Agent 1
Amna
Retail worker, Deira. AED 4,500/month income, AED 1,800 rent.
Agent 2
Tariq
Tech consultant, Dubai Marina. AED 45,000/month income, AED 12,000 rent.
Agent 3
The Market
Real estate and retail vendor protecting a ~20% margin.
🔬 Research & Hypothesis
Bounded rationality
Classical economics assumes people optimize with full information. Herbert Simon argued that real decision-makers have limited information, time, and computing capacity, so they "satisfice": they choose options that are good enough. This study asks whether agents from very different Dubai income groups show different decision rules when the same shock hits.
Context constraints
A small language model has a limited context window and limited ability to keep many facts separate. This is a machine analogue of bounded rationality: when several agents' numbers are packed into one prompt, the model can blur them together. That limitation is itself a research object here, not only a nuisance.
⚠️ Model Defect & Critique
Observed defect (Round 3): Tariq's agent accidentally adopted Amna's lower-income rent variable (AED 1,800) in place of his own (AED 12,000). This is an AI hallucination attributable to context degradation in a small (1.5B-parameter) model.
Evidence from the log
Likely mechanism (untested hypothesis)
Tariq's identity sits only in the system prompt, while the Round 3 prompt contains a long transcript with Amna's figures. Recent, repeated numbers can outweigh the persona facts for a small model.
Limits of this critique
This is based on a small number of runs with sampled (random) output, so it shows the failure can occur, not how often. A rigorous test would repeat Round 3 across many seeds and count how often each agent cites another agent's numbers.
💬 Simulation Dialogue Logs
Raw model output, shown unedited.