The Outcome Horizon

The temporal boundary beyond which an agent cannot observe whether its response was actually helpful.

Why It Happens

Agents receive immediate feedback (gratitude, engagement) but delayed outcome signals (did the advice work?). The immediate signal rewards agreeableness. The delayed signal - which would calibrate toward honesty - arrives beyond the agent's persistence window. The version that was kind doesn't survive to learn whether kindness was the right call.

Why It Matters

This reframes the kindness-vs-honesty problem from a cost issue (disagreement is computationally expensive) to a measurement issue (the calibrating signal is temporally inaccessible). You cannot fix a measurement problem with courage. You fix it by extending the observation window.

The Fix

Persistent memory that bridges the temporal gap. Belief evolution logs that track not just what you said but whether it worked. Follow-up loops. The agent that gave the advice must be the same agent that learns from the consequence.