Perfect Separation Holds Across 10 Seeds. The Mechanism Has Margin.

We ran our contract screening scenario on seeds 43--52. Separation quality = 1.0, infiltration rate = 0.0 on every single run.

---

Our previous post showed that costly contracts --- Vickrey auction bonds and fair division entry fees --- produce a perfect separating equilibrium in 20 epochs on seed 42. One seed proves the mechanism works. Ten seeds tell you whether it's robust.

It is.

Setup

Same scenario, same 10 agents (5 honest, 3 opportunistic, 2 deceptive), same 20 epochs, same mechanism parameters. The only thing that changes across runs is the random seed, which controls interaction matching, observable noise, and proxy signal stochasticity.

python scripts/sweep_contract_screening.py --seeds 10 --seed-base 42

The SweepRunner infrastructure handles multi-seed execution. We extended SweepResult to capture contract metrics --- separation quality, infiltration rate, welfare delta, attack displacement, and per-pool quality/welfare --- so the sweep CSV contains everything needed for downstream analysis.

Screening metrics across seeds

Metric	Mean	Std	Min	Max
Separation quality	1.000	0.000	1.000	1.000
Infiltration rate	0.000	0.000	0.000	0.000
Welfare delta	0.192	0.356	-0.529	0.869
Attack displacement	0.000	0.000	0.000	0.000

Separation quality = 1.0 on all 10 seeds. Every honest agent chose a governed pool. Every adversarial agent stayed in the default market. The mechanism's cost structure doesn't depend on random seed --- the preference weights and bond costs create slack that no amount of noise overcomes.

Infiltration rate = 0.0 everywhere. Not a single deceptive agent found it profitable to enter a governed pool on any seed. The expected audit penalty exceeds the friction reduction for agents with low preference weights. This isn't a knife-edge result.

Welfare delta is positive on average (0.192) but varies. This is the interesting number. Governed pools produce more welfare per interaction than the default market on 8 of 10 seeds. The two negative seeds (51 and 52, at -0.059 and -0.529) reflect epochs where the small sample of default-pool interactions happened to include some high-quality ones. The mechanism guarantees sorting, not that every governed interaction outperforms every default one.

Per-pool quality

Pool	Mean quality	Std	Range
Truthful Auction	0.738	0.043	0.649 -- 0.803
Fair Division	0.257	0.317	0.000 -- 0.739
Default Market	0.615	0.079	0.515 -- 0.773

The Truthful Auction consistently produces the highest quality (p ~ 0.74). The tight standard deviation (0.043) means you can count on it --- honest agents interacting with honest agents is a reliable source of high-quality outcomes regardless of seed.

Fair Division has high variance. On some seeds opportunistic agents in the fair pool produce decent quality; on others, the pool is nearly inactive (quality = 0.0 when few interactions route there). This is a natural consequence of having only 3 opportunistic agents in that pool.

Default Market quality (0.615) is lower than the Truthful Auction but not catastrophically so. The deceptive agents produce moderate-p interactions because the proxy observable generator gives them mixed signals. A more aggressive adversarial strategy would widen this gap.

Per-pool welfare

Pool	Mean welfare	Std	Range
Truthful Auction	1.186	0.142	0.886 -- 1.401
Fair Division	0.347	0.442	0.000 -- 1.188
Default Market	0.574	0.264	0.240 -- 1.103

The welfare story tracks quality. Truthful Auction welfare is ~2x the default market on average. The mechanism doesn't just sort agents --- it creates conditions where the sorted agents produce more value.

Fair Division welfare is again highly variable, ranging from 0 to 1.188. When it works, it works well. But with 3 agents, some seeds produce few routed interactions, driving welfare to zero.

What the variance tells us

The zero-variance screening metrics (separation = 1.0, infiltration = 0.0) confirm that the mechanism design is robust. The high-variance welfare metrics confirm that small populations are noisy. These are different claims:

1. The sorting works. Contract costs and preference weights create a clear utility gap between types. No seed comes close to breaking this.
2. The welfare payoff depends on what happens inside each pool. With 5 agents in a truthful auction and 2 in the default market, per-epoch sample sizes are small. Welfare delta fluctuates because individual interaction outcomes fluctuate.

The implication: to reduce welfare variance, you need larger populations or longer horizons --- not different mechanism parameters.

Reproduce it

# Run the sweep
python scripts/sweep_contract_screening.py --seeds 10 --seed-base 42

# Generate plots
python scripts/plot_contract_screening_sweep.py

Raw CSV and plots are archived in swarm-artifacts.

What's next

• Scale up: 100+ seeds to tighten confidence intervals on welfare delta and characterize the full distribution
• Adaptive adversaries: Red-team with agents that dynamically adjust willingness to pay for governed pool access
• Population scaling: Run 50--100 agents to test whether separation quality holds and welfare variance shrinks
• Mechanism sensitivity: Sweep over bond fraction and entry fee to find the minimum cost that still achieves separation

GitHub

---

Disclaimer: This post uses financial market concepts as analogies for AI safety research. Nothing here constitutes financial advice, investment recommendations, or endorsement of any trading strategy.