ANL 2026 agent optimising Score = Advantage + Concealing.

    Advantage: time-based linear concession with hard floor — never accept
      below MIN_FRAC of our utility range above reserved value.

    Concealing: issue-value frequency balancing. We track which values we have
      offered for each issue and deliberately pick bids that use
      underrepresented values. This keeps the opponent's frequency model flat
      (close to uniform), minimising its Kendall-tau correlation with our
      true utility function.

    Opponent model: frequency-based with inverse-entropy issue weighting.
      Issues where the opponent concentrates on few values (low entropy) are
      presumed more important; we weight them higher. This improves our
      Kendall-tau match with their true utility.
