Controllable traffic simulation agents (behavior latents) trained with UCB-guided surrogate Bayesian optimization will exhibit coalition-based equilibrium deviations (beyond unilateral) in high-density scenarios, with deviation magnitude scaling with the number of coordinated agents (R² > 0.85).

This discovery has a Claude-generated validation package with a full experimental design.

Precise Hypothesis

Spine & Adversarial ReadReady for validation

“This hypothesis tests whether coordinated subsets of Bayesian-optimization-controlled traffic agents produce collective equilibrium deviations whose magnitude scales predictably (R²>0.85) with coalition size.”

• highThe 'coalition deviation' metric may simply reflect a mechanical artifact of giving coordinated agents a larger joint optimization budget (shared surrogate over concatenated latents) versus independent agents each optimizing separately — this is an optimization-power confound, not emergent game-theoretic coalition behavior.
  Protocol requires equalizing total BO iterations per-agent (not per-coalition) and includes a shuffle-membership ablation, but the EVP does not yet specify a formal control where independent agents are given an equivalent joint surrogate without coordination (e.g., shared surrogate, independent objectives) to fully isolate 'shared information' from 'shared objective.' This gap should be closed by adding a third arm (shared-surrogate/independent-objective) before results are trusted.
• mediumWhy UCB-guided BO specifically, and why these particular simulators (Waymax/nuPlan/SMARTS) and this specific density threshold (0.3 veh/m/lane) rather than alternative acquisition functions (Thompson sampling, EI) or other traffic sim platforms — the methodology choice is not justified against alternatives, raising concern that results are specific to this arbitrary configuration rather than general.
  UCB is chosen for its explicit exploration-exploitation tunability (β) which is needed to test robustness across the boundary conditions, and open-source simulators with real-world scenario logs (Waymo/nuScenes) are chosen for reproducibility and realistic density distributions. However, the EVP does not include a comparison arm against Thompson sampling or Expected Improvement acquisition, so the claim that this is a general BO-driven phenomenon (rather than a UCB-specific one) is not yet established — this should be added as a robustness check before generalizing beyond UCB in any publication.
• highWith Verification Confidence scored at 0.00 and Evidence Strength only 0.69 in the source discovery record, there is essentially no existing empirical grounding for this claim — it is a theoretically motivated hypothesis with no pilot data, and the R²>0.85 threshold may have been chosen without any prior distribution of plausible effect sizes to calibrate against.
  This is acknowledged directly: the EVP's abort checkpoints (particularly Checkpoint 1, a cheap pilot at n=5 seeds) are designed precisely to detect early whether any effect exists at all before committing to the full $165K validation spend. No resolution beyond this staged, cost-gated design is claimed — if the pilot shows no effect, the hypothesis should be abandoned rather than pursued to full scale.

Experimental Protocol

ROI Projection

Implementation Sketch

for topology in [merge, intersection, weave]:
  for density in [0.1, 0.3, 0.5]:
    for k in [2,3,4,5,6,8,10]:
      for seed in range(30):
        scenario = init_scenario(topology, density, seed)
        agents = sample_agents(scenario, n_total)
        coalition = agents[:k]; independents = agents[k:]

        # Independent baseline
        for agent in agents:
          z_agent = UCB_BO(surrogate=GP(), acquisition=UCB(beta),
                            objective=agent.reward, n_iters=50)
        joint_value_independent = simulate(agents, latents=z_independent)

        # Coalition condition
        z_coalition = UCB_BO(surrogate=GP(dim=k*d), acquisition=UCB(beta),
                              objective=sum(coalition.rewards), n_iters=50)
        joint_value_coalition = simulate(coalition + independents_rest,
                                          latents=[z_coalition, z_independent_rest])

        delta = joint_value_coalition - joint_value_independent
        log(topology, density, k, seed, delta, trajectories, collisions)

fit_model = best_of([power_law, linear, log_linear]).fit(delta ~ k)
report(R2=fit_model.r2, ci=bootstrap(fit_model, 2000))
run_ablations(randomize_calibration, remove_observability, shuffle_membership)