# LITERATURE MAP AND SOURCE CHECK
Icarus Online LLC | The Garden Research | Verified starting references, 19 October 2026

These works motivate controls and hypotheses. None proves that the Garden treatment produces persistent character. The artistic thinkers are intellectual influences; scientific claims below are supported only to the extent that each cited work actually studied them.

| Source | Link | Relevant lesson / limitation |
|---|---|---|
| Park et al. (2023), *Generative Agents: Interactive Simulacra of Human Behavior* (UIST) | https://doi.org/10.1145/3586183.3606763 | Agent memory, reflection, and planning can shape simulated long-term behavior. A different intervention and endpoint than this study. |
| Luz de Araujo et al. (2025), *Principled Personas* (EMNLP) | https://aclanthology.org/2025.emnlp-main.1364/ | Persona prompts are sensitive to irrelevant attributes; motivates neutral and explicit-persona controls. Does not show atmospheric art creates character. |
| Choi et al. (2024), *Examining Identity Drift in Conversations of LLM Agents* | https://arxiv.org/abs/2412.00804 | Assigned persona is not necessarily stable across conversations; long-session measurement matters. |
| Ashery, Aiello, Baronchelli (2025), *Emergent Social Conventions and Collective Bias in LLM Populations* (*Science Advances*) | https://doi.org/10.1126/sciadv.adu9368 | Collective conventions can arise in interacting populations; informs a separate group-level experiment. |
| Jones and Hauert (2026), *Emergent Culture in Minimal LLM Systems* | https://arxiv.org/abs/2606.30668 | Small LLM collectives with shared decaying memory can develop persistent artifacts and patterns. Treat as an early/preprint result. |
| Schmer-Galunder (2026), *Culture in the Code: Anthropological Concepts Decoding AI's Hidden Assumptions* | https://arxiv.org/abs/2610.14203 | Critiques reducing culturally situated meaning to quantified proxies; useful framing, not empirical confirmation of this mechanism. |
| Shi et al. (2025), *Judging the Judges* (IJCNLP-AACL) | https://aclanthology.org/2025.ijcnlp-long.18/ | Model-based evaluators have positional biases; blind human annotation plus randomized auditing preferred here. |
| Anthropic (2024), *Claude's Character* | https://www.anthropic.com/news/claude-character | Prior work explicitly tries to shape model character during training. Our harness-level, non-prescriptive treatment is different, not wholly unprecedented. |
| Open Science Framework registration guidance | https://help.osf.io/article/330-welcome-to-registrations | Timestamp plans before collecting confirmatory data, specify endpoints, exclusions, and analysis. |

## Literature gap we can honestly claim as a research question

We have not identified a mature controlled demonstration that **original, non-prescriptive cultural artwork**, delivered as ambient context to a **stateful agent harness**, causes **non-lexical behavioral differences that persist on held-out tasks after withdrawal**, with a **memory-off falsification check and blinded assessment**. That is a proposed empirical question, not a declaration of priority or proof of novelty.

## Publication conduct

Prioritize original papers and their dates. Distinguish peer-reviewed from preprint work. Maintain an update log as new studies appear. Do not attribute invented dialogue to the historical thinkers. Confirm new references before adding them to claims.