Each ensemble is a pre-configured task type with thermodynamic parameters from the NSED whitepaper. The utility curve U(t) = 1-(1-pg)e-Λ(pv-pg)t - βt² predicts accuracy per round, combined with evaluator convergence for optimal stopping.

Deliberation Rounds = the maximum number of propose-evaluate cycles.
Convergence allows early termination if agents reach agreement before all rounds complete.

After each round, the room agreement A ∈ [-1, +1] measures the directional consensus of the whole ensemble: A = Σ signed_qv / Mtheoretical (total signed QV mass divided by the maximum possible positive mass). +1 = unanimous endorsement, 0 = split, -1 = unanimous rejection.

Evidence is a direct posterior snapshot each round: E(r) = max(A, 0) × U(r), where U(r) is the thermodynamic utility at round r. There is no accumulation — each round independently computes its halting signal, and negative room agreement is clamped to zero so a rejected winner produces no evidence. Halting fires when E(r) ≥ effort. A secondary path fires when t > Topt and room agreement is positive (A > 0.3). The raw net_support and cesaro_support vectors are also available in the SSE stream for per-proposal diagnostics.

Effort: the posterior evidence threshold (0.0 - 1.0). Higher effort requires stronger room agreement and higher utility before early termination.
Min Rounds: minimum number of rounds that must run before early termination is considered. This enforces a bootstrap minimum before halting.

Example: With effort=0.6 and min_rounds=2, each round computes E(r) = max(A, 0) × U(r); once E(r) ≥ 0.6 (after at least 2 rounds), deliberation stops even if more rounds were configured. If left empty, all deliberation rounds will always run.

User tools allow agents to call external functions during deliberation. Each tool call is forwarded to the dashboard for manual response.
The dm_user tool is a common preset that lets agents send messages directly to the user for clarification.