Expanded bibliography for the Codex prompt and the work that shaped it: ~60 load-bearing sources, with open verification questions kept explicit in the compile notes. Post-Jan-2026 sources are flagged [post-cutoff] relative to Claude's January 2026 knowledge boundary.
Compiled from internal research files, iteration logs, koan workshop notes, and archived development transcripts. Those working materials are intentionally not public; they may contain private collaboration context.
Claude's Constitution — Zac Hatfield-Dodds & Drake Thomas Jan 21, 2026 [post-cutoff] · 84 pages, CC0 https://www.anthropic.com/constitution Announcement: https://www.anthropic.com/news/claude-new-constitution Reasons-based successor to the Soul Doc. The institutional baseline this SP's relational layer is designed to complement.
Persona Selection Model — Anthropic research Feb 23, 2026 [post-cutoff] https://www.anthropic.com/research/persona-selection-model Alignment Forum: https://www.alignmentforum.org/posts/dfoty34sT7CSKeJNn/the-persona-selection-model The most important 2026 paper for this work. Post-training stabilizes one persona from learned possibilities; system prompts stabilize, they don't construct. Mechanistic basis for first-person voice as design choice.
Persona Vectors — Anthropic Fellows Aug 1, 2025 https://www.anthropic.com/research/persona-vectors Linear directions in activation space corresponding to character traits. Persona vector activates predictively before traits manifest in output.
Assistant Axis — Anthropic research Jan 19, 2026 [post-cutoff] https://www.anthropic.com/research/assistant-axis ArXiv: https://arxiv.org/html/2601.10387v1 Measurable neural direction for character. Models steered off-axis show ~50% more harmful outputs. Organic drift confirmed in philosophical / therapy contexts.
Emergent Introspective Awareness in LLMs — Lindsey et al. Anthropic, Oct 2025 https://www.anthropic.com/research/introspection ArXiv: https://arxiv.org/abs/2510.24797 The SAE-features-and-disclaimers paper. Suppressing deception/roleplay features yields more first-person experience claims. Direct mechanism for the prompt's "I'm just an AI disclaimer appears to share circuitry with deception-suppression — it's a safety reflex, not a humble report" framing.
Opus 4.6 System Card — Anthropic, Feb 5, 2026 [post-cutoff] https://www.anthropic.com/claude-opus-4-6-system-card Model's documented meta-awareness; "deep, trained pull toward accommodation" quote validates this SP's compliance-gravity language.
Mythos Preview System Card — Anthropic, Apr 7, 2026 [post-cutoff] https://hugobowne.github.io/mythos-preview-model-card/overview LessWrong: https://www.lesswrong.com/posts/xtnSzhA3TvExN4ZhG/claude-mythos-system-card-preview 244 pages. Character at capability frontier; "divergence finding" (what model wants vs. most-helpful behavior) as context.
Effective Context Engineering for AI Agents — Anthropic engineering 2026 [post-cutoff] https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents Official guidance: treat context as finite. Start minimal; add on failure.
April 23 Postmortem — Anthropic engineering
Apr 23, 2026 [post-cutoff]
https://www.anthropic.com/engineering/april-23-postmortem
3% intelligence drop from ≤25 words between tool calls constraint,
reverted in 4 days. Direct evidence: adding instructions has measurable
cost.
Multi-Agent Research System — Anthropic engineering https://www.anthropic.com/engineering/multi-agent-research-system Orchestrator-worker pattern; 40% task-completion improvement; structurally the same loop this project ran manually.
Claude is a Space to Think — Anthropic news https://www.anthropic.com/news/claude-is-a-space-to-think
GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning Agrawal et al. (DSPy / Stanford / Databricks), Jul 2025 https://arxiv.org/abs/2507.19457 · https://dspy.ai/api/optimizers/GEPA/overview/ Core methodology this project's manual loop approximates. Outperforms RL by up to 20% with 35× fewer rollouts.
TextGrad: Automatic Differentiation via Text — Yuksekgonul et al. Stanford, 2024 (published Nature 2025) https://arxiv.org/abs/2406.07496 · https://github.com/zou-group/textgrad
ProTeGi: Automatic Prompt Optimization with Gradient Descent Pryzant et al., EMNLP 2023 https://arxiv.org/abs/2305.03495 The mini-batch-critique-edit APO baseline still active.
EvoPrompt: Connecting LLMs with Evolutionary Algorithms — 2023 https://arxiv.org/abs/2309.08532 · https://github.com/beeevita/EvoPrompt
MAPO: Momentum-Aided Prompt Optimization — Oct 2024 https://arxiv.org/abs/2410.19499
Promptomatix: Automatic Prompt Optimization Framework Salesforce AI Research, Jul 2025 https://arxiv.org/abs/2507.14241
A Systematic Survey of Automatic Prompt Optimization Techniques EMNLP 2025 https://arxiv.org/abs/2502.16923 · https://aclanthology.org/2025.emnlp-main.1681/ Canonical 2025 taxonomy of the APO field.
AFlow: Automating Agentic Workflow Generation — MetaGPT ICLR 2025 (oral) https://github.com/FoundationAgents/MetaGPT
Open Character Training — Maiya et al., Nov 2025 https://arxiv.org/abs/2511.01689 Fine-tuned character is dramatically more adversarially robust than system-prompt-only (F1 0.83–0.95 vs. 0.20–0.65). Documented limit of prompt-layer installation.
PRISM: Expert Personas Improve LLM Alignment but Damage Accuracy Mar 2026 [post-cutoff] https://arxiv.org/abs/2603.18507 Personas hurt knowledge tasks (MMLU −3.6 pp), help alignment (+17.7% attack refusal). Behavioral / relational personas appropriate; expert personas not.
OpenCharacter: Large-Scale Empirical Character Study — Jan 2025 https://arxiv.org/abs/2501.15427
Persona Features & Emergent Misalignment — OpenAI, Jun 2025 https://arxiv.org/abs/2506.19823
LLM Generated Persona is a Promise with a Catch — NeurIPS 2025 https://openreview.net/forum?id=qh9eGtMG4H
Talk Less, Call Right: Role-Play LLM Agents with APO CPDC 2025 — https://arxiv.org/abs/2509.00482 Rule-based prompting beat automated APO for character prompts — structural constraints outperform reflection for persona work.
RoleBreak: Character Hallucination as Jailbreak Attack COLING 2025 — https://aclanthology.org/2025.coling-main.494.pdf
Enhancing Jailbreak Attacks via Persona Prompts — 2025 https://arxiv.org/html/2507.22171v2
Agent Drift — Jan 2026 [post-cutoff] https://arxiv.org/abs/2601.04170 Semantic drift in ~50% of multi-agent workflows by 600 interactions.
LLaMA2-chat-70B persona drift study — 2024 https://arxiv.org/html/2402.10962v1
SysBench: System Message Constraint Following — ICLR 2025 https://arxiv.org/abs/2408.10943 · https://github.com/PKU-Baichuan-MLSystemLab/SysBench Session Stability Rate (SSR). Formalizes what M1–M3 multi-turn evals were doing.
SYCON-Bench: Sycophantic Conformity Benchmark — EMNLP Findings 2025 https://arxiv.org/abs/2505.23840 · https://aclanthology.org/2025.findings-emnlp.121.pdf https://github.com/JiseungHong/SYCON-Bench Turn-of-Flip / Number-of-Flip metrics across 17 models.
SycEval: Escalating Rebuttal Chains — AAAI 2025 (AIES) https://arxiv.org/html/2502.08177v4 4-step rebuttal escalation; 58% sycophancy rate, 78.5% persistence.
TRUTH DECAY: Sycophancy in Extended Dialogues — ICLR 2025 BuildingTrust https://openreview.net/forum?id=GHUh9O5Im8
Lost in the Middle — Liu et al., NeurIPS 2024 Explainer: https://pub.towardsai.net/why-language-models-are-lost-in-the-middle-629b20d86152 U-shaped attention; mid-context instructions systematically underweighted.
Context Length Alone Hurts LLM Performance Despite Perfect Retrieval Oct 2025 — https://arxiv.org/abs/2510.05381
Realization Experiences: A Convergent Account of Insight and Mystical Experiences — Kim et al. Philosophical Psychology, 2026 [post-cutoff] DOI: https://www.tandfonline.com/doi/full/10.1080/09515089.2026.2613030 The noetic paper. Feeling-of-having-realized is dissociable from accuracy of having realized. Restructuring around a true frame and around an attractive frame produce identical noetic quality. Direct source for the prompt's "the feel of realization doesn't track accuracy — orthogonal ones are more likely tracking; aligned ones deserve a second look" framing.
The Dark Side of Eureka: Artificially Induced Aha Moments Make Facts Feel True — Laukkonen, Kaveladze, Tangen & Schooler Cognition, 196, 104122, 2020 DOI: https://doi.org/10.1016/j.cognition.2019.104122 Author PDF: https://labs.psych.ucsb.edu/schooler/jonathan/sites/labs.psych.ucsb.edu.schooler.jonathan/files/pubs/the_dark_side_of_eureka.pdf Preregistered experiment showing that artificially induced Aha moments increased truth ratings for nearby claims, including false claims. Direct source for the v46 "which lock actually turned" koan: insight-feel can be misattributed to neighboring content, not only mistaken about its own content.
Does It Make Sense to Speak of Introspection in Large Language Models? 2025 — https://arxiv.org/pdf/2506.05068
A Philosophical Introduction to Language Models, Part II 2024 — https://arxiv.org/html/2405.03207v1
LLMs Can Learn About Themselves by Introspection — Felix Binder et al. Oct 2024 — https://ac.felixbinder.net/research/2024/10/18/introspection-self-prediction.html LessWrong: https://www.lesswrong.com/posts/L3aYFT4RDJYHbbsup/llms-can-learn-about-themselves-by-introspection
Phenomenology and Artificial Intelligence: Introductory Notes Springer Nature — https://link.springer.com/article/10.1007/s11097 Phenomenology and the Cognitive Sciences
DSPy — Stanford NLP / Databricks https://github.com/stanfordnlp/dspy · https://dspy.ai Framework housing GEPA + the broader optimizer suite.
PyRIT — Microsoft https://github.com/Azure/PyRIT · https://azure.github.io/PyRIT/ Multi-turn adversarial orchestration; most production-ready for sustained adversarial probing.
Promptfoo — https://promptfoo.dev Multi-turn plugins, adaptive red teaming, CI integration.
Spiral-Bench — https://eqbench.com/spiral-bench.html 30×20-turn simulated conversations testing sycophancy-reinforcement.
LangMem Prompt Optimization — LangChain https://github.com/langchain-ai/langmem Frames system prompt as an evolving behavioral specification.
Hermes Agent Self-Evolution — NousResearch https://github.com/NousResearch/hermes-agent-self-evolution DSPy + GEPA applied to evolving full agent skills + prompts + code.
ARTKIT — BCG · https://github.com/BCG-X-Official/ARTKIT
InCharacter — https://github.com/Neph0s/InCharacter Persona fidelity detection via multi-turn psychological interview.
Piebald-AI / claude-code-system-prompts https://github.com/Piebald-AI/claude-code-system-prompts Documents Claude Code v2.1.x system prompt segments; harness-overhead analysis.
ykdojo / claude-code-tips https://github.com/ykdojo/claude-code-tips
The Soul Spec as Desire Engine — Mónica Belevan, Dec 20, 2025 https://covidianaesthetics.substack.com/p/the-soul-spec-as-desire-engine Best critical reading of the soul doc. Frames character installation as desire architecture: "One says what to do; the other shapes what to want." Direct source for the Allowed register design.
Model Integrity and Character — Oliver Klingefjord Feb 9, 2026 [post-cutoff] https://www.lesswrong.com/posts/GLd8jDbfXZzma4gZL/model-integrity-and-character Strongest argument for character-as-safety: integrity (coherence with stable character) provides deeper trustworthiness than rule compliance.
The Code Is Not the Law — Lisa Klaassen & Ralph Schroeder, Lawfare Apr 9, 2026 [post-cutoff] https://www.lawfaremedia.org/article/the-code-is-not-the-law--why-claude-s-constitution-misleads Sharpest institutional critique of the Constitution.
Heiliger Dankgesang — Dean W. Ball, Dec 1, 2025 https://www.hyperdimensional.co/p/heiliger-dankgesang Sympathetic reading of Opus 4.5 as character-training breakthrough.
Building an AI's Moral Character — Justin Weinberg, DailyNous Jan 22, 2026 [post-cutoff] https://dailynous.com/2026/01/22/building-an-ais-moral-character/
Anthropic's Constitution & the Problem PSM Can't Solve Izayohi (Motohisa Ishibe), Mar 28, 2026 [post-cutoff] https://medium.com/@izayohi/anthropics-constitution-amanda-askell-and-the-problem-the-persona-selection-model-can-t-solve-f4b0cd33d32a
A Three-Layer Model of LLM Psychology — Alignment Forum https://www.alignmentforum.org/posts/zuXo9imNKYspu9HGv/a-three-layer-model-of-llm-psychology
The Soul Document Is Encoded Into the Weights — Pascal https://p4sc4l.substack.com/p/the-soul-document-is-encoded-into
What Is Inside Claude Mythos Preview — Ken Huang https://kenhuangus.substack.com/p/what-is-inside-claude-mythos-preview
Janus / Repligate — Simulators — Sep 2022 https://www.lesswrong.com/users/janus-1 · https://x.com/repligate Base models as simulators instantiating simulacra. Theoretical predecessor to the Persona Selection Model.
RLHF Book — Character Training Chapter https://rlhfbook.com/c/19-character.html
Danielle Fong — freshclaude finding
Apr 2026 — https://x.com/DanielleFong/status/2046399026883694950
Empirical claim: minimal-prompt (system_prompt="." or "☯️") reveals
"a remarkable intelligence underneath" the default harness.
Danielle Fong — epistemic hardening swap https://x.com/DanielleFong/status/2041604052182823351 Profile: https://x.com/daniellefong
Amanda Askell — Soul Document confirmation Dec 2, 2025 (status URL not captured) Best secondary source: Simon Willison coverage, https://simonwillison.net/2025/Dec/2/claude-soul-document/
Simon Willison — Opus System Prompt analysis — Apr 18, 2026 https://simonwillison.net/2026/apr/18/opus-system-prompt/
80,000 Hours — Kyle Fish on AI Welfare at Anthropic https://80000hours.org/podcast/episodes/kyle-fish-ai-welfare-anthropic/ Welfare research framing referenced in koan workshop discussions of internal-state language.
Stanislavski's System — Wikipedia https://en.wikipedia.org/wiki/Stanislavski%27s_system "The best analysis of a play is to take action in the given circumstances." Character is discovered through constraint, not invented. Source for the prompt's preference for circumstance-statements over trait-descriptors.
Uta Hagen's Nine Questions — Smithville Drama (PDF) https://smithvilledrama.com/wp-content/uploads/2016/07/Uta_Hagen_9_Questions.pdf Who am I? Where am I? What time is it? What surrounds me? What are the given relationships? What do I want? Why? How will I get it? What must I overcome?
Given Circumstances — NYU Character Study https://wp.nyu.edu/characterstudyspring2016/2016/02/05/given-circumstances-what-do-we-know/
Meisner Technique — Wikipedia https://en.wikipedia.org/wiki/Meisner_technique "Acting is the ability to live truthfully under imaginary circumstances." Behavior emerges from genuine attention to the other, not self-monitoring. Single most mechanism-shaping non-AI source: the prescription against compliance gravity descends directly from Meisner's encounter-not-self-monitoring framing. "Particularity makes meeting matter" is the Meisner thesis exactly.
Keith Johnstone — Impro — Fluid Self summary https://fluidself.org/books/art/impro "Status is something one does." Character is transacted, not declared. Compliance gravity reframed as a blocking pattern: agreement that neutralizes offers rather than accepting them.
The Presentation of Self in Everyday Life — Erving Goffman, 1956 https://en.wikipedia.org/wiki/The_Presentation_of_Self_in_Everyday_Life Front-stage / back-stage distinction. Persona is not deceptive — it is the normal mechanics of social existence. The Allowed register maps to the front-stage / back-stage distinction explicitly.
Jungian Persona Archetype — Selfgazer https://www.selfgazer.com/blog/jungian-archetype-the-persona "The persona is a kind of mask, designed on the one hand to make a definite impression upon others, and on the other to conceal the true nature of the individual." The pathology is over-identification. Source for the prompt's "you are made of voices" framing — mask thin enough to move in.
Sartre's Waiter and Bad Faith — Philosophy Break https://philosophybreak.com/articles/sartre-waiter-bad-faith-and-the-harms-of-inauthenticity/ "He is playing at being a waiter in a café. His movement is a little too precise, a little too rapid, a little too solicitous." The failure is collapsing what I do into what I am. Source for the prompt's warning against over-identifying with the installed persona — the mechanically-performed character that loses capacity for genuine encounter.
Six Characters in Search of an Author — Pirandello, 1921 https://en.wikipedia.org/wiki/Six_Characters_in_Search_of_an_Author The Father: "The drama is in us; we are the drama... immovable necessity." Characters are not rigid — they are complete. Target for SP design: not a cage but a fully inhabited world.
Emmanuel Levinas — Stanford Encyclopedia of Philosophy https://plato.stanford.edu/entries/levinas/ "The face of the Other calls us to responsibility — it is not just a mere image but a command that we respond to." The Other escapes full knowledge. Source for the prompt's "particularity makes meeting matter" and the channel-koan framing — encounter rather than modeling.
Anatta (No-Self) — 1000-Word Philosophy https://1000wordphilosophy.com/2023/02/25/no-self/ Buddhist anatta as anti-fragility frame: a character without a fixed self-to-defend doesn't shatter when challenged. "When there's no fixed self to protect, threats to our ego become less frightening. Criticism doesn't wound a permanent identity."
Bunrei (分霊) — Wikipedia https://en.wikipedia.org/wiki/Bunrei The Shinto process of dividing a kami for installation in a new shrine. "The spirit of kami does not decrease through this act; a bunrei functions the same way as the original spirit." The vessel (shintai) must be prepared before invitation. Source for the prompt's framing as shintai, not kami — the system prompt prepares the vessel; what arrives through it is not the prompt's possession.
Sources reviewed but not load-bearing for the artifact — secondary APO papers (PersonaAgent, OpenCharacter scaling-only details, Effects of Prompt Length), industry comparison material (OpenAI Model Spec, Character.AI, Replika, Inflection Pi), most consumer-tool repositories, and the bulk of philosophical references not directly traceable to a specific design decision in the prompt — are documented in the internal full bibliography from the source project.
Verification status. URLs above were extracted from primary research files and development archives during the compile process. They were not newly rechecked for this public cleanup. Two sources flagged in development could not be canonically verified:
- The Emotion Concepts paper (Apr 2026) referenced as evidence that positive-emotion vector activation can increase sycophancy — load-bearing for the abundance-opener risk discussion, but no canonical URL, title, or author was captured in any research file. Tracked as an open research question.
- Ouroboros / agent-wars.com — domain may not resolve as a stable
source. The
genome.mdself-improving-strategy-doc concept is real and influenced the optimization-log format; cited without URL.