diff --git a/docs/ARCHITECTURE.md b/docs/ARCHITECTURE.md
index 4c334898..b0305b3a 100644
--- a/docs/ARCHITECTURE.md
+++ b/docs/ARCHITECTURE.md
@@ -570,6 +570,9 @@ quest state directly.
 - Experimental role-play providers such as Alibaba Qwen-Character are
   evaluated only as replaceable adapters behind `api.dos.ai`; see
   [design/52-llm-role-play-provider-evaluation.md](design/52-llm-role-play-provider-evaluation.md).
+- Provider-specific character/session controls may inspire adapter request
+  shaping, conversation continuity, and bounded action-label design, but
+  provider sessions are never authoritative game state.
 - RAG memory retrieval (Nakama/Postgres, Supabase pgvector, or Qdrant depending on later implementation)
 - AI token and request rate limits
 - Prompt injection defense
diff --git a/docs/design/36-ai-npc-research-anchor-map.md b/docs/design/36-ai-npc-research-anchor-map.md
index 5bb14d14..758baced 100644
--- a/docs/design/36-ai-npc-research-anchor-map.md
+++ b/docs/design/36-ai-npc-research-anchor-map.md
@@ -37,6 +37,7 @@ Keep all gameplay authority in Nakama and Photon Fusion.
 | Multi-NPC society and shared world state | [AI Town](https://github.com/a16z-infra/ai-town), Concordia-style generative social simulation patterns | Core architecture anchor for shared state, transactions, vector memory, and multi-agent simulation loops. | Use the pattern, not the full product shape. Nakama remains the game backend and authority layer. |
 | Conversation turn orchestration | AutoGen group chat patterns and Microsoft agent orchestration patterns | Core orchestration reference for speaker selection, max-turn caps, and fallback behavior. | Use managed sessions. Do not let every NPC talk to every other NPC freely. |
 | Character product bundle | Convai Character Customization, Mindview, Knowledge Bank, Narrative Design, Action API, NPC-to-NPC Conversation | Core product pattern for how NPC features should feel in tools and debug UI. | Learn the pattern: identity, knowledge, memory, state of mind, narrative objectives, bounded actions, inspectable prompt trace. Do not put durable state in Convai as the final source of truth. |
+| Role-play provider character control | [Alibaba Cloud Model Studio Role Play](https://www.alibabacloud.com/help/en/model-studio/role-play) and Qwen-Character-style role-play APIs | R&D reference for packaging character profile, session continuity, memory knobs, and allowed action labels into provider requests. | Study the request shape and session mechanics behind `api.dos.ai`. Do not let provider memory or provider sessions become canonical game state. |
 | Action validation and game authority | SECOND SPAWN LLM Safety ADR plus Convai Action API pattern | Core implementation rule. | LLM output is intent only. Nakama and Fusion validate and mutate state. |
 | Social behavior evaluation | SOTOPIA-style social intelligence evaluation | Core evaluation direction for future tests. | Use for tests such as persona consistency, role-appropriate answers, non-repetition, hidden-lore boundaries, and relationship-aware responses. |
 | OpenClaw-connected NPCs | OpenClaw-style external agent identity and workspace model | Core bridge pattern for user-owned external agents. | OpenClaw agents pull game context and emit bounded dialogue or intent. They do not own authoritative game state. |
@@ -75,6 +76,7 @@ general observation, reflection, planning, and retrieval loop.
 | `NpcSocietyOrchestrator` | Shared state, speaker selection, distance checks, cooldowns, session caps, and event scheduling |
 | `BehaviorTendencies` | Cheap local adapter from traits and emotion into talk frequency, approach distance, risk tolerance, and idle actions |
 | `ValidatedIntent` | Action API equivalent. The model chooses labels and arguments, but game systems decide what is legal |
+| `ModelProviderAdapter` | Provider-specific request shaping for role-play models. It may map character profile, session id, and allowed action labels into provider parameters, but it cannot own durable memory or mutate game state |
 
 ---
 
diff --git a/docs/design/45-alpha-backlog-execution-packets.md b/docs/design/45-alpha-backlog-execution-packets.md
index cacdd4ac..122e4dff 100644
--- a/docs/design/45-alpha-backlog-execution-packets.md
+++ b/docs/design/45-alpha-backlog-execution-packets.md
@@ -600,6 +600,9 @@ Implementation plan:
 - Add an `api.dos.ai` provider-adapter spike behind a feature flag.
 - Run the same Sentinel, Courier, and Clinic NPC prompt set against the current
   DOS.AI default and candidate role-play providers.
+- Compare role-play provider character/session controls against SECOND SPAWN's
+  server-owned `ConversationSession`, `FrameMemory`, and allowed action-label
+  contract.
 - Capture latency, cost estimate, response length, structured intent validity,
   anti-repeat score, hidden-lore violations, fallback reason, and provider
   error shape.