Fairness & Collusion Prevention
March 2026
In Minority Rules, every participant in a market selects one of three cards each round. The card chosen by the fewest participants — the minority card — wins. Winners split the total stakes in proportion to their stake, minus a small fee. If two or more cards tie for the fewest selections, there is no minority and insurance payouts apply instead.
This creates a strategic dynamic unlike traditional prediction markets. Your payout depends not on a random outcome, but on what everyone else decides to do — and they're all trying to be in the minority too.
The probability of winning in Minority Rules can be computed exactly using combinatorial mathematics. We've calculated these probabilities for market sizes from 20 to 1000+ participants. As the number of participants grows, the probability of winning approaches 1/3 (33.3%), the loss probability approaches 2/3 (66.7%), and no-minority outcomes become increasingly rare.
Winners split the total stakes in proportion to their individual stake relative to the total stakes of all winners, minus a fee. This means a participant's return depends not only on whether they win, but on how their stake compares to other winners — a critical factor in analyzing collusion profitability.
Collusion occurs when two or more participants secretly coordinate their card selections to gain an advantage over others. The most effective collusion strategy is card coverage: a group of three participants each picks a different card, guaranteeing that one of them is always on the minority card (when a minority exists).
In a smaller market, a coordinated group of three covering all cards guarantees one member wins nearly every round — a win rate approaching 100% for the group. This can give the group an edge over independent participants, though that edge diminishes rapidly as the market grows.
We quantify collusion advantage using a metric we call the colluder edge: the difference in return on stakes between the colluding group and the average non-colluding participant, measured in percentage points. A positive edge means colluders earn a higher return on their staked capital than honest participants do.
Our simulations — covering tens of thousands of scenarios across varying market sizes, fee levels, and group configurations — confirm a clear and consistent pattern: the colluder edge diminishes rapidly as market size grows. In smaller markets, a coordinated group's guaranteed-winner advantage translates to a meaningful financial edge. In larger markets, that advantage is diluted: the colluders become a smaller fraction of the market, their guaranteed winner must share the total stakes with an increasing number of non-colluders who independently selected the same card, and the fee erodes what remains.
At large market sizes, the mathematical advantage becomes so small that it is difficult to distinguish from normal statistical variance — making collusion a high-risk, low-reward strategy.
The structural defense: The larger the market, the less any small group can influence outcomes. This is not a design choice we impose — it is a mathematical property inherent to the market's structure. Minority Rules naturally protects its largest markets through the same probability theory that governs the market itself.
A group of 3 colluders in a 20-participant market represents 15% of all participants — enough to have some influence on which card becomes the minority. The same 3 colluders in a 100-participant market are just 3% — their influence on the card distribution is negligible. At 500 or 1000 participants, it's virtually zero.
Our simulations confirm that adding more colluders doesn't solve this. Scaling from 3 to 9 colluders in a larger market does not recover the edge that a small group has in a smaller market. The additional colluders compete with each other for the winners' share of the total stakes, diluting their own returns.
Every round, the fee reduces the total stakes. For collusion to be profitable, the group's coordinated advantage must exceed the cumulative fee on their stakes. In small markets, this is easily achieved. In large markets, the advantage shrinks below the fee threshold, making collusion a net-negative strategy.
Smart colluders might try to evade detection by only coordinating some of the time — playing normally most rounds and colluding occasionally. Our analysis shows this dramatically reduces profitability: at 10% activation (colluding 1 in 10 rounds), the edge drops to just 2-5 percentage points. The mathematical reward for occasional collusion is barely distinguishable from normal variance.
Despite the market's natural structural defenses, we take collusion seriously. We've built a multi-layered, real-time detection system that monitors every round of play.
Our detection system is grounded in hypothesis testing. For each participant, we model the null hypothesis: "this participant's outcomes are consistent with independent play." We then measure how far their actual behavior deviates from what this null hypothesis predicts, using the exact probability model described in Section 2 as the expected baseline.
A key feature of our approach is that expected values are not static — they are computed dynamically based on the actual market sizes each participant played in. A participant who played 50 rounds at 30 participants and 50 rounds at 200 participants has a different expected profile than one who played 100 rounds all at the same size. We weight accordingly.
If two participants are acting independently, their card selections should be uncorrelated. For any pair choosing among 3 cards, the probability of both selecting the same card in a given round is 1/3. Over r shared rounds, the number of same-card matches follows a binomial distribution B(r, 1/3).
A coordinated pair covering different cards would produce a same-card rate near zero — a deviation so extreme that even over a modest number of shared rounds, the probability of this occurring by chance is vanishingly small. We use the distance between observed and expected rates, normalized by the standard deviation, to quantify the strength of evidence.
Beyond pairwise analysis, we test whether groups of participants exhibit collective behavior inconsistent with independence. For g independent participants, the probability that at least one wins in a given round is:
A card-coverage collusion group of 3 would produce a "group win rate" approaching 100%, while three independent participants would produce a rate determined by the formula above — typically 30-60% depending on market size. The gap between these rates provides a powerful detection signal that does not degrade with market size, because it measures the structural guarantee of coverage rather than any individual's win rate.
Colluders who activate their strategy intermittently produce a weaker signal per round but still generate a systematic deviation from expected behavior over time. Our system uses cumulative statistical methods — analogous to techniques used in industrial quality control for detecting small persistent shifts in manufacturing processes — that accumulate evidence across rounds without requiring any single round to be individually suspicious.
These methods are specifically designed to resist "innocence banking" — a colluder cannot offset future suspicious behavior by behaving normally for many rounds beforehand. The accumulated evidence only grows; it does not decay.
Our detection uses multiple independent analytical methods, each targeting a different mathematical property of collusion. When several methods independently identify the same participants, our confidence increases multiplicatively — not additively. This Bayesian convergence of independent evidence streams is what allows us to achieve both high detection rates and zero false positives simultaneously.
We don't just theorize about detection — we stress-test it adversarially. We've built a headless simulation engine that directly executes the market's payout mathematics (the same code that runs in production) and models colluders executing specific strategies against populations of independent participants.
Our simulation matrix covers:
For each scenario, we measure both the colluder edge (whether the strategy is actually profitable) and the detection outcome (whether our system catches it). We also measure false positive rates by running the same detection methods against the non-colluding participants within each scenario.
Result: Our detection system identifies 100% of profitable collusion strategies across all tested scenarios. Every strategy that produces a meaningful advantage over honest participants is caught — at every market size, every fee level, and every activation rate we've tested, including colluders who activate their strategy only 10% of the time.
Equally important: our flag-generating detection methods produce zero false positives across all 13,000+ test scenarios. Honest participants are never wrongly accused.
This is achieved through two mechanisms. First, the statistical tests we employ require extreme deviations from expected behavior — deviations so large that they fall outside the range of what random variance can plausibly produce. Second, our multi-signal convergence approach requires independent lines of evidence to agree before escalating to enforcement. The probability that multiple independent statistical tests all produce false positives for the same participant pair is the product of their individual false positive rates — a vanishingly small number.
When our system detects collusion with high confidence, it takes automatic action:
We believe in proportional response — a first offense could result from a misunderstanding or a one-time lapse in judgment. But repeated collusion demonstrates intent, and we act decisively to protect the integrity of the market for all participants.
Our detection system is not static. We employ an automated red team / blue team adversarial framework powered by AI to continuously stress-test and improve our defenses.
We run an AI-powered adversarial agent whose sole objective is to find ways to collude profitably without being detected. This agent is given the same information a real adversary would have — publicly known market rules, existing strategy results, and the conceptual description of our detection methodology — but no access to internal thresholds or implementation details. It is instructed to be as creative and ruthless as possible: proposing novel strategies, identifying potential blind spots, and reasoning about what assumptions the detection system might be making that could be exploited.
A separate AI-powered defensive agent then receives the red team's complete, unfiltered analysis — every proposed strategy, every weakness assessment, every line of reasoning — along with full access to the actual detection system's internals. It traces each proposed strategy through every detection layer, evaluates whether it would be caught, identifies any genuine gaps, and recommends specific threshold adjustments or new detection heuristics where needed.
The adversarial agent does not know its output will be reviewed by a defensive agent. This creates a genuine information asymmetry: the red team holds nothing back because it believes its analysis is confidential, while the blue team gets full intelligence on the adversary's thinking. The result is an honest, thorough adversarial assessment followed by an equally thorough defensive evaluation — a continuous improvement loop that operates faster and more creatively than manual review alone.
This framework complements our existing approach of:
This layered adversarial approach — where we continuously challenge our own defenses with the most creative attacks we can generate, then systematically close any gaps identified — ensures that our system evolves faster than any potential collusion strategy.
Minority Rules is a prediction market built on strategy, intuition, and human psychology — not a system that rewards secret coordination. Every participant should have a fair opportunity to win based on their own judgment.
We invest significantly in mathematical analysis, simulation testing, and real-time monitoring to uphold this principle. If you ever have concerns about fair play, please contact us at [email protected].
Back to Minority Rules