Synthetic Environment · RQ4 · Unity (C#)
Prison Sandbox
Closed Social System · Information Scarcity · Consequence Permanence
Research Instrument Context
Prison Sandbox is not documented here as a game pitch. In the doctoral project it is the synthetic world for RQ4: a closed prison ecosystem with a finite, non-replenishing population, hidden per-character reliability, and structurally irreversible consequence. Rather than rebuilding the whole live system for simulated inhabitants, the final phase reimplements only the component most likely to depend on the inhabitants being real, the formation of a per-character read, and asks whether season accounts built on it hold up under the same criteria as the live study.
Role in the Doctoral Project — The RQ4 Diagnostic
A Diagnostic, Not an Expansion
If reads formed against simulated characters support a season account that holds up under the RQ3 criteria (event selection, causal framing, significance attribution, and accuracy against the record), the pattern generalises. If they do not, the failure identifies what real stakes were supplying, and that is itself the finding.
An Experiment the Live Site Cannot Run
If an account’s judgments follow from the position of whoever formed them, several chroniclers with different histories toward the same characters should give different accounts of one record, disagreeing about who was responsible for what. Circulating rival histories of real people’s season is not defensible; with simulated inhabitants the same test carries no such cost.
The Run Record Holds Reasons, Not Just Events
Every system that resolves off-screen writes to a permanent run record: actions, rumour propagation and whether it was true, deaths and the official version, and each character’s short-lived wants (whims) with how they ended. A log of events supports a summary. A record that also holds wants, grievances and debts supports an explanation.
Termination, Resolution, Ending
Termination is mechanical (escape, transfer or death). Resolution does not happen; the gangs are still at war the day after the player goes over the fence. The ending is the chronicle written from the record afterwards. The same separation frames the live study, where a sixty-day clock ends a season that does not resolve.
Why This Environment — Design Rationale
Closed Population — Variable Isolation
A fixed cast of fifty prisoners, eight guards and five staff, with a limited reserve pool. Once the reserve is exhausted, nobody replaces the dead, and a depleted prison can collapse. Because the population cannot be topped up, the social network stays measurable and its evolution can be attributed to what agents did rather than to population flux.
Escape Objective — Information-Seeking Behavior
The player's only goal is escape. No authored path exists. This naturally produces information-seeking behavior: the player must talk to agents, evaluate their reliability, and build partial models of a social system they cannot fully observe. Information reliability becomes a design mechanic rather than a background parameter.
High-Stakes Irreversibility — Consequence as Mechanic
Killing an NPC permanently removes their narrative thread. Choosing one faction closes routes to others. There is no save-and-reload. Consequence permanence becomes a felt experience rather than a configurable parameter, so any account of a run is an account of losses that actually stuck.
Hidden Agendas — Reliability Stratification
Each agent carries private information states and motivations invisible to the player. Information shared by one agent may conflict with information shared by another for reasons that are agentically coherent. This models the strategic information distortion — lying, selective disclosure, exaggeration — that makes a record worth narrating.
Design Variable Mapping
| Closed inmate population | A fixed, non-replenishing cast lets social network structure be measured precisely. Population closure is a methodological requirement, not a narrative convenience. | Closure |
| Faction-stratified information access | Inmates, guards, and faction leaders have non-overlapping information states. Information reliability varies systematically by social position — modeling the trust-stratified propagation the framework formalises. | Reliability · Trust |
| Permanent NPC death | Irreversible removal of an agent and their associated narrative thread. The loss is structurally significant: quests, alliances and information that lived with that character are gone for the run. | Permanence |
| Contradictory escape routes | Multiple agents offer structurally incompatible information about escape paths. The player cannot validate information externally. This operationalises information reliability variance as the primary navigation constraint. | Reliability |
| Faction allegiance tracking | Choices made with one faction permanently alter access to others. Consequence permanence shapes the social state space that remains available. | Permanence |
Designed Agent Population — Social Variable Carriers
The Informant
Reliability: Stratified
Trades information selectively. Accurate about some domains, systematically misleading about others. Models partial information reliability — the agent architecture's core construct.
The Guard Captain
Reliability: Corrupted
Holds accurate system information but withholds or distorts based on self-interest. Social position gives them credibility their information reliability does not warrant.
Gang Leaders (4)
Reliability: Group-bounded
Accurate within faction knowledge, misleading about rival factions. Trust propagates through group membership.
Long-term Inmates
Reliability: Rumor-based
Hold degraded second-hand information. Models how information reliability decays through social propagation chains — the structural decay the Watcher measures.
Implementation Status
The full game design is specified and a prototype is in development; a seeded agent simulation of its information layer already runs (see Agent Simulation). In the doctoral plan, the RQ4 port of the per-character read takes place in Year 4, after the live-site questions have been answered.