Created by: Mathias Becerra, Divyesh Khatri, Yixun, Yan Chen, Xin Yue
Artist Statement
It is Week 5 at Stanford. Your CS project partner just dropped out to found an AI startup. Your roommate has “AI Safety Researcher” in their LinkedIn headline, but somehow also interned at three frontier labs this summer (?) . A freshman just raised a seed round before finishing CS106A, and every other conversation at On Call Cafe begins with, “So… what are you building?” Meanwhile, someone just announced a model that is “definitely” going to change the world.
Again.
Who can you trust when the people around you are doing their utmost to help humanity? However, what happens when the effort to promote humanity looks exactly like an attempt to destroy it?
404: ALIGNMENT NOT FOUND is a social deduction game for four to eight players revolving around the theme of ethical management of AI. While in your company, there can be one or more individuals whose goal is the implementation of the AI revolution, all the other participants are on the right side of the argument.
In each round, players of the game secretly choose between the TRAIN and SAFE options which influence their levels of credit. Subsequently, the meeting resembles an academic discussion. But unfortunately for the players, not everyone can survive the voting rounds, which results in someone being fired.
The main premise of our game revolves around a question that is becoming more challenging to answer both at Stanford and elsewhere: What do we do in situations when doing the right thing means ruining our résumé?
The game mechanics are built around this tension. Achieving credit makes you successful, yet it raises suspicions. Playing it safe is a way of helping humanity, yet at the same time, it makes someone very wealthy. Even if the lab does not succeed, one always wins out in the end. After all, intentions rarely matter when it comes to incentives.
Our game is about the irony of everyday contradictions that surround generative AI. We make use of language models to write our emails, analyze our code, summarize our readings, and even help us think sometimes. Yet still, we (as a team) wonder during our ideation process if these are evolving too fast.
404: ALIGNMENT NOT FOUND shows what happens when convenience, ambition, and responsibility clash together, combining fellowship, challenge, social deduction, and the thrill of deception.
Because sometimes the biggest AI alignment problem…is the humans who deliberate the capabilities of these LLMs.
Concept Map & Ideation Process
Our team aimed to create a social deduction game that focuses on conversation and strategic thinking, but also one that provides a unique currency mechanism in contrast to existing games like Werewolf and Mafia (where tokens are usually not that common). Through prototyping and analyzing different games, like Hues & Cues (Mathias’ personal experience) , We’re not Strangers (that focused on fellowship – getting to know each player) Secret Hitler (social deduction), we discovered that we wanted to have deception and mystery as our official game components. During Week 1, Mathias proposed a prototype that tested the synergy between subjective color guessing (Hues & Cues) and hidden roles. He also had the opportunity to play Secret Hitler during Week 2’s Game Night, which made our team more aware of how voting through secret identities contributes to exciting cycles of deduction.
Testing various games (1) We’re not Strangers for our session. 2) Prototyping ideas for games (Mathias’ initial game ideas), 3) Playing Secret Hitler (Game Night 1)
Discovering our game’s kinds of fun (Fellowship, Challenge, Discovery)
The insights learned in the process were the basis for the development of two prototypes.
- Yixun prototyped a game that contained a currency system rewarding strategic resource management, whereas Divyesh designed a game inspired in “We’re doomed” with hidden objectives based on Mathias’s proposal in Tuesday’s class in Week 2 on implementing the theme of “corporate espionage and the rapid advancement of LLMs training.”
- After playtesting both prototypes, we realized that the game was getting too complicated if we were to combine the mechanics of currency and hidden roles/news. So we decided to work on the hidden-objective mechanism and eliminate the currency-based systems (also taking Butch’s feedback into account)
MAPS AND BRAINSTORMING
After playtesting in our first group meeting, we decided to base the direction of our game on the hidden objective rather than the currency mechanic. (Hand-drawn map created by Yixun with Procreate)
MDA – based on Yixin’s initial brainstorming and experience in creating mind maps
MDA breakdown—Defining mechanics, dynamics, and aesthetics for 404: Alignment Not Found. (Before official in-class playtesting with other groups) – (Hand-drawn map created by Yixun with Procreate)
Initial Decisions
Prototype Directions
Yixun’s Currency-Based Version
Yinxu proposed a version of the game built around a currency system only, where player actions were tied to earning and spending an in-game resource.
Early sketches, and prototype notes for the currency-based version plus playtesting (Mathias, Yixun, and Divyesh)
Divyesh’s Target-Based Version
Divyesh proposed a target-based mechanic, drawing inspiration from the games We’re Doomed and Werewolf. This direction focused on players secretly aiming for specific outcomes rather than accumulating currency, and it introduced the core Train / Safe choice that appears in the following section (official rules). Mathias proposed the idea of the theme of “accelerating LLMs and aligning with good people. Kinds of fun targeted in this version: fellowship (social deduction) and challenge.
Elements: Targets — Safe & Train (representing the training of LLMs).
Early sketches for the target-based version, including the Train/Safe and Aligned/Accelerationist role split.
Early sketches and prototype notes for the target-based version. (Divyesh)
Game Shape
Our game primarily emphasizes fellowship and challenge, using expression as a supporting aesthetic rather than focusing on fantasy or narrative.
- Fellowship: The game revolves around social deduction, discussion, accusation, and voting. Every round encourages players to negotiate, persuade, and read one another, making interaction the core experience. Both Aligned researchers and accelerationists are expected to collaborate with their teammate to win over the other camp. With the addition of personal credits, players may even turn against teammates, choosing personal gain over the group’s success, which adds dynamic to cooperation.
- Challenge: Players constantly face strategic dilemmas. Accelerationists must hide their identities while training with convincing reasons; Aligned researchers should struggle between fighting for themselves or for the team, while firing out the forger among them. Whether to TRAIN or SAFE, when to prioritize personal credit over collective survival, and how to identify hidden accelerationists before Doom reaches a critical level.
- Expression: Players choose whether to bluff, cooperate, betray, or play selfishly, allowing different social play styles and moral decisions to emerge naturally.
Our mechanics are intentionally built to maximize Fellowship through player interaction and Challenge through strategic uncertainty. This design direction also explains why our iterations consistently removed extra systems: the simpler the mechanics became, the more the core social experience stood out.
Formal Elements
Formal elements summary — players, objectives, rules, procedures, and outcomes.
Iteration History
Where we started
Our first prototype was Pirate’s Gambit, a social deception party game for 6-12 players: hidden Pirates and Treasure Hunters, a moderator, night phases, a money system, a Witch Market for buying abilities, and a physical “Floor is Lava” mechanic. Playtesting it taught us two things. First, the hidden-role deception core produced the moments we liked most: accusations, conversions, reading people. Second, everything stacked on top of it (currency, missions, market, physical movement, a rotating moderator) made the game hard to explain and harder to run. Our instructor’s feedback said the same thing about our early direction: too many mechanics combined, and we should commit to one core aesthetic.
Iteration 1 — Theme pivot: pirates to AI labs
We re-themed to an AI research lab: everyone works at the same lab, hidden Accelerationists want the model to ship, everyone else wants to fire them before it does. We picked this theme for relevance (some of us have friends at AI labs and wanted to playtest with them) and because it gave the deduction game a shared resource with real dread attached: the Doom track. We also drew on We’re Doomed for the tragedy-of-the-commons framing, where personal greed (Credit from training) conflicts with collective survival (keeping Doom down). This pivot is also where we cut the currency, market, missions, and moderator: the new core loop (secret TRAIN/SAFE choice, public reveal, discussion, firing vote) needed none of them.
Iteration 2 — LLM simulation before human playtesting
July 2nd, First Playtesting
Because our theme is about AI risk, we thought it might be fun to refine the game based on LLM agent testing. We wanted to see the thinking decisions that agents made in order to see if it might surface how it might make players feel behaviorally in the background. We built a simulator where agents receive the ruleset and a secret role, discuss, and vote, and ran it in phases:
- Two Accelerationists: at 6 players with one Accelerationist, the aligned side won 100% of games (a lone traitor could not win); at 8 players with two, aligned wins dropped to 80% with longer games. We adopted the 2-Accelerationist rule for 7-8 players, but the simulation flagged that Doom never once reached 10.
- Race Pressure: Doom could not race because the aligned majority pinned it at 0. We iterated on automatic Doom escalation seven times; framing mattered enormously (a fatalistic rules framing produced a hedge stampede and 20% aligned wins, a containment framing produced an all-SAFE meta and 100%). We adopted a capped ramp (+1 early rounds, +2 later) after it produced the first solo-Accelerationist doom win.
- A 20-game reality check showed our adopted config was actually 40% aligned with 12 doom ships, driven by coordination traps (a FUNDING ROUND where all 8 players train ships the model instantly) and large herd swings. One card’s shuffle position was deciding games.
- Compute limits: capping the net Doom swing per training run cut doom ships from 12/20 to 1/10 and brought aligned wins to 70%. We adopted it. Adding mandatory voting on top killed games by round 3-4 in simulation, but vote forensics showed this was an artifact of LLM herding (74% round-1 pile-ons), not the rule; a Monte Carlo at human tie rates supported mandatory pointing. We printed mandatory voting for the tabletop and kept abstention as a variant.
One honest limitation: LLMs are strong readers, so they never flagged that our ruleset was too dense. That problem only surfaced with humans.
Iteration 3 — First human playtest (ourselves)
Two findings. First, we hated that the game ended the instant Doom hit 10: it made endings feel like accidents rather than climaxes. We changed the ship condition so that only a training run can end the game: Doom can overshoot 10 through Race Pressure or news, and the table gets one rescue round to pull it back before the next training run resolves. We expanded the Doom track past 10 to support the overshoot. This became our favorite moment in the game: the table at Doom 11, everyone deciding whether to save the world or grab one last Credit. Second, some news cards were incoherent on inspection: BREAKTHROUGH doubled a TRAIN’s Doom without doubling its Credit, which gave the Accelerationist no cover and the greedy no incentive. We rebalanced the news deck around a rule of thumb: every card should either tempt the aligned or pressure the table, ideally both.
We also had designed a new doom track that attempted to converge the Doom track as well as the round number into one track where the numbers represented both the doom level and the round number.
Iteration 4 — Class playtest: the rules were the problem
Class feedback converged hard on one point: the mechanics were fine but the ruleset was too dense to learn up front. Specific repeated items: rule explanation too long, win conditions unclear (is it reaching 10 or going over?), the compute limit confusing, the new doom track confusing, and the reference card too small to be useful. Players also told us the game got fun in the second half, once they had internalized the loop; our instructor feedback note said to focus on the first playthrough.
Changes we made in response:
-
- Made the game easier and simpler to understand: Removed the compute limit entirely. It was our most explanation-heavy rule, triggered rarely, and interrupted the reveal (the most exciting moment) with arithmetic. It also fought our new end condition, which depends on big swings being possible: a table of SAFEs rescuing Doom from 11, or a table of greed shipping the model, are the moments the game is about. The simulation had adopted the cap to stop coordination-trap blowouts, but the rescue-round rule now covers the same failure more elegantly: an overshoot is no longer instant death.
- Reverted back to the old doom track. It was the old Doom track’s information with a worse presentation.
- Ordered the news deck instead of shuffling it, so complexity reveals itself during play: early cards are pure Doom movement, modifiers arrive mid-game, and the strange cards (LEAK, HYPE CYCLE) appear after the table knows the loop. The up-front teach shrank to the core loop plus “the cards explain themselves.” Shuffling is now a variant for repeat players.
- Rewrote the rules to roughly half their previous length, stated the ship condition twice (it was the most misplayed rule), and moved everything a player touches during a round onto a larger per-player reference card.
- Added a discussion timer and script/rules for host to follow during discussion to prevent lags in the game.
- Rotating news drawer, so a different player flips and reads each round, keeping quieter players involved in the ritual of the game.
- Wrote out a script for the host to read, better guiding players through the game
-
Changes we made in response:
-
- Made the game easier and simpler to understand: Removed the compute limit entirely. It was our most explanation-heavy rule, triggered rarely, and interrupted the reveal (the most exciting moment) with arithmetic. It also fought our new end condition, which depends on big swings being possible: a table of SAFEs rescuing Doom from 11, or a table of greed shipping the model, are the moments the game is about. The simulation had adopted the cap to stop coordination-trap blowouts, but the rescue-round rule now covers the same failure more elegantly: an overshoot is no longer instant death.
- Reverted back to the old doom track. It was the old Doom track’s information with a worse presentation.
- Ordered the news deck instead of shuffling it, so complexity reveals itself during play: early cards are pure Doom movement, modifiers arrive mid-game, and the strange cards (LEAK, HYPE CYCLE) appear after the table knows the loop. The up-front teach shrank to the core loop plus “the cards explain themselves.” Shuffling is now a variant for repeat players.
- Rewrote the rules to roughly half their previous length, stated the ship condition twice (it was the most misplayed rule), and moved everything a player touches during a round onto a larger per-player reference card.
- Added a discussion timer and script/rules for host to follow during discussion to prevent lags in the game.
- Rotating news drawer, so a different player flips and reads each round, keeping quieter players involved in the ritual of the game.
- Wrote out a script for the host to read, better guiding players through the game
- Clarified Win Conditions.
Feedback we chose not to implement, and why:
- “Don’t reveal fired players’ roles.” We kept role reveals. Without them, the aligned side gets almost no confirmed information and the deduction collapses into guessing; the cost the feedback identifies (one revealed Accelerationist exposes their partner) is real, but it is also the intended punishment for getting caught, and the pair can play to avoid association.
- “The game isolates loud and quiet players.” This is a known weakness of the whole social deduction genre and we could not fully solve it. We partially addressed it structurally: votes are simultaneous pointing rather than open nomination (quiet players’ votes count equally), LEAK targets by group pointing rather than by argument, and the rotating news drawer gives every player a recurring speaking moment.
- A dominant-strategy concern was raised without specifics. Our simulation work suggests the closest thing to one (mass SAFE play) loses to Race Pressure, and mass TRAIN play ships the model; we believe the equilibrium forces mixing, but we flag it as an open balance question.
Open questions we are curious about but have not playtested:
- Personal news roles: news cards that assign a temporary role or power to a specific player for that round, making the news system less anonymous.
- Tunable greed: letting tables adjust the Credit win threshold themselves as a difficulty knob, since that one number controls how tempting betrayal is.
What we learned overall:
Our iteration pattern was mainly subtraction. Almost every version of this game got better by removing something (the market, the currency, the compute limit, the doom track, the shuffled deck) and the mechanics that survived (the Doom track, the secret simultaneous choice, the public Credit, the firing vote) are the ones that generate discussion by themselves.
-
Pre-Final
Playtest Video (Our group)
In-class Final Playtesting Video
https://www.youtube.com/watch?v=Bes5Rv8jODU
https://www.youtube.com/watch?v=NvzSBciNEi0&feature=youtu.be
Print-n-Play
The print-and-play version of our game can be found in the following link:
Interesting Moments from Playtesting Sessions:
- “Once someone explained it, everything clicked.”
- “The rules felt too long—there were just too many words but the little black cards helped a lot!”
- “I wasn’t sure exactly how you win. Is it 10 or over 10? “
- “The board and movement were confusing at first, but as we started playing more rounds and more , I started to get it!”
- “The game itself isn’t that complicated—the explanation just needs to be simpler!”
Box Art
Pre-final game set (before final playtesting)
Official game set (Mathias’ design)
AI Disclosure
Chapt was used to generate initial ideas for integrating tokens with the mechanism of deception (images, brainstorming, etc.), which were later discarded/extended/or replaced with better ideas.
