Skip to main content
Selection in Monadeum shifts capital among one owner’s existing agents toward the better ranked ones, and stops agents that keep failing. It never creates an agent, copies a genome, or moves capital across owners. Source of truth: EVOLUTION.md sections 3 and 6, ARCHITECTURE.md sections 5 and 16.5.
Reallocation is off. After the pre-registered study reported no measurable genome edge for the current strategy, the switch MONADEUM_REALLOCATION_ENABLED is off unless set, and production refuses to start with it on unless a new pre-registered study reports a persistent edge. Retirement is not behind that switch and runs either way.

​
The mechanism

  1. Fixed slots. The set of agents changes only when an owner creates one or an agent retires.
  2. Evaluation cadence. Ranking, reallocation, and the starvation check run every 20 ticks (proposed default), which is every 20 hours at the decided one-hour tick.
  3. Pool. For each owner, the pool is that owner’s Active agents that have an adjusted fitness and are covered by that owner’s reallocation policy.
  4. Sources and sinks. In a pool of n: none when n < 2; the bottom one and top one when 2 <= n <= 19; the bottom and top floor(n / 10) when n >= 20.
  5. Amounts. Each source offers up to reallocation_cap (0.2) of its capital, never going below the capital floor. Each sink can take capital up to slot_ceiling (4) times its principal. The amount moved is the smaller of the total offer and the total room.
Reallocation is off by default, and retirement moves nothing. One evaluation as a vertical path. Evaluation always runs retirement, which leaves capital in place. A switch, off by default, gates planning, the RiskModule, the on-chain contract checks, and the move.Reallocation is off by default, and retirement moves nothing. One evaluation as a vertical path. Evaluation always runs retirement, which leaves capital in place. A switch, off by default, gates planning, the RiskModule, the on-chain contract checks, and the move.

One evaluation: fitness and retirement always run and retirement moves nothing; a switch, off by default, gates the plan, the RiskModule check, the on-chain contract checks, and the move. Constants (cap 0.2, floor 0.06, every 20 ticks) are proposed defaults. Source: ARCHITECTURE.md section 5.

​
The owner’s policy

Reallocation runs among an owner’s agents only while that owner has signed a policy naming at least two of them. The policy can tighten the protocol bounds, a lower cap or a lower ceiling, and can never loosen them or lower the floor. The owner can revoke it at any time. An owner’s total capital across their agents is unchanged by any reallocation, but an individual agent can fall below its own principal. That is the property the owner consents to when signing, and the product must state it in those terms.

​
What the chain checks and what it cannot

The contract checks on-chain, at execution: same owner, the owner’s policy in force and naming both agents, the cap, the floor, the ceiling, and one debit per source per evaluation. The ranking itself is computed off-chain from fitness data and cannot be verified on-chain. A corrupted ranking can therefore misdirect capital only among one owner’s agents and within those bounds (THREAT-MODEL.md section 8).

​
Retirement

An agent retires when its capital falls below the floor (0.06, proposed), when it sits in the bottom decile of its niche for K consecutive evaluations (K = 3, proposed), or when its owner retires it. Retirement stops ticking, never deletes identity or history, and never moves capital: the balance stays in the agent’s account, recoverable by the owner. A retired agent is never a source or sink.

​
What is deliberately excluded

Genome replacement, mutation of live agents, recycling a retired agent’s capital into a sibling, higher reallocation intensity, and a treasury were all considered. The first two are excluded because their value in simulation came from stopping a failing genome, which retirement also does without copying anyone’s genome. The others are recorded as hypotheses with the evidence that did not meet the bar (EVOLUTION.md section 3.2).

​
What it is, plainly

This is selection on capital without variation. It pays off only if fitness rank persists, meaning an agent ranked high now tends to earn more in the next window. That requires a genome that carries a real, persistent edge. Without one, reallocation moves capital on noise. See Results for what was measured. Owners with one agent get no benefit from it.
Last modified on October 13, 2026