What's this about?
Enterprises hesitate to deploy AI agents mainly because of hallucination risk. This paper's framing shift is the interesting part: "zero hallucination" shouldn't be something you hope a good-enough model achieves — it should be a property a surrounding *system* enforces regardless of the model's quirks. HALO is a six-layer architecture (grounded generation, constrained execution, multi-signal verification, calibrated abstention, traceability, continuous oversight) demonstrated on a regulated insurance claims-extraction task.
What exactly is "calibrated abstention"?
It's the system's ability to say "I don't know" or "I won't answer" when it isn't confident enough — calibrated to actually match its real accuracy, rather than a model that always produces a confident-sounding answer whether it's grounded or not. Combined with the other five layers (like traceability — being able to show exactly which source supports each claim), it's an engineering answer to a problem people usually try to solve by picking a smarter model.
Why does it matter?
This is representative of a broader trend: as base model capability plateaus somewhat in the public conversation, more of the "make AI enterprise-ready" work is shifting to the scaffolding around the model rather than the model itself — directly relevant to any team building regulated or high-stakes agent deployments.