- title
- Observer-Sovereign Rendering(OSR)整合技術白皮書
- english_title
- Observer-Sovereign Rendering (OSR) Integrated Technical Whitepaper
- author
- Neo.K
- ai_collaboration
- GPT-5.6 Sol
- institution
- EveMissLab/一言諾科技有限公司
- version
- v0.1
- date
- 2026-09-18
- language
- zh-TW
- document_type
- Integrated Technical Whitepaper / Engineering Specification
- status
- Architecture Baseline / MVP Handoff
- source_series
- OSR-00 · OSR-01 · OSR-02 · OSR-03 · OSR-04 · OSR-05 · OSR-06 · OSR-07
Observer-Sovereign Rendering(OSR)整合技術白皮書
從觀察者條件化顯現到 APR 驅動的 Local–Cloud AI 原生渲染 Runtime
Observer-Sovereign Rendering (OSR) Integrated Technical Whitepaper
From Observer-Conditioned Manifestation to an APR-Governed Local–Cloud AI-Native Rendering Runtime
作者: Neo.K
AI 協作: GPT-5.6 Sol
機構: EveMissLab/一言諾科技有限公司
版本: v0.1
日期: 2026 年 9 月 18 日
狀態: Architecture Baseline / MVP Handoff
0. 文件目的
本白皮書將 OSR-00 至 OSR-07 的理論系列壓縮為一個可交付本地端 Agent、Codex、Claude Code 或其他工程 AI 直接實作的技術基線。
本文件的工作不是再提出新的 OSR 理論,而是固定:
- Runtime component graph;
- canonical data contracts;
- OSR-IR;
- APR integration boundary;
- rendering backend/plugin ABI;
- truth / provenance contract;
- local–cloud carrier contract;
- temporal reconstruction path;
- PNCW / MRMIC-NVCL integration boundary;
- benchmark harness;
- failure / fallback semantics;
- MVP implementation phases;
- acceptance gates;
- non-goals。
本白皮書的最小工程目標是:
並在 runtime 中由 APR 持續治理:
1. 核心產品命題
OSR 不把 AI-native rendering 定義為:
用 AI 替代 rasterizer。
也不定義為:
每一幀送到雲端重新生成。
OSR 的工程命題是:
本身成為第一級 runtime computation。
對每個 observer、scope、時間與資源狀態,系統決定:
- 要不要重新處理;
- 哪些內容必須顯現;
- 使用哪種 realization;
- 表示深度多少;
- 更新頻率多少;
- 可否 reuse;
- 是否需要 local neural;
- 是否值得 cloud inference;
- 是否需要 verification;
- 失效後如何 fallback;
- output 具有什麼 truth status。
2. 第一代 MVP 核心假說
第一代 MVP 不嘗試證明:
或:
只驗證:
第一階段尤其測:
是否可在某些低至中等動態場景中,形成:
且:
同時:
3. 非目標
v0.1 不做:
- 自研完整遊戲引擎;
- 自研 production path tracer;
- 仿製 NVIDIA DLSS 5;
- 逆向任何 proprietary neural renderer;
- 60 FPS general-purpose cloud image API;
- production multiplayer anti-cheat;
- production robotics / medical deployment;
- 全平台 renderer backend;
- 自動把 AI inference 寫回 canonical world;
- production-grade provider SLA;
- universal OSR quality score。
第一代應先證明 architecture semantics 與 controlled prototype utility。
4. 與現有 EveMissLab Runtime 的分工
OSR 不重寫已存在的 runtime。
4.1 APR
APR 已提供:
- Persistent World State;
- EvidenceStore;
- Goal;
- Budget;
- PolicyController;
- adapters / sources;
- plugin registry;
- hosted semantic inspectors;
- image generation plugin boundary;
- closed-loop observe / verify / recover。
OSR 直接消費 APR state。
不把:
- raster;
- ray;
- cloud renderer;
硬寫進 APR core。
4.2 PNCW
PNCW 已提供:
以及:
- versioned projection contract;
- deterministic identity;
- source / surface authority separation;
- freshness / integrity verification;
- fail-closed;
- idempotent visibility commit。
OSR 不重寫可見性狀態機。
4.3 MRMIC / NVCL
MRMIC/NVCL 已提供:
- Canvas-first visual world;
- geometry / projection;
- provider-owned resource boundary;
- secure principal;
- native resource portal;
- mounted / visible / focused / controlOwner separation;
- visual observe / repair loop。
OSR 不重寫 Canvas。
5. OSR 新增的工程層
OSR 真正新增:
Observer Adapter
↓
Manifestation Obligation Builder
↓
Perceptual Rendering Compiler (PRC)
↓
OSR-IR
↓
Heterogeneous Field Solver
↓
Backend Execution
↓
Truth / Verification Gate
↓
Temporal Reconstruction
↓
Projection Bridge
也就是:
6. 系統總架構
┌──────────────────────────────────────────────────────────┐
│ Canonical World │
│ engine state / scene graph / physics / IDs │
└───────────────────────┬──────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ APR World/Evidence Layer │
│ goal / belief / freshness / uncertainty / risk / budget │
└───────────────────────┬──────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ OSR Observer Layer │
│ camera + goal + profile + authority + runtime state │
└───────────────────────┬──────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ Manifestation Obligation Builder │
│ minimum / preferred / deadline / provenance / forbidden │
└───────────────────────┬──────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ Perceptual Rendering Compiler │
│ capability matching / hard gates / cost / utility │
└───────────────────────┬──────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ OSR-IR / Heterogeneous Rendering Field │
│ raster / ray / reuse / local neural / cloud / semantic │
└──────────────┬────────────────────────────┬──────────────┘
│ │
▼ ▼
Local Execution Cloud Intelligence
│ │
└─────────────┬──────────────┘
▼
┌──────────────────────────────────────────────────────────┐
│ Truth Firewall / Verification │
│ structural / identity / temporal / provenance / policy │
└───────────────────────┬──────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ Local Temporal Reconstruction / Compositor │
└───────────────────────┬──────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ PNCW Projection Lifecycle │
│ PROJECTED → VERIFIED → VISIBLE │
└───────────────────────┬──────────────────────────────────┘
│
▼
┌──────────────────────────────────────────────────────────┐
│ MRMIC / Engine Surface / Final Display │
└──────────────────────────────────────────────────────────┘
7. Runtime Package Boundary
建議新 repo:
osr-runtime/
或:
Observer-Sovereign-Rendering/
v0.1 建議結構:
osr-runtime/
├─ osr_core/
│ ├─ observer.py
│ ├─ obligations.py
│ ├─ compiler.py
│ ├─ ir.py
│ ├─ field_solver.py
│ ├─ truth.py
│ ├─ receipts.py
│ └─ errors.py
├─ osr_apr/
│ ├─ world_adapter.py
│ ├─ policy_adapter.py
│ ├─ evidence_adapter.py
│ └─ runtime_metrics.py
├─ osr_backends/
│ ├─ base.py
│ ├─ raster_mock.py
│ ├─ reuse.py
│ ├─ local_neural.py
│ ├─ cloud_image.py
│ └─ semantic.py
├─ osr_temporal/
│ ├─ warp.py
│ ├─ confidence.py
│ ├─ compositor.py
│ └─ cache.py
├─ osr_truth/
│ ├─ contracts.py
│ ├─ verifier.py
│ ├─ provenance.py
│ └─ writeback_firewall.py
├─ osr_bridge/
│ ├─ pncw.py
│ ├─ mrmic.py
│ └─ engine_protocol.py
├─ contracts/
│ ├─ observer-state-v1.schema.json
│ ├─ manifestation-obligation-v1.schema.json
│ ├─ backend-capability-v1.schema.json
│ ├─ osr-ir-v1.schema.json
│ ├─ truth-record-v1.schema.json
│ ├─ cloud-request-v1.schema.json
│ └─ cloud-response-v1.schema.json
├─ engines/
│ ├─ godot_reference/
│ └─ simulator/
├─ benchmarks/
│ ├─ fixtures/
│ ├─ scenarios/
│ ├─ metrics/
│ └─ runner/
├─ tests/
├─ examples/
├─ docs/
└─ artifacts/
8. 語言與程序邊界
v0.1 建議:
- OSR control plane:Python 3.10+;
- APR:直接 Python import;
- engine adapter:engine-native;
- local high-frequency temporal path:engine / native side;
- cloud adapter:Python;
- contracts:JSON Schema;
- benchmark data:JSONL / CSV / PNG / MP4。
重要:
Python 不負責 120 Hz per-pixel warp。
9. 控制面與資料面的分離
Control Plane
負責:
- observer;
- obligation;
- APR;
- compiler;
- backend selection;
- truth contract;
- cloud request;
- receipts。
Data Plane
負責:
- RGB;
- depth;
- normal;
- motion;
- masks;
- GPU textures;
- warp;
- compositor。
Control plane 盡量傳:
不是每幀把 4K texture 複製進 Python。
10. ObserverState v1
{
"schema": "osr-observer-state/v1",
"observer_id": "player-1",
"revision": 12,
"camera": {
"id": "main",
"revision": 144
},
"goal": {
"mode": "dialogue",
"targets": ["npc_07"]
},
"profile": {
"style": "default",
"accessibility": []
},
"authority": {
"world_read": true,
"cloud_export": true,
"hidden_state_read": false
},
"risk": {
"level": "low"
},
"budget_ref": "budget-822"
}
11. ManifestationObligation v1
{
"schema": "osr-manifestation-obligation/v1",
"id": "obl-face-007",
"scope": {
"kind": "object_region",
"object_id": "npc_07",
"subregion": "face"
},
"minimum": {
"representation_depth": 2,
"identity_fidelity": 0.99
},
"preferred": {
"representation_depth": 4
},
"deadline_ms": 33,
"mandatory": false,
"forbidden_modes": [],
"truth": {
"preserve_identity": true,
"world_writeback": false
},
"provenance_required": true
}
12. Manifestation Obligation 不包含 Backend
不能寫:
use_cloud = true
在 obligation 層。
因為:
Backend 由 compiler 決定。
13. BackendCapability v1
{
"schema": "osr-backend-capability/v1",
"backend_id": "cloud-image-A",
"revision": "2026-09-18",
"modes": ["cloud_neural"],
"inputs": [
"rgb",
"depth",
"motion",
"mask",
"text_control"
],
"outputs": [
"pixel_anchor"
],
"region_support": true,
"temporal_state": false,
"deterministic": false,
"network_required": true,
"world_write_authority": false,
"cost_model": "provider-metered",
"limits": {
"max_width": 2048,
"max_height": 2048
}
}
14. Backend ABI
所有 backend 實作:
class RenderBackend:
def capability(self) -> BackendCapability:
...
def estimate(self, request, runtime_state) -> CostEstimate:
...
def execute(self, request) -> BackendResult:
...
def validate_result(self, result) -> ValidationResult:
...
Async backend 可額外:
submit()
poll()
cancel_local_interest()
注意:
不代表 remote operation 一定取消。
15. APR Plugin 重用策略
OSR 優先重用 APR PluginRegistry。
建議 component kinds:
osr_backend
osr_verifier
osr_partition
osr_cost_model
osr_scheduler
osr_engine_adapter
osr_temporal
APR discovery 原則繼續保留:
Plugin registration 不:
- 開網路;
- 讀 credential;
- 載入大型模型;
- 啟動 GPU 工作;
- 啟動 engine process。
16. OSR-IR v1
{
"schema": "osr-ir/v1",
"plan_id": "plan-...",
"world_revision": 9221,
"observer_revision": 12,
"policy_revision": 8,
"capability_revision": "cap-331",
"obligations": [],
"nodes": [],
"edges": [],
"verification": [],
"fallbacks": [],
"commit": [],
"truth": [],
"receipts": {
"plan": true,
"execution": true,
"truth": true
}
}
17. OSR-IR Node
{
"id": "node-face-cloud",
"probe_mode": "cloud_neural",
"backend_id": "cloud-image-A",
"scope_ref": "obl-face-007",
"inputs": [
"base_rgb:face",
"depth:face",
"motion:face",
"mask:face"
],
"outputs": [
"anchor:face"
],
"budget": {
"cloud_calls": 1
},
"deadline_ms": 1200,
"blocking": false,
"truth_output_class": "generated_appearance"
}
18. Dependency Edge
{
"from": "node-face-cloud",
"to": "node-face-verify",
"relation": "requires_output"
}
19. Plan Identity
deterministic mode 下:
20. Compiler Pipeline
normalize obligations
↓
resolve world + runtime capabilities
↓
partition scopes
↓
generate probe candidates
↓
hard-gate invalid candidates
↓
estimate cost / utility / truth risk
↓
solve minimum contracts
↓
allocate preferred upgrades
↓
insert verification
↓
attach fallback
↓
schedule
↓
emit OSR-IR
21. Compiler Hard Gates
候選在 scoring 前先檢查:
- authority;
- privacy;
- export;
- mandatory visibility;
- truth preservation;
- deadline impossibility;
- unsupported carrier;
- world revision;
- provider capability;
- deterministic-only mode;
- multiplayer authorization。
不合法:
不是:
22. Cost Model
第一代每個項目可以只是 measured scalar。
不需要一開始建立 perfect learned predictor。
23. Utility Model
不採 universal score。
Compiler 內部可以為單一 domain 使用 utility heuristic:
但 benchmark 不用同一 scalar 宣稱整體勝負。
24. Global Budget Closure
流程:
- mandatory obligations;
- safety reserve;
- minimum contracts;
- exploration reserve;
- preferred upgrades;
- stability penalty;
- global resource closure。
若 minimum 都不可滿足:
25. APR Integration Contract
OSR 從 APR 取得:
Goal
WorldState
Evidence
Freshness
Uncertainty
Risk
Budget
再補 engine runtime:
GPU utilization
VRAM
frame time
network RTT
bandwidth
cloud budget
backend availability
形成:
26. APR → OSR Need Adapter
class OSRNeedAdapter:
def build_observer_state(apr_state, engine_state):
...
def build_obligations(observer, world, scene):
...
def rerender_need(scope, history):
...
27. NO_RERENDER
v0.1 必須把:
NO_RERENDER
實作成正常 decision。
而不是:
special optimization
條件:
- freshness pass;
- significant delta low;
- risk low;
- reuse valid;
- minimum contract remains satisfied。
28. Refresh Levels
L0 = no rerender
L1 = local refresh
L2 = regional refresh
L3 = global refresh
Full refresh 不是 default。
29. Significant Delta
v0.1 不必先做 AI delta predictor。
先以 engine-native state delta:
- transform;
- visibility;
- material revision;
- animation revision;
- camera;
- gameplay event;
建:
這比 pixel-only 更可靠。
30. Truth Classes
canonical_world_fact
derived_fact
inferred_appearance
generated_appearance
Reuse / style / semantic overlay 作 lineage mode,不另升 authority。
31. TruthRecord v1
{
"schema": "osr-truth-record/v1",
"scope": "npc_07.face",
"source": "engine:npc_07",
"source_revision": 9221,
"truth_class": "generated_appearance",
"transform_chain": [
"base_raster",
"cloud_image_A",
"local_temporal_warp"
],
"backend_revision": "opaque",
"confidence": null,
"freshness_ms": 180,
"verification": [
"identity",
"structural"
],
"reversible": true,
"world_writeback": false,
"output_hash": "sha256:..."
}
32. Truth Firewall
核心 API:
truth_firewall.can_display(result, obligation)
truth_firewall.can_export(carrier, observer, policy)
truth_firewall.can_write_world(result)
v0.1:
can_write_world(generated_appearance) = false
can_write_world(inferred_appearance) = false
無 exception。
33. Reverse Channel
未來若 vision inference 要更新世界:
Inference
→ WorldUpdateProposal
→ Evidence
→ Authorized Verification
→ World Commit
v0.1 OSR MVP 不實作 auto world promotion。
34. Preservation Contract
每個 cloud / neural scope:
{
"preserve": {
"object_count": true,
"identity": true,
"pose": true,
"silhouette": true,
"mandatory_objects": ["weapon_1"]
},
"allow_generate": [
"microtexture",
"skin_scattering",
"hair_lighting"
],
"forbid": [
"semantic_category_change",
"object_addition",
"object_removal"
]
}
35. Verification Ladder
V0 none
V1 structural
V2 identity / semantic
V3 temporal
V4 world / policy
v0.1 不對所有 pixel 做 V4。
36. Verifier ABI
class Verifier:
def verify(self, candidate, reference, contract) -> VerificationResult:
...
第一代 verifier 優先 deterministic:
- mask IoU;
- object count;
- pose;
- required-object presence;
- hash / revision;
- simple identity fixture。
之後再加 VLM verifier。
37. Cloud Plane 原則
v0.1 cloud 不允許成為 minimum manifestation 唯一來源。
38. Cloud Request v1
{
"schema": "osr-cloud-request/v1",
"request_id": "rid-...",
"world_revision": 9221,
"camera_revision": 144,
"scope": "npc_07.face",
"carrier_schema": "pixel-anchor-input/v1",
"carrier_hash": "sha256:...",
"preservation_contract": {},
"generation_contract": {},
"deadline_ms": 3000,
"export_policy_revision": 3
}
39. Cloud Response v1
{
"schema": "osr-cloud-response/v1",
"request_id": "rid-...",
"source_world_revision": 9221,
"backend_id": "general-cloud-image-A",
"backend_revision": "opaque",
"output_class": "pixel_anchor",
"output_hash": "sha256:...",
"latency_ms": 1730,
"provenance": {
"provider": "provider-A"
}
}
40. Cloud Carrier v0.1
第一版只需要:
base RGB ROI
depth ROI
object mask
motion ROI
text / structured control
如果 provider 不能消費 depth / motion:
第一代 general cloud adapter 可以只用:
RGB + mask + prompt/control
但 contract 仍保留其他欄位。
41. 第一代 Cloud Output
只做:
不先做 latent protocol。
理由:
- 可視;
- 易 debug;
- provider 通用;
- 無 shared decoder;
- benchmark 容易。
42. 第二代 Output
再加入:
residual
latent
material latent
lighting latent
43. Revision Binding
每個 cloud result 必須綁:
- world revision;
- camera revision;
- scope;
- request ID。
回來後先:
revision check
↓
scope check
↓
truth check
↓
verification
↓
commit
44. Late Result
若:
已不相容:
VISIBLE_COMMIT = DENY
但可:
CACHE_ONLY
如果 carrier 仍有 asset-level reuse value。
45. Idempotency
同一:
最多 visible commit 一次。
duplicate result:
IGNORE
46. Temporal Reconstruction v0.1
第一版不訓練新 model。
流程:
cloud anchor
+ current depth
+ motion vectors
+ current base frame
↓
depth-aware warp
↓
confidence mask
↓
blend with current base
↓
final region
47. Temporal Confidence
v0.1 可以 heuristic。
48. Composition
之後才替換成 learned compositor。
49. Disocclusion
若:
history invalid
則:
base fill
→ optional local enhance
→ optional cloud refresh
不得用 hallucinated old anchor 補 canonical structure。
50. Camera Cut
立即:
invalidate temporal region history
keep base renderer
schedule optional new anchors
51. Anchor Cache
Key:
scene_revision
object_id
scope
view_cell
lighting_state
style_contract
backend_revision
Value:
anchor
truth_record
quality
timestamp
validity
52. Semantic Render Cache
未來可以更一般:
第一代先 object/scope cache。
53. PNCW Bridge
OSR plan node 完成:
READY
backend 產出:
PROJECTED
truth / verification pass:
VERIFIED
visibility commit:
VISIBLE
如果 PNCW 尚未作 runtime dependency:
v0.1 可先提供 adapter interface + mock lifecycle。
54. PNCW Bridge Interface
class ProjectionBridge:
def requested(self, manifest):
...
def projected(self, output):
...
def verified(self, verification):
...
def visible(self, commit):
...
55. MRMIC / Surface Bridge
v0.1 不要求真正把 game frame 畫進 MRMIC。
只需:
- contract adapter;
- artifact projection demo;
- truth/provenance surface。
真正 engine display 由 engine 自己完成。
56. EngineAdapter v1
class EngineAdapter:
def world_revision(self):
...
def observer_state(self):
...
def list_scopes(self):
...
def capture_base(self, scope):
...
def capture_depth(self, scope):
...
def capture_motion(self, scope):
...
def capture_mask(self, scope):
...
def apply_region_output(self, scope, output):
...
def runtime_metrics(self):
...
57. Reference Engine
v0.1 建議:
理由:
- open source;
- 容易建立小場景;
- 可直接保存完整 fixture;
- 不需處理商業引擎 license / opaque renderer;
- 適合 repo-based Agent 開發。
OSR core 保持 engine-neutral。
58. Godot Adapter 第一版
不需要改 renderer source。
只需要取得:
- base viewport;
- depth;
- motion 若可取得;
- object mask;
- camera;
- object transform;
- scene revision。
若 motion buffer 第一版取得困難:
可用:
先做 object-level warp fixture。
59. Offline Simulator 先行
真正接 Godot 前,
先做:
SyntheticWorldAdapter
輸出:
- deterministic frames;
- masks;
- motion;
- depth;
- scripted world deltas。
這可以先把:
- PRC;
- truth;
- cloud mock;
- temporal;
- benchmark;
全部驗證。
60. Offline Simulator 不是多餘
它允許:
避免一開始 debug engine / network / AI / compiler 四層。
61. MVP Phase 0|Contract Closure
交付:
- schemas;
- Python dataclasses;
- validation;
- deterministic IDs;
- receipts;
- mocks。
Acceptance:
all schemas round-trip
invalid truth writeback rejected
same input → same plan ID
62. MVP Phase 1|Offline OSR Compiler
交付:
- ObserverState;
- obligations;
- PRC;
- backend capability;
- OSR-IR;
- field solver;
- NO_RERENDER;
- fallback。
Backend 全 mock。
Acceptance:
- mandatory preserved;
- budget closure;
- no-feasible-plan;
- deterministic plan;
- cloud-denied scope never selects cloud。
63. MVP Phase 2|Truth Firewall
交付:
- TruthRecord;
- preserve / generate contract;
- structural verifier;
- writeback firewall;
- truth receipts。
Acceptance:
- generated output cannot mutate world;
- addition / omission fixture rejected;
- base fallback works;
- provenance survives reuse chain。
64. MVP Phase 3|Temporal Reconstruction Offline
交付:
- anchor;
- warp;
- confidence;
- blend;
- disocclusion;
- camera-cut invalidation。
Acceptance:
- one anchor → multiple output frames;
- stale history rejected;
- confidence fallback to base;
- no blank frame。
65. MVP Phase 4|Godot Base Scene
場景:
room
character
reflective object
hazard
background
交付:
- base RGB;
- object masks;
- depth / approximate depth;
- scripted camera;
- scripted object motion;
- scene revision。
66. MVP Phase 5|Real General Cloud Anchor
第一個真 provider 只做:
face / character ROI
不做全畫面。
流程:
Godot ROI
→ export gate
→ cloud image adapter
→ result
→ truth/structure verify
→ local compose
Acceptance:
- real call;
- result receipt;
- latency recorded;
- cost recorded;
- failure preserved;
- invalid result fallback。
67. MVP Phase 6|APR Runtime Integration
不是 mock APR。
真接:
WorldState
EvidenceStore
Goal
Budget
Policy
PluginRegistry
新增 OSR adapter。
Acceptance:
- fresh stable scene → NO_RERENDER;
- uncertainty → refresh;
- risk event → mandatory refresh;
- budget pressure → preferred quality drops before minimum。
68. MVP Phase 7|Benchmark Harness
實作 OSR-07 最小版:
B0 Local Raster
B2 Dynamic Quality
B3 Local Neural/mock
B4 Naive Cloud
O0 OSR Local
O1 OSR Cloud Anchor
測:
- quality;
- structure;
- temporal;
- latency;
- cost;
- cloud call efficiency;
- truth violations;
- reuse。
69. MVP Phase 8|Evidence Package
每次 run:
run_manifest.json
plan_receipts.jsonl
execution_receipts.jsonl
truth_receipts.jsonl
metrics.csv
frames/
summary.md
70. MVP Phase 9|E1 Decision
最後只回答:
若沒有:
停止擴張。
先修 architecture。
若有:
才進 dedicated neural renderer / latent / higher-frequency cloud。
71. 本地 3070 驗證策略
RTX 3070 不當:
large neural renderer target
而當:
base render
buffer producer
temporal reconstructor
compositor
local verifier
第一代測:
能否找到相對於:
新的 Pareto operating point。
72. General Cloud 第一代用途
不要測:
60 cloud calls/sec
先測:
1 anchor / 2–10 sec
scene anchor
dialogue face
photo mode
slow camera
73. 最適合 MVP 的場景
優先:
例如:
- 對話;
- RPG;
- strategy;
- photo mode。
不先做高速 FPS。
74. 第一個 Demo
建議:
OSR Dialogue Lab
場景:
- NPC 站在室內;
- 玩家可慢速繞行;
- 光源可移動;
- NPC 有 face region;
- 背景 static;
- 一個 hazard toggle。
Modes:
Local Base
Naive Cloud
OSR Cloud Anchor
OSR Offline/Network-Failure
75. Demo UI
顯示:
FPS
GPU frame time
current renderer per region
cloud request state
anchor age
truth class
verification status
cloud cost
network latency
NO_RERENDER count
76. Developer Overlay
Region 顏色:
raster green
reuse blue
local neural purple
cloud red
semantic white
verification-failed orange
stale gray
僅 debug。
77. Benchmark Fixtures
至少:
static
dialogue
slow orbit
fast motion
disocclusion
hazard
camera cut
network jitter
cloud outage
truth addition
truth omission
78. 指標 v0.1
必要:
第一版不必全做 LPIPS / FLIP。
最小:
- SSIM / simple perceptual;
- mask IoU;
- identity fixture;
- warp error;
- p50/p95 latency;
- cloud calls;
- bytes;
- WVR;
- reuse;
- fallback。
之後再加 FLIP / LPIPS。
79. Cloud Cost
所有 provider adapter 必須輸出:
estimated_cost
pricing_basis_date
usage
若未知:
estimated_cost = null
不可猜。
80. Network Metrics
記:
- request submit;
- first response;
- complete response;
- validation;
- commit。
可分:
81. Tail Latency
至少:
如果 sample 太少:
明確標:
insufficient sample count
82. Failure Semantics
統一 enum:
CAPABILITY_UNAVAILABLE
NO_FEASIBLE_PLAN
EXPORT_DENIED
TIMEOUT
REMOTE_ERROR
INVALID_RESPONSE
STALE_RESPONSE
TRUTH_VIOLATION
VERIFICATION_FAILED
BUDGET_EXCEEDED
LOCAL_BACKEND_FAILED
83. Failure 不自動 Retry
v0.1:
APR 可下一輪重新決策。
84. Fallback Semantics
fallback 只能:
- 滿足 minimum;
- 或清楚標記 DEGRADED;
- 否則 BLOCKED。
不可 silent success。
85. Runtime State
Scope state:
STABLE
SUSPECT
REFRESHING
VERIFYING
VISIBLE
DEGRADED
BLOCKED
86. Cloud Request State
CREATED
SUBMITTED
RETURNED
STALE
REJECTED
VERIFIED
COMMITTED
FAILED
87. Credential Boundary
Credential:
- 不進 OSR-IR;
- 不進 receipts;
- 不進 benchmark artifacts;
- 不進 engine payload。
由 provider adapter lazy read。
沿用 APR plugin 的安全習慣。
88. Privacy
ExportPolicy:
{
"allow_pixels": true,
"allow_depth": true,
"allow_semantics": false,
"allow_object_ids": false,
"allow_text": false
}
89. Sensitive Scope
Scope 可以標:
private
restricted
public
compiler hard gate cloud。
90. Cloud Provider 不可變成 Source Authority
Cloud result 永遠:
truth_class = inferred_appearance | generated_appearance
world_write_authority = false
第一代固定。
91. Deterministic Offline Tests
CI 不呼叫真 cloud。
provider 用 fixture。
真 provider test:
explicit opt-in
call cap
cost cap
no retry
92. Test Layers
Unit
- schema;
- compiler;
- truth;
- cache;
- revision;
- fallback。
Integration
- APR adapter;
- engine simulator;
- cloud mock;
- temporal。
Live
- real engine;
- real provider;
- benchmark。
93. Golden Test 01
Stable scene:
input unchanged
→ NO_RERENDER
→ no cloud call
→ output remains valid
94. Golden Test 02
Face close-up:
preferred depth ↑
→ cloud candidate selected
→ base visible first
→ cloud returns
→ identity verify
→ late upgrade
95. Golden Test 03
Cloud timeout:
base remains visible
→ cloud node fails
→ no blank frame
→ receipt records timeout
96. Golden Test 04
World revision changes during cloud request:
response stale
→ visible commit denied
97. Golden Test 05
Cloud adds forbidden object:
structural verify fail
→ fallback base
→ truth violation receipt
98. Golden Test 06
Hazard microchange:
pixel delta small
significant delta high
→ mandatory refresh
99. Golden Test 07
GPU pressure:
preferred local enhancement drops
minimum remains
100. Golden Test 08
Export denied:
cloud quality high
but privacy policy denies
→ cloud never called
101. Golden Test 09
Camera cut:
temporal cache invalidated
→ base immediate
→ optional new cloud anchor
102. Golden Test 10
Same deterministic input:
same semantic OSR plan
103. Acceptance Gate v0.1
Phase 0–3:
100% deterministic offline tests pass
truth writeback violations = 0
stale visible commits = 0
Phase 4–6:
engine base continuity = 100%
cloud failure blank-frame = 0
export-denied calls = 0
Phase 7:
至少找到一個 controlled domain:
使 OSR 有 practical Pareto improvement。
104. 不預先保證 存在
如果不存在:
這是有效研究結果。
不要為了「OSR 必須成功」改 threshold。
105. E0 Gate
只證明:
- architecture runs;
- contracts work;
- failure semantics work。
106. E1 Gate
至少:
- 一個 domain;
- 一個 baseline;
- practical gain;
- hard constraints pass。
107. E2 Gate
未來才要求:
- interactive p95;
- temporal;
- tail jitter;
- sustained session。
108. E3 Gate
未來:
- long duration;
- production security;
- provider SLA;
- real users;
- regression;
- deployment。
109. GitHub / Repository Strategy
建議新 repo 不塞進 APR。
原因:
而:
APR 作 dependency / sibling project。
110. Repository Names
建議優先:
osr-runtime
或:
observer-sovereign-rendering
避免只叫:
renderer
失去研究識別。
111. README 第一頁要寫的限制
必須明確:
Research prototype.
Not a DLSS replacement.
Not a production cloud-gaming stack.
Not proven to improve all workloads.
General cloud image APIs are anchor backends, not assumed real-time renderers.
112. License / External Assets
MVP scene 使用:
- 自製;
- CC0;
- permissive;
- engine primitives。
不要直接拿商業遊戲 asset 上傳 cloud。
113. Cloud Provider Test Data
第一代 cloud live test 使用:
- 自製 character;
- synthetic scene;
- no personal data;
- no proprietary assets。
114. Security
Cloud output 永遠 untrusted input。
必須驗:
- MIME;
- dimensions;
- size;
- request ID;
- revision;
- hash;
- schema。
不執行 remote code。
115. Resource Limits
每個 cloud request:
max input bytes
max output bytes
timeout
call cap
budget cap
全部 explicit。
116. Logging
不 log:
- credentials;
- secret pixels 若 privacy 禁止;
- raw provider error body 若可能含敏感內容。
log:
- error class;
- latency;
- request ID;
- bounded provider code;
- byte counts;
- cost metadata。
117. Benchmark Evidence Ready
所有結果都可回到:
world fixture
plan receipt
execution receipt
truth receipt
frame
metric
這是 OSR-07 的核心。
118. 白皮書到 MVP 的交接句
本白皮書完成後,本地 Agent 不需要再問:
OSR 是什麼?
它只需要從:
開始建立 contracts 與 deterministic simulator。
119. 第一個實作順序
contracts
→ dataclasses
→ mock registry
→ compiler
→ truth firewall
→ offline temporal
→ simulator benchmark
→ engine adapter
→ real cloud anchor
→ APR integration
→ benchmark
不要反過來先接 cloud API。
120. 第一個 Commit 建議
feat: bootstrap OSR contracts and deterministic offline runtime
內容只要:
- package;
- schemas;
- IDs;
- errors;
- mock backend;
- tests。
121. 第二個 Commit
feat: add manifestation obligations and perceptual render compiler
122. 第三個 Commit
feat: add truth firewall and provenance receipts
123. 第四個 Commit
feat: add temporal anchor reconstruction simulator
124. 第五個 Commit
feat: add APR adapter and no-rerender policy
125. 第六個 Commit
feat: add Godot reference engine adapter
126. 第七個 Commit
feat: add opt-in general cloud anchor backend
127. 第八個 Commit
feat: add OSR benchmark harness and E1 acceptance report
128. Runtime Invariants
第一代至少固定以下 invariants。
I1
I2
I3
I4
I5
I6
I7
是合法 action。
I8
I9
I10
129. 第一代性能假說
不把任何數字寫成保證。
只測:
在可接受成本內。
130. External Technical Context
2026 年 DLSS 5 已公開展示:
- real-time generative neural rendering;
- engine frame grounding;
- color / motion vectors;
- surface / lighting attributes;
- deterministic output;
- temporal stability;
- traditional rendering + neural appearance。
OSR 不複製其模型。
OSR 利用這個產業進展確認:
OSR 的獨立研究問題則是:
131. Long-Term Architecture
若 MVP 成立,後續可以:
v0.2
- stronger temporal;
- FLIP / LPIPS;
- better cost model;
- multiple cloud adapters。
v0.3
- latent return;
- local decoder;
- persistent cloud session;
- speculative anchor。
v0.4
- ray / path backend;
- LoD/reuse;
- higher-frequency cloud。
v0.5
- dedicated neural renderer;
- engine-native GPU path;
- E2 interactive benchmark。
v1.0 candidate
只有在:
證據成熟後考慮。
132. 與未來硬體的關係
如果未來 local GPU 足夠強:
OSR 不失效。
Compiler 只會把:
因為:
133. 與未來 Cloud 的關係
若 cloud latency 更低:
OSR 也可提高:
架構不需要重寫。
134. 與 AI Agent 的關係
AI agent observer 不一定需要 RGB。
OSR 未來可輸出:
- semantic carrier;
- object graph;
- state;
- task view。
同一 world 同時:
但:
共享。
135. 與 AR / XR 的關係
OSR 的:
- observer;
- manifestation obligation;
- truth class;
- reversible mediation;
- local/cloud;
- reprojection;
都可自然延伸 XR。
但第一代不展開。
136. 整合技術定義
本文將 OSR Runtime 定義為:
其中:
- :world/evidence interface;
- :observer;
- :manifestation obligations;
- :compiler;
- :OSR-IR / plan;
- :heterogeneous rendering field;
- :truth / provenance;
- :distributed local-cloud runtime;
- :evaluation / evidence layer。
137. 一個完整 Step
1. engine advances world
2. APR updates world/evidence/budget
3. OSR builds observer state
4. obligations are updated
5. significant delta determines affected scopes
6. PRC patches OSR-IR
7. local base renders
8. optional cloud requests are submitted
9. temporal path maintains high-frequency display
10. cloud result returns
11. truth / revision / structure verification
12. compatible result late-upgrades region
13. receipts are emitted
14. benchmark/evidence records metrics
138. 這就是第一代 MVP
第一代不需要:
- 重新發明 neural renderer;
- 重新發明 path tracing;
- 等全新 GPU。
它需要的是:
139. 最終工程判定
完成 v0.1 後只問三件事:
139.1 架構能不能跑?
139.2 有沒有至少一個有用工作域?
139.3 值不值得繼續投入更低延遲 dedicated neural backend?
140. Go 條件
至少:
- E0 全通過;
- 一個 ;
- truth / privacy invariants pass;
- cloud sparse 比 naive cloud 有意義;
- local base continuity;
- compiler overhead 未抵銷收益。
141. No-Go / Redesign 條件
如果:
- cloud output 不可穩定 preserve structure;
- temporal reconstruction 無法維持;
- compiler overhead 太大;
- general cloud 成本不合理;
- local-only 已全面支配;
則:
不硬推 cloud。
OSR 仍可保留:
142. 最終結論
OSR 技術白皮書的真正核心不是:
「我們有一套新的 AI renderer。」
而是:
這個 contract 將:
- observer intent;
- world state;
- APR;
- renderer diversity;
- local/cloud;
- temporal reuse;
- truth authority;
- verification;
- evidence;
放入同一個可執行系統。
對第一代 MVP 而言,最值得驗證的也不是「AI 是否能畫出最漂亮的一幀」,而是:
如果答案是 yes,
OSR 就得到 E1 級實證。
如果答案是 no,
OSR-07 已經規定怎麼讓它失敗,而不是靠理論敘事逃避。
因此,本白皮書之後的正確下一步不是再寫 OSR-08。
而是:
參考文件
OSR 第一卷
- OSR-00 — 從萬能光到觀察者主權渲染:理論譜系重構與研究綱領。
- OSR-01 — 我們到底想看什麼?觀察者、目的與顯現場。
- OSR-02 — 從射線到觀察探針:廣義觀察算子與感知渲染編譯器。
- OSR-03 — 異質渲染場:Raster、Ray、Neural、Semantic 與 Reuse 的統一選擇。
- OSR-04 — APR 治理的自適應渲染:感知預算、表示深度與動態重算。
- OSR-05 — 世界真值與顯現真值:生成式渲染中的一致性、幻覺與表示主權。
- OSR-06 — 分散式 AI 原生渲染:Local World、Cloud Intelligence 與 Temporal Reconstruction。
- OSR-07 — 可反證的觀察者渲染:品質、延遲、成本、一致性與結構保真。
EveMissLab Runtime
- APR — Adaptive Perceptual Reading / Agentic Perception Runtime. https://github.com/kakon77777-commits/APR
- MRMIC/NVCL. https://github.com/kakon77777-commits/MRMIC_NVCL
- PNCW — Projection-Native Computational World. https://github.com/kakon77777-commits/PNCW
External Technical Context
- NVIDIA ADLR. DLSS 5: Generative Neural Rendering. 2026. https://research.nvidia.com/labs/adlr/DLSS5/
- NVIDIA GeForce. DLSS 5 3D-Guided Neural Rendering. 2026. https://www.nvidia.com/en-us/geforce/news/dlss-5-3d-guided-neural-rendering/
- NVIDIA CloudXR SDK. Network Setup / Network Quality. https://docs.nvidia.com/cloudxr-sdk/
- Perel, O., et al. TRON: Tracing Rays to Orchestrate a Neural Renderer for 3D Gaussian Reconstructions. 2026. https://research.nvidia.com/labs/sil/projects/tron/
- Wang, Y.-C., et al. Real-Time Level-of-Detail Rendering with ReSTIR. SIGGRAPH 2026.
- Hong, P., et al. Multi-Layer Reservoir Splatting for Temporal Reuse under Disocclusion. SIGGRAPH 2026.
- Andersson, P., et al. FLIP: A Difference Evaluator for Alternating Images. HPG 2020.
附錄 A|MVP 最短交接
GOAL:
Build E0→E1 OSR prototype.
HARD INVARIANTS:
- generated output cannot mutate world
- cloud failure cannot blank minimum view
- stale cloud result cannot visible-commit
- denied scope cannot leave device
- verification fail must fallback/block
- deterministic offline plan test must pass
PHASE ORDER:
0 contracts
1 compiler
2 truth firewall
3 offline temporal
4 Godot scene
5 real cloud anchor
6 APR integration
7 benchmark
8 evidence package
9 E1 decision
DO NOT:
- build full engine
- build custom path tracer first
- optimize 60 FPS cloud first
- add latent protocol before pixel-anchor MVP
- hide failures
附錄 B|一句話版本
OSR v0.1 的工程實作不是打造另一個 DLSS,而是建立一個 APR 驅動、真值可治理、local-first、cloud-optional 的渲染控制 Runtime:先由觀察者與世界狀態產生顯現義務,再由 PRC 編譯成異質 Render Plan,本地永遠維持最低可用世界畫面,雲端只對高價值 scope 提供稀疏 appearance intelligence,本地以 temporal reconstruction 擴展並驗證後才 commit,最後用 OSR-07 的 benchmark 決定它到底值不值得繼續。