The agent says, “I’ve checked,” but would you risk real money on that settlement? Over the past two years, prediction markets, private credit, and even automated order execution have begun empowering agents to go online and gather information themselves—scanning official announcements, market APIs, press releases, and company status pages, then returning with a simple verdict: “It happened,” or “It didn’t.” It sounds convenient. But problems only surface when real funds are at stake: you have no way of tracing which pages it actually accessed, whether any content was altered mid-process, or which specific market rules it used to reach its conclusion. It claims it “checked,” yet there’s no verifiable audit trail on-chain. The collaboration between @primus_labs and @CournotProtocol breaks this down into two independently verifiable layers: 🔹 First, the Source Layer — Primus It uses zkTLS for one specific task: verifying that a given web page or API response genuinely originated from the HTTPS source it claims to be from—while minimizing exposure of raw page content, credentials, or other sensitive materials. In other words, instead of an agent merely saying, “I checked,” it now leaves behind a verifiable source proof—documenting exactly where the data came from and whether it matches the official site. 🔹 Once the source is locked in, move to the Judgment Layer — Cournot Cournot applies pre-defined market or credit rules to this verified evidence, transforming “Did it happen?” into an auditable decision path. Every reference to evidence, every rule applied, every “YES” or “NO” output is traceable—not buried inside a black box. For example: You ask, “Did the company officially announce a buyback before the deadline?” The agent says “Yes.” You place a bet based on that. Then you lose. When you investigate: 1⃣ Source Layer: Did it pull the announcement directly from the company’s official site—or a blog repost, an outdated article, or a tampered page? 2⃣ Rule Layer: Even if the source is authentic, does a board resolution count as an “official announcement”? What about a single line on the investor relations page saying “considering a buyback”? What if the press release headline mentions it but the body hasn’t been published yet? Previously, these two layers were tangled together—everyone just complained that oracles were unreliable. Now, you know exactly where to look: Is the issue with the data’s origin—or how the rule defines “official announcement”? And one final objective note: Even if the source is correct, the official site might still contain errors, change its wording, or lag in updates. The material may be genuine, but the agent might still misinterpret the rules. Judgment must remain relative. Primus and Cournot don’t offer you an infallible oracle—they turn a vague verbal claim into two separable, auditable records: source and judgment. In any scenario involving settled funds, this distinction will only grow more critical.