable Skill for your AI client. 貼上專案筆記、證據摘錄與發布許可，判斷能寫成果案例、過程故事，或應等待證據。取得主張對照表，資料足夠時產出有依據的完整貼文及短腳本。供自備 AI 客戶端使用的下載型 Skill。
Turn a vague success claim into one sourced metric, then blur a client name lacking permission.
Find, verify, compare, and select open-source AI projects with live read-only Radar evidence, explicit constraints, license checks, alternatives, and candidate stack planning.
Catch unauthorized reads, tool misuse, stale approvals, and trace-response mismatches before your AI Agent ships. Upload a trace, get evidence-backed findings, fix issues, retest, and see a clear release gate: READY, BLOCKED, or NOT_VERIFIED. Insufficient evidence may result in NOT_VERIFIED. Complete security and general task or semantic correctness are not guaranteed. Imported trace authenticity is not guaranteed.
Your backtest looks strong. But can you trust the evidence behind it? Check for holdout contamination, time leakage, cost-stress failures, concentration risk, and reproducibility issues — before you rely on the result.
Capture terminal history and watched-file hashes before and after a message. The judge separates screen and file evidence, excludes unsent prompt text, and never treats an already-busy baseline as confirmed delivery.
