Comparator escalation for persistent relational state
The general distinct-architecture thesis was weakened. A bounded engineering-interface claim remains plausible.
Hypothesis
A directional, agent-local relational state will outperform a strong generic-state comparator on long-horizon social tasks.
Setup
Compare an explicit relational-state agent with increasingly capable generic-state and history-retrieval agents across matched social tasks.
Baseline / comparator
Generic persistent state with equivalent access to interaction history and a stronger reconstruction prompt.
Success criterion
A durable advantage across task families under matched context, compute and memory budgets.
Result
Stronger comparators reproduced much of the claimed advantage. Explicit relational state remained easier to inspect and replay, but was not uniquely necessary for the tested capabilities.
Verdict
The general distinct-architecture thesis was weakened. A bounded engineering-interface claim remains plausible.
Interpretation
The useful question moved from ‘is RBS necessary?’ to ‘when is maintained relational state a better interface than reconstruction or retrieval?’
Belief update
Confidence decreased in architecture-specific novelty and increased in studying capability-cost, auditability and state ownership directly.
Next experiment
Equalise memory budgets more strictly, add rare-event retention tests, and evaluate group-level state rather than extending the dyadic architecture by default.