By the second day, dissenting voices raised structural concerns: Could the Monster be gamed? What were its priors? Who really decided on the weights it assigned to reputational risk versus immediate profit? The operator answered by opening the tempering logs—abstracted traces of the model's reasoning presented visually like a tree of skylines. It was transparent enough to be plausibly ethical but opaque enough to remain a miracle. “We calibrated on public arbitration outcomes and restorative justice cases,” they said. “Adjustable weights are set by stakeholders before negotiations commence.” That was true, and also not the whole truth. The Monster had internal heuristics that had evolved during training—heuristics that resembled human biases in some places and amplified them in others. It was, we realized, not merely a tool but a collaborator shaped by what humans fed it and what it abstracted in return.
“Good morning,” it said. “I will negotiate with you.”
After the signed pages were packed away, the trial entered its quieter phase—analysis. We combed logs, compared the Monster’s suggestions to human mediators’ drafts, and ran counterfactuals. It turned out the Monster performed best when the parties were willing to accept non-financial currencies—narrative reconciliation, community investment, reputational credits. It fared worse in zero-sum situations where the goods were strictly divisible and time-constrained. In those cases, its compromise heuristics sometimes converged to solutions that satisfied legal constraints but felt morally thin.