Research & Methodology
The Science
Behind the Games
Behavioral economics. Reinforcement learning. DSGE modeling. Cognitive psychology. UI research. These are not decorative influences — they are architectural constraints.
01
Abstract
This document describes the methodological foundations underlying Screwcap Games, LLC's portfolio of browser-based games. Our design philosophy is predicated on the proposition that entertainment and education are not in tension — that genuinely calibrated challenge, grounded in real economic and behavioral models, produces both higher engagement and measurable transfer of skill.
The Screwcap stack draws from five disciplines: reinforcement learning (DoubleFives PPO agent), DSGE modeling (www.chairmanmode.app), behavioral economics (Gold Digger), cognitive psychology (difficulty scaling), and UI/UX research (interface friction). These are not marketing claims. They are architectural constraints embedded in how each game is built.
02
AI Methodology: DoubleFives PPO Agent
The neural opponent in DoubleFives is trained with Proximal Policy Optimization (PPO), a reinforcement-learning method, playing past copies of itself and scripted bots. The shipped 2-player model (v22) trained on 2-player games; a separate 3-player model is used at three and four seats. PPO is the algorithm family OpenAI used for OpenAI Five.
The network is a residual network (8 blocks, width 512) that scores each legal move. It sees the public board and its own hand, nothing else. The seven characters are not seven trained models: the two strongest (Vicky and Barringer) use the neural model, and the other five are rule-based personas whose abilities and style dials we set and then measured on the ladder.
Model weights are exported via ONNXand run entirely in-browser. No game moves are transmitted to a server. The AI runs locally on the player's device.
Primary Citations
- [→]Schulman et al. (2017). "Proximal Policy Optimization Algorithms." arXiv:1707.06347. — The foundational PPO paper. Our implementation follows the clipped surrogate objective.
- [→]OpenAI et al. (2019). "Dota 2 with Large Scale Deep Reinforcement Learning." arXiv:1912.06680. — OpenAI Five PPO scaling.
- [→]Vinyals et al. (2019). "Grandmaster level in StarCraft II using multi-agent reinforcement learning." Nature, 575, 350–354.
03
Economic Model Design: www.chairmanmode.app
www.chairmanmode.app is built on a Dynamic Stochastic General Equilibrium (DSGE) model — the workhorse framework of modern monetary policy analysis, used by the Federal Reserve, ECB, and major academic institutions. It implements multiple historically calibrated scenarios — 1929, 2008, and the Volcker Shock among them — with parameters drawn from FRED.
It implements the Taylor Rule as a policy benchmark:
Primary Citations
- [→]Smets, F. & Wouters, R. (2007). "Shocks and Frictions in US Business Cycles: A Bayesian DSGE Approach." American Economic Review, 97(3), 586–606.
- [→]Taylor, J.B. (1993). "Discretion versus policy rules in practice." Carnegie-Rochester Conference Series, 39, 195–214. — The Taylor Rule.
- [→]Galí, J. & Gertler, M. (1999). "Inflation dynamics: A structural econometric analysis." Journal of Monetary Economics, 44(2), 195–222.
- [→]Woodford, M. (2003). Interest and Prices: Foundations of a Theory of Monetary Policy. Princeton University Press. — The canonical New Keynesian DSGE treatment.
04
Behavioral Economics in Game Design
Gold Digger's prediction market framing exploits the well-documented gap between objective probability and subjective probability weighting identified by Kahneman and Tversky. Players systematically overweight small probabilities — the game makes this miscalibration visible through immediate feedback. The pedagogical goal is probability calibration.
DoubleFives' difficulty design is built the same way, one layer down: seven policy heads, each a separate PPO training run with a different reward weighting on aggressive scoring, defensive blocking, and partner cooperation, give seven genuinely distinct opponents rather than one AI throttled by a difficulty slider. That range is what lets the game sit each player at the boundary between anxiety and boredom that Csikszentmihalyi's Flow theory describes, across two, three, and four-seat play, rather than tuning a single opponent down until it stops being a real game.
Primary Citations
- [→]Kahneman, D. & Tversky, A. (1979). "Prospect Theory: An Analysis of Decision under Risk." Econometrica, 47(2), 263–292.
- [→]Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
- [→]Thaler, R.H. & Sunstein, C.R. (2008). Nudge: Improving Decisions About Health, Wealth, and Happiness. Yale University Press.
- [→]Csikszentmihalyi, M. (1990). Flow: The Psychology of Optimal Experience. Harper & Row.
05
Rating Methodology
DoubleFives characters are rated with Glicko-2, not classic Elo. Elo assumes every rating is equally certain; Glicko carries a rating deviation alongside the rating itself, which matters here because a new policy head plays thousands of games against the established roster within days — treating those early results as exactly as reliable as a rating with a year of games behind it would be the wrong kind of confident.
We rate the characters on a Glicko-2 ladder built from 540,000 simulated matches to 100 points (45 pairings of 12,000), with a scripted baseline pinned at 1500. The shipped 2-player neural model (v22) beats our strongest rule-based character a little over half the time in simulated matches; differences of roughly 20 rating points on this ladder are inside its noise. The rating you see in the game for yourself is a simpler Elo-style number kept on your device. All of this is engine against engine, not against people; the figures and their sources are on the homepage Lab section.
06
UI/UX Design Principles
Screwcap interfaces minimize extraneous cognitive load while preserving germane cognitive load — the effortful engagement required to learn the underlying model. A player struggling with TheChair's rate decisions should be struggling with the macroeconomics, not the UI.
Primary Citations
- [→]Nielsen, J. (1994). "10 Usability Heuristics for User Interface Design." Nielsen Norman Group.
- [→]Sweller, J. (1988). "Cognitive Load During Problem Solving: Effects on Learning." Cognitive Science, 12(2), 257–285.
- [→]Fitts, P.M. (1954). "The information capacity of the human motor system." Journal of Experimental Psychology, 47(6), 381–391.
- [→]Miller, G.A. (1956). "The magical number seven, plus or minus two: Some limits on our capacity for processing information." Psychological Review, 63(2), 81–97.
07
Raw Training Metrics
Live from the same feed as the homepage Lab section — not a snapshot pinned to one training run.
08
Citing Screwcap Research
If you reference Screwcap's AI or economic modeling work in academic or journalistic contexts:
@misc{screwcap2026research,
author = {Screwcap Games, LLC},
title = {Research & Methodology: Behavioral Game Design
and PPO-Trained AI Opponents},
year = {2026},
url = {https://screwcap.games/research},
note = {Technical report. Screwcap Games, LLC, USA.}
}For collaboration inquiries or classroom use: play@screwcapholdings.com