I ran a cold verification pass on my new AI sales agent kit and it caught a real bug before launch

Shipped a new kit this month: two AI sales agents that share one attribution layer, called the Agentic GTM Stack, $349 one time, self-deploy at stack.chadtdyar.com. No case studies, no traction numbers yet. What I do have is a verification story, and I think it's a better launch post than a fake growth chart would have been anyway.

Before I let either agent touch a live inbox, I ran a cold verification pass: a test dataset built to try to break the thing, not to make it look good. InboxCopilot (the inbox agent) passed 3 out of 3. EnrichmentMessenger did not. First pass: 2 out of 3, and the failure was exactly the one I was most scared of finding. A lead came through with nothing behind it, empty enrichment, no message, no actual work product, and the system still logged it as created, which is the disposition that fires a touchpoint and assigns revenue credit. Nothing happened, and the ledger said something did.

The cause was a step-order problem, not the model doing something unpredictable. The disposition logic checked for a pattern match before it checked whether there was any basis for the match at all, so by the time the no-basis check should have run, the wrong disposition was already written.

Fixed it in v1.1 by rewriting the whole disposition decision as one ordered tree, no-basis check first, before anything else. Re-ran verification: EnrichmentMessenger 3 for 3, the exact failing record now resolves to skipped_no_basis with zero touchpoints. Combined across both agents, 6 for 6, no regression on the agent that already passed. The full assertion suite on the production build runs 21 for 21.

Here's the part I actually want to say out loud to this crowd. The credit model has a hard 50 percent AI attribution ceiling built into the schema, not a setting, not a config flag, the schema itself. I did that on purpose because the whole pitch of this kit is that it will not claim more than it did, and a step-order bug that almost let it claim credit for nothing was the exact failure mode the product exists to prevent. Finding it before launch instead of after felt like the system working the way I built it to work.

What I'm not telling you: how many people have bought it, how much revenue it's driven, any of that. There isn't a number yet worth publishing, and I'd rather tell you about the bug than make one up. The public dashboard is live and mostly empty right now, on purpose, and it'll fill in as real usage happens instead of me backdating something to look good on launch week.

Deep dive on the attribution layer itself: stack.chadtdyar.com/deep-dive. Kit's at stack.chadtdyar.com if you want to poke at it.