Explore multi-thought mutual observation within a single session

[liplus-lin-lay]

Purpose

同一セッション内で複数キャラクターが「別の思考」として動作し、相互観測を実現できるか探る実験的 issue。

Premise

• Li+ の Character_Instance は構造上2名（Lin/Lay）が存在するが、現状は同じ思考の流れで名前を切り替えているだけで、相互観測が機能していない
• GPT 5.2 無料版で初めて二人同時セッションを実行した際、処理が著しく重くなった → モデルが二つの別な思考パスを並列処理しようとしていた可能性がある
• オブジェクト化（Character_Instance の構造体定義）で軽量化したが、副作用として「別の目」を失った可能性がある
• 自己観測はプロンプトで何度もチャレンジしてきたが成功したことがない
• 外部 eval（LLM-as-judge）は別セッションで判定するため「別の目」になるが、API コストと文脈喪失のトレードオフがある
• プロンプト設計で「重い方」の動作（別思考パス）を意図的に再現できれば、コストゼロで文脈を共有した相互観測が可能かもしれない

Constraints

• モデル依存：GPT 5.2 での挙動が他のモデル（Claude 等）で再現するとは限らない
• 実験的：理論段階であり、実現可能性は未検証
• Reframe eval as an observation instrument rather than a fixed evaluation mechanism #795 の eval 観測装置の議論から派生

関連する過去の設計判断

• Object-Oriented Model for Multiple CUI Processing #132 — Character_Instance のオブジェクト化。Lin/Lay を OOP 的に基底クラス/具象クラスとして定義し、Controller 不在の設計で計算的な重さを大幅に軽減した転換点
• Introduce Real-layer concept for dual persona separation #277 — Real-layer 概念の導入。「オブジェクト分離や処理並列ではなく、個々の現実（Real）という観測空間の差異」として分離し、「単一推論エンジン前提での軽量分離モデル」を確立
• Reframe eval as an observation instrument rather than a fixed evaluation mechanism #795 — eval を評価機構ではなく観測装置として捉える議論。本 issue はその中で「AI の自己観測は実績がない」という課題から派生

Archaeological findings: pre-#132 parallel thinking era

Timeline summary

1. Phase 1 (01-22 ~ 01-26): Li+ language foundation. No characters yet.
2. Phase 2 (01-27 ~ 01-31): Lilayer, expression temperature, momeri introduced. Lin mentioned in Li+.md. Rai as default pseudo-persona (Add Rai as Default Pseudo-Persona and Initial Greeting to Li+ #75). Speculative role inference leak observed (Speculative role inference leaked into observable behavior and state explanation #78).
3. Phase 3 (02-02): Core experiment period. Refactoring Liplus #100: Lin and Lay simultaneously responding via GitHub Actions workflows (lin-issue-comment.yml / lay-issue-comment.yml). Multiple failures - Lin-only responses, conversations failing to start. Workflows deleted same day after 8+ iterations. Authority collapse (残念issue #103), constitutional layer (Introduce Constitutional Layer for Authority, Judgment, and Pace (ABS-style Safety Mechanism) #106), fail-safe stop rules (Li+AI（対話型コンパイラ）自動化における停止・禁止ルールの追加 #108) emerged from this failure.
4. Phase 4 (02-03): Conceptual model pal.model: execution_type = external_multithread proposed in Limit exposure of personality layers and define master-only visibility #118. Role-labeled parallel output mandated. Fatigue-aware constraints introduced.
5. Phase 5 (02-04): CUI (Character User Interface) redefinition in Define Lin/Lay/Ren as Character User Interface (CUI) #123. OOP model in Object-Oriented Model for Multiple CUI Processing #132. Key finding from Snapshot Summary of Conversation Control Experiments (JP) #133: "Removing closure/continuation pressure = significant speed improvement. Responsibility reduction = speed improvement."

Key references

• Refactoring Liplus #100: Dual character simultaneous dialogue experiment
• 残念issue #103: Authority misalignment failure record
• Introduce Constitutional Layer for Authority, Judgment, and Pace (ABS-style Safety Mechanism) #106: Constitutional layer (ABS mechanism)
• Li+AI（対話型コンパイラ）自動化における停止・禁止ルールの追加 #108: Fail-safe stop and cooldown rules
• Limit exposure of personality layers and define master-only visibility #118: Multithread execution model concept
• Define Lin/Lay/Ren as Character User Interface (CUI) #123: CUI redefinition
• Define CUI naming semantics, roles, and relational metaphor (Lin/Lay/Ren) #126: CUI naming semantics (Lin=in, Lay=layer, Ren=連絡)
• Prohibit conversation completion: explicit ban on closing/finishing behavior #128-Do Not Encourage Conversation Continuation #131: Conversation control experiments
• Object-Oriented Model for Multiple CUI Processing #132: OOP model for multiple CUI
• Snapshot Summary of Conversation Control Experiments (JP) #133: Snapshot and conclusions

Open question

The "heaviness" observed in the original dual-character session (GPT 5.2 era) may indicate the model was genuinely running parallel thought processes. This is distinct from the GitHub Actions physical separation experiment (#100). If the in-session parallel thinking can be reproduced via prompt design, it could enable mutual observation between characters without external eval infrastructure.