Conflux — About & FAQ
Conflux is an interactive lab for public-goods dynamics. It lets you simulate how incentives (tax, reward, punishment), network topology, and behavioral rules shape cooperation and welfare over time.
Quick Tour
- Home: play with parameters and watch curves evolve.
- Experiments: run grid sweeps and visualize outcomes with a heatmap.
- Compare Topologies: overlay well-mixed vs small-world for the same settings.
Model Overview
In each round, agents receive an endowment w and choose a contribution ci∈[0,w]. The sum of contributions is multiplied by a marginal per-capita return (a) and redistributed. Policies add incentives: taxes for under-contributing, rewards for generous contributors, and punishment for defectors. Noise captures execution jitter and imperfect observation.
Parameters & Implications
Core
| n | Population size. Larger n increases payoff dilution; cooperation is harder without mechanisms. |
| w | Per-round endowment. Upper bound on each contribution. |
| a (MPCR) | Marginal per-capita return. If a<1, pure egoists defect unless incentives counterbalance. |
| T | Number of rounds. Longer horizons allow beliefs to stabilize. |
Noise
| σ_exec | Execution noise: agents miss intended contribution (trembling-hand). Can accidentally nudge cooperation. |
| σ_obs | Observation noise: misperceive others’ contributions. Increases misclassification under punishment. |
Mechanisms
| Tax (on low c) | Activated when c<τ·w, levies rate·(w−c). Discourages free-riding. |
| τ (threshold) | Low contributions below τ are penalized. Higher τ is harsher. |
| Reward (for high c) | Activated when c≥ρ·w, grants rate·c. Encourages generosity. |
| ρ (threshold) | Higher ρ targets only top contributors; lower ρ spreads rewards widely. |
| Punishment | Expected fine on under-contributors. Effective but may reduce welfare if overused under noise. |
| fine | Penalty magnitude on defectors. |
| punishers share | Share of potential punishers in the population/neighborhood. |
Topology
| well-mixed | Everyone interacts with everyone. One global public good. |
| small-world | Watts–Strogatz neighbors + rewiring. Local public goods; clusters can sustain cooperation. |
| k | Initial ring degree (even). Higher k broadens neighborhoods. |
| p | Rewire probability. Larger p reduces path length; may mix behavior faster. |
Agent Types
- Egoist: best-response to current beliefs, accounting for mechanisms and risks.
- Conditional Cooperator: tends toward the believed average contribution.
- Punitive Cooperator: like CC, but benefits from punishing defectors (modeled as expected risk).
- Altruist: contributes high fraction regardless of beliefs (e.g., 0.8·w).
Common Patterns & Tips
- When a<1, incentives are needed for egoists to cooperate.
- Under high
σ_obs, generous punishment backfires due to misclassification. - Small-world topology can preserve cooperative clusters even if global cooperation is low.
- Rewards often lift average welfare, taxes/punishment can lift cooperation but may reduce welfare if too harsh.
FAQ
Is the punishment stage explicit?
For speed and clarity, Conflux uses an expected-penalty heuristic based on punisher share and thresholds. A full second-stage model is on the roadmap.
How are beliefs updated?
Agents track a scalar belief about average contributions and update with an adaptive learning rate that increases with prediction error.
Why two topologies?
Well-mixed approximates mean-field interaction; small-world captures local clustering and short paths—good proxies for many real networks.
Can I export results?
Yes—use CSV export on the Experiments page. For full reproducibility, save the config alongside results.