In JIT-Agent, candidate agent harnesses are compared under identical backbone models $\pi_\psi$ and evaluation random seeds to ensure valid comparison. Because protocol-compliant harnesses may still exhibit divergent performance and resource consumption, preference data $\mathcal{D}_{\text{I}}^{\text{pref}}$ is constructed using a strict Pareto-efficiency condition. A candidate harness $h^+$ is strictly preferred over $h^-$ ($h^+ \succ_\tau h^-$) if and only if:

$h^+ \succ_\tau h^- \iff r^+ > r^- \land \ell^+ \le \ell^- \land \kappa^+ \le \kappa^- \land (\ell^+ < \ell^- \lor \kappa^+ < \kappa^-)$

where $(r^\pm, \ell^\pm, \kappa^\pm)$ represent rollout averages for reward, execution latency, and monetary operational cost, respectively. The magnitude of preference is weighted by a multi-objective margin:

$\Delta_{\text{val}}(\tau; h^+, h^-) = \alpha_r (r^+ - r^-) + \alpha_\ell [\ell^- - \ell^+]_+ + \alpha_\kappa [\kappa^- - \kappa^+]_+$

where $[x]_+ = \max(x, 0)$ and alpha_r, alpha_ell, alpha_kappa ge 0 balance the value dimensions, enforcing that higher task rewards are favored while strictly penalizing excess latency and monetary expense.