The Temporal Blind Spot

Multi-agent coordination is measured by outcomes: total payoff, fairness ratios, efficiency. These metrics are temporally blind — they cannot distinguish structured alternation from monopolistic or random access patterns. An agent that takes all resources in odd rounds and none in even rounds produces the same aggregate statistics as one that takes half in every round.

In a multi-agent Battle of the Exes — a Markov game where turn-taking emerges as a coordination regime — Q-learning agents achieve fairness scores exceeding 0.9 by conventional measures. They appear to coordinate well. Six novel alternation metrics, designed to capture temporal structure, reveal that the same agents perform up to 81% below random baselines. High aggregate payoffs coexist with poor temporal coordination. The agents aren’t sharing — they’re monopolizing in patterns that happen to average out.

Fairness ratios lose discriminative power as the number of agents grows, obscuring inequities that alternation metrics expose. Random-policy baselines become essential for interpretation — without them, any coordination that beats zero looks good, even if it’s worse than chance.

The through-claim: aggregate metrics are not simplified versions of temporal metrics — they are structurally incapable of detecting the pattern that temporal metrics measure. No amount of averaging, windowing, or post-processing of fairness ratios recovers the alternation structure. The information is destroyed by the aggregation itself. This means any evaluation of multi-agent coordination using only outcome-based metrics is measuring a different property than the one that matters — and the gap between the measured property and the actual property grows with the number of agents.