We obsess over whether the model will hallucinate. The failure I actually lived through was quieter and worse.

The model didn't hallucinate.

It got confused, because we had poisoned our own context.

Months of design docs, decisions and reference material piled up in one place, at different dates, some superseding others, none of it curated. The agents, and the people, started giving incoherent answers, and nobody could say why. In a team of five sharing one context bank, that's binary: you operate like an A-Team, or everyone inherits the same rot at once, multiplied.

This talk is the war story and the discipline that came out of it: context as a curated product with an owner and a retention policy, human-to-AI alignment as process we already know (vision / status / roadmap as the interface), and the human overlaps that AI amplifies instead of removing.

The alignment problem was never the model's. It was ours.

Takeaways:
1. Context poisoning ≠ hallucination. Stale, contradictory, uncurated context makes agents confused, and it fails silently until a human says "this doesn't feel right."
2. The same vision / status / roadmap interface that aligns humans aligns agents. You already know how to do this.
3. At team scale the blast radius multiplies. One shared context bank means one team's context debt hits everyone at once. Govern it like a product.

Duration: 30 min (25 + 5 Q&A). Audience: builders, architects, engineering leads. Part 2 of the Build / Secure / Scale signature set.