The fastest way to lose an on-call's trust is to hand them a confident answer that is wrong. During an incident, a plausible but unsupported root cause sends people down the wrong path at the worst possible time. So the bar for an AI incident commander is not fluency. It is whether the human can check the claim in seconds.
Every claim carries its evidence
Ember never returns a cause on its own. It returns a cause with the specific signals it keyed on: the log lines, the metric, the deploy that landed two minutes before the alarm. The confidence is not a vibe, it is a function of how well the evidence lines up. If the evidence is thin, the confidence is low and Ember says what it would check next.
cause Redis connection pool exhausted after deploy v2025.7.3 78%
evidence 15:02:11 pool timeout after 5000ms (redis)
redis connected_clients=200 maxclients=200
deploy v2025.7.3 landed 14:58, raised pool to 200Why grounding beats a bigger model
It is tempting to think the answer is just a larger model. But a larger model that is not grounded is a more convincing wrong answer. The win comes from the discipline: read the real telemetry, cite it, and rank by how well it supports the claim. That discipline is in the product, not only the weights.
If Ember cannot point at the line that convinced it, neither should you.
Grounding is also what makes Ember safe to trust with the fastest path. When the cause is cited and the confidence is honest, a drafted rollback is a decision the on-call can make in seconds instead of minutes.