Verification is a pipeline stage, not a filter

Three systems, one instinct

In Syrus, claim extraction and fact-checking are their own agents in the sequence rather than a post-hoc pass over the finished draft. In the survey analyser, a self-learning feedback loop and a dual-stream architecture exist for the same reason. In the posture device, a five-minute self-correction window and a ten-second stability filter sit between detection and alarm.

Different domains, same move: the system is not finished when it has an answer. It is finished when the answer can survive being questioned.

Why bolting it on afterwards fails

A verification step added at the end can only score what it is handed. It has no access to the intermediate state - which sources were considered, which claims were load-bearing, what the model was uncertain about. By the time you are checking a finished paragraph, the information you needed to check it properly has already been discarded.

Making verification a stage means the pipeline keeps that state around because a later stage depends on it. The architecture forces the honesty.

Alert fatigue is the same bug

The posture device makes this concrete. A binary good/bad alarm is a system with no verification stage - every detection becomes a notification, users stop trusting the notifications, and the device ends up in a drawer. Adding a correction window and a stability filter is not a UX nicety; it is the system checking its own claim before it acts on it. That single change removed 40% of false positives.

What it costs

Latency and complexity, mostly. Six agents are harder to reason about than one prompt, and a five-minute window means the device is deliberately slower to react than it could be. That is the trade, and it is worth making anywhere the cost of a confident wrong answer exceeds the cost of a slower right one - which, in research tooling and health hardware, is essentially always.