Social Experiment Methodology

The framing

CashPop is a Telegram game. It is also a continuously-running, planetary-scale, real-time experiment on human coordination, focal beliefs, strategic depth, and common-knowledge propagation. The protocol's commercial structure (advertising-funded, token-redistributing) is what makes the experiment financially sustainable. The experiment, in turn, is what makes the protocol an enduring scientific commons.

This page documents what we measure, how we anonymize, what we publish, and the academic partnerships involved.

Variables observed

For every Round, we log:

Variable	Type	Resolution
Round ID	integer	unique
Question ID	integer	maps to public reservoir entry
Round start timestamp	UTC ms	exact
Participant count	integer	exact
Commit count by phase ms-bucket	array	binned at 1-second intervals
Reveal count by phase ms-bucket	array	binned at 1-second intervals
Majority outcome		exact
Vote ratio	float	exact
Demographic decomposition	aggregate	binned at ≥100 participants per bin

We do NOT log:

• Individual user identity at any level
• Individual user vote (only aggregate ratios)
• IP addresses (only country code at >100 participant aggregation)
• Cross-Round behavior of any specific user (each user appears as an anonymous, non-cross-correlated token in each Round)

Anonymization protocol

All published data passes through a k-anonymity floor of k=100: no aggregation cell contains fewer than 100 participants. Cells below k are merged with neighboring cells or suppressed.

For demographic decompositions (country, age band, language), we apply differential privacy with ε = 1.0: a calibrated Laplace noise mechanism is added to each cell count.

Cryptographic proof: the anonymization pipeline is open-source (github.com/cashpop-protocol/data-pipeline), and a third-party verifier can verify that the published dataset is consistent with the anonymization rules.

What we publish

Quarterly aggregated datasets

Each quarter, we release:

1. Round-level aggregate: per-Round statistics (count, ratio, demographics) — k-anonymized.
2. Question reservoir snapshot: questions used, response distributions, focal-point estimates.
3. Time-series: weekly DAU, MAU, Round count, prize pool distribution.
4. Cross-country focal-point matrix: per question category, the cosine similarity of response distributions across countries.

All datasets: CC-BY-4.0 license. Published at datasets.cashpop.meme.

Annual research reports

Each year, the protocol publishes a peer-reviewed-quality research report. We have committed to:

• At least one quarterly paper preprint targeting SSRN / arXiv (econ.GT, cs.GT).
• At least one annual report co-authored with an academic partner (target: behavioral economics group at a research university).
• An open data review board including external academic reviewers.

The first such report — covering the Q1 2027 dataset — is targeted for publication Q2 2027.

Methodologies being explored

Active research lines (subject to peer review):

1. Cross-cultural Schelling-point divergence index. A metric for how culturally-specific focal points are, as a function of question type and cultural distance.
2. Level-k mixture estimation at population scale. Using Round-level commit-time and answer data to estimate the level-k distribution of a population.
3. Common-knowledge propagation latency. How fast do beliefs about beliefs propagate through Telegram social graphs? Measurable via cross-Round consistency for synchronized question batteries.
4. Wisdom-of-crowds calibration vs. ground truth. For factual questions, comparing CashPop majority to verified ground truth.
5. Strategic-depth heterogeneity by demographic stratum. Are seniors more level-1, are crypto-natives more level-3? Empirically testable.
6. Belief-revision under information shocks. Run question batteries before and after major news events; measure shift.

Why this matters

Three reasons:

1. Scientific. Beauty Contest experiments have been conducted with <200 subjects in controlled labs for 30 years. CashPop runs ~10,000 subjects per Round, in field conditions, across cultures, with cryptographic auditability. The empirical leverage is unprecedented.

2. Civic. Common knowledge structures (Aumann, Rubinstein, Vives) underpin everything from market microstructure to election outcomes. Better measurement of how common knowledge forms and dissolves is a public good.

3. Reputational. The protocol's long-term legitimacy depends on its data-commons identity. CashPop is not extracting from users; it is co-producing a scientific resource with them.

Limitations we acknowledge

• Selection bias. CashPop users are not a representative sample of humanity. They are Telegram users with crypto curiosity and time to play games. Our results generalize to this population, not to humanity at large.
• Incentive contamination. Players are paid (in POP). This changes the strategic game compared to no-incentive lab studies. We document and quantify this effect.
• Question reservoir bias. The reservoir is LLM-generated and debiased to a best effort, but residual biases exist. We publish the reservoir openly so external researchers can re-analyze with their own debiasing.

References

• Aumann, R.J. (1976). Agreeing to Disagree. Annals of Statistics 4(6).
• Rubinstein, A. (1989). The Electronic Mail Game. AER 79.
• Vives, X. (2008). Information and Learning in Markets. Princeton.
• Surowiecki, J. (2004). The Wisdom of Crowds. Doubleday.
• Galton, F. (1907). Vox Populi. Nature 75.