03 / Method & limits
How the evidence was selected.
This is a small, purposive sample of documented mechanisms. The unit is a mechanism described by a linked source, coded into one primary domain. It is not a representative sample of players, runs, or every feature of a game.
Draft status: documentary research and preliminary coding were assisted by AI. The student’s review of the classifications and argument is still pending. No in-game replication or personal play diary is claimed.
Resets means the specified state is removed at the stated boundary. Persists means that state is carried over. Mixed / conditional means the source describes exceptions or a condition. Unknown means the evidence does not establish the answer.
Read the project statement (draft)
What Survives a Run? asks how cross-run information and persistent states change the choices and narrative access available after a reset. It compares Hades, Returnal, and Slay the Spire, with Outer Wilds explicitly included as a non-roguelike time-loop contrast.
The atlas uses fourteen documented mechanisms rather than invented player histories. The search prioritized developer articles, official interviews and patch notes, then used gameplay guides where official descriptions lacked detail. Each record links evidence and preserves its trigger, consequence, coding and limitation. Visitors can filter the sample, compare its matrix and inspect sources.
A central difficulty was separating documented gates from unsupported claims about persistence or character memory. The selected Returnal source links Xenoglyph translation to language samples, but does not establish retention after death. That record remains unknown. One Hades character's response cannot demonstrate universal NPC memory.
A separate fictional demonstration has two independently chosen instances: an archive and a corridor. Each ends in escape or death before another can begin. A failure can reveal information that changes a later encounter, while the Warden meets the player anew. A clue from each instance unlocks an additional central-door ending. These characters, decisions and outcomes are authored, excluded from the documentary dataset, and provide no evidence about actual games.
Research and preliminary coding were AI-assisted. The sample leaves out mechanics, platform differences and player interpretations. It describes selected rules, not narrative depth or causal effects. Empty cells indicate missing sampled evidence rather than absent mechanics.
Student review and the course's Claude data-discovery step remain to be completed or verified.
Read the coding rules
- Player knowledge
- Information a player can learn and use, distinguished from an in-game recorded flag.
- Narrative state
- Dialogue, recognition, story or relationship state described across attempts.
- System unlocks
- Lasting access to content, abilities or progression systems.
- Run loadout
- The temporary resources and equipment held in a particular attempt.
A record has one primary domain for readability; the original mechanism may span several. “Not applicable” is different from “unknown.” A source describing one NPC cannot establish what every NPC remembers.
Limits of this comparison
Sources differ in purpose, detail, date and game version. A developer explanation and a guide do not offer the same type of evidence. This atlas describes documented rules; it does not measure subjective agency or prove that a mechanism causes a better story. The sample omits many games, scenes and exceptions.
Outer Wilds, when included, is a time-loop comparison rather than a roguelike. The fictional model is separate from the coded dataset and cannot validate real-game claims.
Conceptual lens: Brent Van Mol, “Looping Across Genres” (2026), an extended abstract on cyclical narrative structures. It supplies a framing question, not this atlas’s empirical dataset.
