Visual Theory Evolutionary Knowledge Runtime
視覺理論演化知識執行層技術白皮書
——將 VUSD 的意圖、理由、共享域、反事實與演化理論落地為 AI 可操作的版本化知識 Runtime
VUSD Series — Technical Whitepaper 01
作者: Neo.K
機構: EveMissLab/一言諾科技有限公司
版本: v0.1
日期: 2026-08-31
定位: Technical Whitepaper / Knowledge Runtime / EveAtelier Integration
相依理論: VUSD Paper 01–05
目標系統: EveAtelier / SEDB-Visual / Style Atlas / AADS / RABCL / MRMIC-NVCL / RVGR
摘要
VUSD 五篇理論建立了一個從視覺作品走向視覺理解的完整理論鏈:
其中:
- :Creator / User Intent;
- :Context / Constraint;
- :Visual Decision;
- :Perceptual / Relational Mechanism;
- :Understanding Shared-Domain State;
- :Observer;
- :Experienced Meaning。
VUSD 同時區分:
其中:
- :Creator Rationale;
- :Contextual Cause;
- :Functional Rationale;
- :Evolutionary Rationale。
Paper 04 再加入:
並以反事實作為視覺理解的重要可操作檢驗。
Paper 05 則將整個模型時間化:
並明確拒絕把目前的 Operator Set 凍結成永久視覺本體。
因此本白皮書提出:
Visual Theory Evolutionary Knowledge Runtime
VTEKR
其定位不是:
藝術百科全書
風格標籤資料庫
prompt library
固定美術規則表
而是:
可由 AI 與人類共同讀寫、可追溯證據、可版本化、可反事實驗證、可演化又不任意漂移的視覺理論知識執行層。
核心架構不變量為:
以及:
1. Runtime 的問題不是「存更多知識」
一般知識庫可以保存:
印象派是什麼
某畫家常用什麼顏色
三分法是什麼
互補色是什麼
但 VUSD 所需的 Runtime 必須回答更複雜的查詢:
這個畫家在哪個時期開始反覆使用這種構圖?
這是作者自己說的,還是後世研究者的解釋?
這個視覺決策在什麼 domain 中通常提高 salience?
有哪些反例?
如果把它移除,預期哪些 Shared-Domain State 會改變?
2026 年的 AI evaluator 與 2030 年 evaluator 是否仍給相同判斷?
這個 operator 是不是已經 split 成兩個?
目前這個 style label 在不同年代的意義有沒有 drift?
因此:
2. VTEKR 的責任邊界
VTEKR 負責:
知識實體
證據來源
理論關係
版本
時間作用域
Observer Scope
Operator Lifecycle
Counterfactual Record
Promotion / Deprecation
Migration
VTEKR 不直接負責:
像素編輯
Canvas state
Provider execution
GPU generation
最終人類審美決定
因此:
3. 與 EveAtelier 既有權限層對齊
建議權限分工:
AADS
= Visual Intelligence / planning authority
SEDB-Visual + VTEKR
= Visual knowledge / theory authority
RABCL
= Workflow compiler
MRMIC
= Persistent canvas authority
NVCL
= Observation-action runtime
RVGR
= Reflexive generation runtime
Provider
= Concrete execution backend
所以:
4. Canonical Runtime Object Families
VTEKR v0.1 至少需要十二個核心 Object Family:
ArtifactRecordCreatorRecordIntentRecordContextRecordVisualDecisionRecordRationaleRecordSharedOperatorRecordObserverRecordMeaningProjectionRecordEvidenceRecordCounterfactualRecordTheoryRecord
外加:
OperatorVersionRecordMigrationRecordPromotionRecordDisagreementRecord
5. Source / Observation / Analysis / Theory 必須分離
最重要的資料不變量:
例如一封畫家書信:
SOURCE
其中一句:
我故意把背景壓暗,讓人物更突出。
是:
SOURCE CLAIM
系統整理成:
CreatorIntentObservation
再建立:
IntentRecord
若跨作品發現:
dark background 在 portrait domain 中常提高 figure-ground separation。
才是:
Functional Rationale / Theory Candidate
不能直接把來源文字塞成普適理論。
6. ArtifactRecord
建議:
{
"artifactId": "artifact:...",
"title": "...",
"creatorRefs": [],
"time": {},
"medium": [],
"sourceRefs": [],
"rights": {},
"semanticRefs": [],
"revision": 1
}
VTEKR 不一定保存 image bytes。
可只保存:
asset URI
hash
metadata
semantic graph ref
7. CreatorRecord
{
"creatorId": "creator:...",
"names": [],
"activePeriods": [],
"roles": [],
"sourceRefs": [],
"trajectoryRefs": []
}
核心:
不是固定風格向量。
8. IntentRecord
{
"intentId": "intent:...",
"artifactRef": "artifact:...",
"actorRef": "creator:...",
"statement": "...",
"status": "DOCUMENTED",
"evidenceClass": "EXPLICIT_AUTHOR_STATEMENT",
"sourceRefs": [],
"timeScope": {},
"confidence": 0.94,
"alternatives": []
}
9. Intent Status
標準:
UNKNOWN
INFERRED
SUPPORTED
DOCUMENTED
DISPUTED
不能用:
TRUE
FALSE
簡化全部作者意圖。
10. Intent Evidence Classes
沿用 Paper 02:
EXPLICIT_AUTHOR_STATEMENT
AUTHOR_NOTE
AUTHOR_LETTER
AUTHOR_INTERVIEW
COMMISSION_BRIEF
DESIGN_SPEC
CONTEMPORARY_DOCUMENT
HISTORICAL_DOCUMENT
SCHOLARLY_INTERPRETATION
HUMAN_ANALYST_INFERENCE
MODEL_INFERENCE
UNKNOWN
11. ContextRecord
{
"contextId": "context:...",
"artifactRef": "artifact:...",
"medium": [],
"technology": [],
"historical": [],
"cultural": [],
"economic": [],
"political": [],
"audience": [],
"evidenceRefs": []
}
12. VisualDecisionRecord
{
"decisionId": "decision:...",
"artifactRef": "artifact:...",
"operatorRef": "visual.op....",
"targetRef": "region:...",
"parameters": {},
"relations": [],
"observationRefs": [],
"confidence": 0.83
}
這一層只回答:
做了什麼。
不應偷偷放入:
為什麼。
13. RationaleRecord
{
"rationaleId": "rationale:...",
"type": "FUNCTIONAL",
"fromRefs": [],
"toRefs": [],
"statement": "...",
"evidenceRefs": [],
"counterevidenceRefs": [],
"scope": {},
"confidence": 0.66,
"status": "PROVISIONAL"
}
14. 四類 Rationale
CREATOR
CONTEXTUAL
FUNCTIONAL
EVOLUTIONARY
對應:
15. SharedOperatorRecord
{
"operatorId": "vusd.shared.reveal_conceal",
"currentConceptVersion": "0.1",
"class": "RELATIONAL",
"status": "ACTIVE",
"definition": "...",
"inputSchemaRef": "...",
"outputSchemaRef": "...",
"observerConditioned": false,
"scope": {},
"evidenceRefs": [],
"lineageRefs": [],
"versionRefs": []
}
16. Operator Class
Paper 03 的四類:
STRUCTURAL
PERCEPTUAL
RELATIONAL
CONTEXT_CONDITIONED
另可有:
COMPOSITE
DOMAIN_SPECIFIC
META
17. Bootstrap Operator Namespace
例如:
vusd.shared.attention
vusd.shared.salience
vusd.shared.contrast
vusd.shared.distance
vusd.shared.reciprocity
vusd.shared.reveal_conceal
vusd.shared.uncertainty
vusd.shared.rhythm
vusd.shared.dominance
vusd.shared.tension.directional
vusd.shared.tension.relational
18. Operator Identity 與 Label 分離
核心:
因此:
{
"operatorId": "vusd.shared.reciprocity",
"labels": {
"zh-TW": "互惠指向",
"en": "Reciprocity"
}
}
未來可換名稱,
不破壞 identity。
19. OperatorVersionRecord
{
"operatorId": "vusd.shared.reciprocity",
"conceptVersion": "0.2",
"previous": "0.1",
"definition": "...",
"scopeChange": "...",
"evidenceDelta": [],
"migrationNotes": []
}
20. Concept Version ≠ Runtime Version
分開:
concept_version
schema_version
runtime_version
不應共用一個版本號。
21. ObserverRecord
{
"observerId": "observer:...",
"type": "HUMAN_PROFILE",
"culture": [],
"time": {},
"expertise": [],
"genreFamiliarity": [],
"task": [],
"preference": {},
"capabilities": {}
}
22. AI Observer
{
"observerId": "observer:ai:gpt-...",
"type": "AI_MODEL",
"provider": "...",
"model": "...",
"version": "...",
"promptSchema": "...",
"evaluationSchema": "..."
}
因此同一模型不同版本:
不同 ObserverRecord
或具有 lineage 的 observer trajectory。
23. Observer ≠ Human-only
VTEKR 保持:
但也不宣稱:
AI has human phenomenal experience
24. MeaningProjectionRecord
{
"projectionId": "projection:...",
"artifactRef": "artifact:...",
"observerRef": "observer:...",
"sharedStateRef": "sharedstate:...",
"meanings": [
{
"label": "mystery",
"weight": 0.63
}
],
"source": "MODEL_PREDICTION",
"confidence": 0.58
}
25. Projection Source
HUMAN_REPORT
HUMAN_ANALYST_INFERENCE
MODEL_PREDICTION
SCHOLARLY_INTERPRETATION
HISTORICAL_RECORD
26. SharedDomainState
{
"sharedStateId": "sharedstate:...",
"artifactRef": "artifact:...",
"regionRef": null,
"operatorStates": {
"vusd.shared.salience": 0.82,
"vusd.shared.reciprocity": 0.71
},
"extractorRef": "observer:ai:...",
"revision": 4
}
27. Shared State 不要求全部 scalar
有些 operator 應是:
scalar
vector
categorical
graph
distribution
例如:
direction graph
grouping graph
reveal/conceal map
不能全部硬壓成:
28. EvidenceRecord
{
"evidenceId": "evidence:...",
"type": "SCHOLARLY_INTERPRETATION",
"sourceRef": "source:...",
"supports": [],
"contradicts": [],
"time": {},
"scope": {},
"confidence": null
}
Evidence 本身不需要假裝有「真實度分數」。
29. Evidence Provenance ≠ Confidence
核心:
模型可:
confidence 0.95
但若 evidence 是:
MODEL_INFERENCE
仍不能升成:
DOCUMENTED AUTHOR INTENT
30. TheoryRecord
{
"theoryId": "theory:...",
"statement": "...",
"status": "PROVISIONAL",
"scope": {},
"supportingEvidenceRefs": [],
"counterEvidenceRefs": [],
"counterfactualRefs": [],
"operatorRefs": [],
"version": "0.1"
}
31. Theory Status
CANDIDATE
PROVISIONAL
ACTIVE
DOMAIN_ACTIVE
DISPUTED
HISTORICAL
DEPRECATED
32. DisagreementRecord
若:
Theory A
contradicts
Theory B
不要消掉其中一個。
{
"disagreementId": "disagreement:...",
"theoryRefs": ["theory:A", "theory:B"],
"relation": "CONTRADICTS",
"scopeOverlap": {},
"resolutionStatus": "UNRESOLVED"
}
33. Plurality by Design
核心:
可以保存:
形式主義解釋
歷史解釋
市場解釋
符號解釋
創作者自述
34. CounterfactualRecord
{
"counterfactualId": "cf:...",
"artifactBefore": "artifact:A",
"artifactAfter": "artifact:B",
"intervention": {},
"minimalClosure": [],
"lockedSemantics": [],
"expectedDeltaShared": {},
"observedDeltaShared": {},
"observerProjectionRefs": [],
"evidenceType": "GENERATED_VARIANT",
"status": "OBSERVED"
}
35. Counterfactual Evidence Type
MODEL_SIMULATION
GENERATED_VARIANT
MANUAL_EDIT
CONTROLLED_EXPERIMENT
HUMAN_AB_TEST
HISTORICAL_VARIANT
36. Prediction 與 Observation 分離
所以:
expectedDelta
與:
observedDelta
必須分欄。
37. Counterfactual Prediction Error
可作為 rationale calibration evidence。
38. Counterexample 是一級資料
Runtime 必須保存:
支持案例
反例
失敗 domain
observer disagreement
不是只存「成功規則」。
39. Persistent Residual
對 theory:
若持續偏大:
可能 operator 缺失
scope 太寬
theory 定義錯
observer model 漂移
40. Candidate Operator Discovery
觸發:
persistent residual
pattern cluster
cross-model recurrence
human analyst proposal
counterfactual discovery
建立:
CANDIDATE_OPERATOR
41. Candidate Operator 最低資料
{
"candidateId": "candidateop:...",
"temporaryLabel": "operator_candidate_7831",
"signature": {},
"exampleRefs": [],
"counterexampleRefs": [],
"proposedScope": {},
"proposedRelations": [],
"origin": "AI_DISCOVERED"
}
42. Operator Discovery ≠ Promotion
硬規則:
43. PromotionRecord
{
"promotionId": "promotion:...",
"targetRef": "candidateop:...",
"fromStatus": "CANDIDATE",
"toStatus": "PROVISIONAL",
"evidenceRefs": [],
"reviewRefs": [],
"decision": "PROMOTE",
"actor": "knowledge-governance:..."
}
44. Promotion Gate
至少:
distinctiveness
definition stability
evidence sufficiency
counterfactual utility
scope clarity
counterexamples considered
compatibility
review
45. Self-Proposal ≠ Self-Grant
AI 可以:
propose operator
propose theory
propose scope update
但:
46. Capability ≠ Authority
沿用:
模型「能」修改理論,
不代表它「有權」直接修改 canonical knowledge。
47. Theory Mutation Authority
建議獨立 Authority:
KnowledgeGovernor
可由:
human review
policy
multi-model evidence
automated tests
共同驅動。
48. Ordinary Runtime Write Path
普通 AADS / Evaluator:
read canonical theory
write observations
write candidate hypotheses
write counterfactual results
不可:
直接覆寫 canonical operator
49. Canonical Knowledge Commit Gate
可類比 MRMIC World write:
TheoryCommitGate
作為 canonical knowledge 的 ordinary writer。
50. Candidate ≠ Canonical
所有:
new theory
new operator
new migration
先進 Candidate space。
51. Knowledge Revision
每個 canonical object 有:
revision
Mutation request 必須帶:
expectedRevision
52. Stale Revision Fail Closed
避免兩個 AI 同時把 operator 改成不同定義。
53. Concept Revision 與 Evidence Append 分離
新增 evidence:
append evidence
不一定要:
bump concept version
只有:
definition
scope
semantic relation
改變時才需要 concept version。
54. Operator Lifecycle State Machine
CANDIDATE
↓
PROVISIONAL
↓
ACTIVE
↘
DISPUTED
↘
HISTORICAL
↘
DEPRECATED
另有:
FORKED
MERGED
55. Fork
{
"operation": "FORK",
"source": "vusd.shared.tension.v0.1",
"targets": [
"vusd.shared.tension.directional",
"vusd.shared.tension.relational"
]
}
56. Merge
{
"operation": "MERGE",
"sources": ["operator:A", "operator:B"],
"target": "operator:C"
}
57. Deprecation
Deprecated object:
仍可查
仍可重現歷史分析
不建議新分析預設使用
58. Historical ≠ Deprecated
HISTORICAL
表示:
在特定歷史範圍仍有描述價值。
DEPRECATED 則表示:
概念模型本身已不建議使用。
59. MigrationRecord
{
"migrationId": "migration:...",
"fromRef": "operator:old",
"toRefs": ["operator:newA", "operator:newB"],
"mapping": "PARTIAL",
"unresolved": [],
"strategy": "REVIEW_REQUIRED"
}
60. Unresolved Mapping
若不能安全轉換:
UNRESOLVED
而不是猜。
61. Unknown-field Preservation
Schema 新版本增加欄位時,
舊 Runtime 應盡量:
read known fields
preserve unknown fields
round-trip unknown data
避免資料被舊 client 洗掉。
62. Extension Namespace
建議:
vusd.core.*
vusd.shared.*
vusd.domain.character.*
vusd.domain.film.*
vusd.experimental.*
vendor.*
lab.*
63. Domain Layer
例如:
vusd.domain.character.bodyline_salience
vusd.domain.character.gaze_reciprocity
vusd.domain.character.garment_reveal_conceal
建立在 Shared Core 上。
64. Domain Constraint ≠ Ontological Closure
Runtime schema 必須支持:
new domain
new operator family
而不是寫死:
character
landscape
film
65. Artist Model
未來 Artist Archive:
{
"artistId": "creator:...",
"trajectory": [
{
"timeScope": {},
"decisionDistributionRefs": [],
"sharedStateDistributionRefs": [],
"intentEvidenceRefs": [],
"rationaleRefs": [],
"historicalContextRefs": []
}
]
}
66. Artist ≠ One Style Vector
核心:
67. StyleProfile
{
"styleProfileId": "styleprofile:...",
"creatorRef": "creator:...",
"timeScope": {},
"surfaceFeatures": {},
"decisionDistribution": {},
"sharedDomainDistribution": {},
"rationaleRefs": [],
"counterfactualBoundaryRefs": []
}
68. Style Atlas Query
未來 Style Atlas 可問:
找同樣 surface 但不同 relational syntax
找同樣 composition rhythm 但不同 palette
找同樣 shared-domain profile 的不同畫家
找某畫家的早期 vs 晚期 operator drift
69. Style Similarity 多層化
不能只用一個 embedding cosine。
70. Style Boundary
Paper 04 的 Counterfactual Boundary:
改哪些東西仍像同系列?
可存:
{
"boundaryId": "boundary:style:...",
"lockedDimensions": [],
"tolerances": {},
"observedFailures": []
}
71. Observer Trajectory
observer v1
→ observer v2
→ observer v3
保存:
model/version
training cut-off
evaluation schema
time
72. Judgment Record
{
"judgmentId": "judgment:...",
"artifactRef": "artifact:...",
"observerRef": "observer:...",
"theorySnapshotRef": "snapshot:...",
"result": {},
"time": "..."
}
73. Theory Snapshot
每次正式 evaluation 可綁:
operator versions
theory versions
observer version
形成可重現:
VisualTheorySnapshot
74. Re-evaluation ≠ Historical Rewrite
新 evaluator 可重評,
但舊 judgment 不刪。
judgment:v1
judgment:v2
並列。
75. Judgment Lineage
{
"previousJudgmentRef": "...",
"changeReason": "observer_model_update"
}
76. Audience Profile
{
"audienceId": "audience:...",
"timeScope": {},
"culture": [],
"market": [],
"genreFamiliarity": [],
"appealProfile": {}
}
77. Appeal 不固定
所以:
Audience Appeal Lock
必須綁特定:
source artifact
audience profile
time scope
78. AADS Query Interface
AADS 可查:
get_visual_rationales(artifact)
get_operator_candidates(target_shared_state)
get_locked_semantics(artifact)
get_counterfactual_history(artifact)
get_style_boundary(style)
get_theory_snapshot(domain)
79. AADS 不直接讀原始資料庫表
建議透過:
VTEKR Query API
避免 AADS 耦合 schema internals。
80. Suggested Query API
resolve_artifact_context()
infer_visual_decisions()
resolve_rationales()
resolve_shared_state()
resolve_observer_projection()
search_counterfactuals()
search_theories()
get_operator_version()
81. Mutation API
普通 client:
append_observation()
append_evidence()
propose_rationale()
propose_operator()
record_counterfactual()
record_judgment()
治理 client:
promote_operator()
fork_operator()
merge_operator()
deprecate_operator()
migrate_knowledge()
82. RABCL Integration
RABCL 可以編譯:
Intent
→ Shared Target
→ Visual Operator Plan
但 RABCL 不應:
改寫 theory ontology
83. RABCL Workflow Node Example
{
"nodeType": "VUSD_SHARED_TARGET",
"target": {
"reciprocity": "retain",
"reveal_conceal": "retain",
"subject_salience": "increase"
},
"theorySnapshotRef": "snapshot:..."
}
84. RVGR Integration
RVGR 在生成中:
observe generation
extract shared-state drift
compare target
propose rewrite
但:
85. RVGR Observation Record
{
"generationRef": "...",
"checkpoint": 12,
"sharedState": {},
"drift": {},
"observerRef": "..."
}
86. RVGR 可以回報 Theory Residual
如果:
expected effect
反覆不成立,
RVGR 可:
propose theory residual
但不能自行 promotion。
87. MRMIC / NVCL Integration
MRMIC 保存:
canonical canvas / document state
NVCL 提供:
observe / act
VTEKR 保存:
why / relation / theory / evidence
因此:
88. Asset State ≠ Knowledge State
圖片刪除、
版本切換,
不一定刪除歷史:
counterfactual evidence
judgment
theory
89. Rights / Provenance
ArtifactRecord 需保留:
rights
source
redistribution status
因為畫家建檔、作品建檔與公開展示不是同一件事。
90. Reference Role
Style Atlas / SEDB-Visual 可沿用:
STYLE_CORE_REFERENCE
IDENTITY_REFERENCE
FACE_REFERENCE
PROPORTION_REFERENCE
POSE_REFERENCE
COSTUME_REFERENCE
COLOR_REFERENCE
LINE_REFERENCE
LIGHTING_REFERENCE
COMPOSITION_REFERENCE
NEGATIVE_REFERENCE
91. Reference ≠ Theory
一張 reference:
可以作 evidence / example
但不應:
自動成為 theory
92. Visual Theory Snapshot
建議:
{
"snapshotId": "snapshot:...",
"domain": "character_art",
"operatorVersions": {},
"theoryVersions": {},
"schemaVersion": "0.1",
"createdAt": "..."
}
93. Deterministic Replay
若:
artifact
observer
snapshot
evaluation schema
固定,
應盡量可重放 evaluation。
生成式 evaluator 仍可能有隨機性,
需記:
seed / sampling settings
若支援。
94. Explanation Record
每個正式 judgment 建議輸出:
Finding
Rationale
Evidence
Scope
Uncertainty
Counterfactual suggestion
而不是只:
score
95. Natural-Language Rationale
VUSD 強調:
所以每個 Operator / Theory 可有:
formal schema
plain-language explanation
examples
counterexamples
historical notes
96. Language Versioning
自然語言 explanation 本身也可版本化。
例如:
explanationVersion
不一定等於 conceptVersion。
97. Multi-language
同一 theory:
zh-TW
en
ja
...
應共享:
theoryId
避免每種語言變不同 ontology。
98. Historical Interpretation Layer
藝術史內容至少:
creator evidence
contemporary interpretation
later scholarship
modern model inference
分欄。
99. Anachronism Flag
ANACHRONISTIC_TERM
MODERN_INTERPRETATION
HISTORICAL_TERM
例如用:
cinematic
描述攝影術之前作品時,
應標:
modern analogy
100. Unknown 必須是一級狀態
以下皆合法:
UNKNOWN_INTENT
UNKNOWN_RATIONALE
UNKNOWN_OPERATOR
UNKNOWN_OBSERVER_EFFECT
UNRESOLVED_MIGRATION
101. Unknown ≠ Null Everywhere
不同:
null
unknown
not_applicable
not_observed
not_supported
應分開。
102. Semantic State Values
建議:
KNOWN
UNKNOWN
UNOBSERVED
NOT_APPLICABLE
DISPUTED
UNRESOLVED
103. Storage Model
VTEKR 適合:
relational core
+
graph relations
+
document/evidence store
+
vector retrieval
不建議只靠一種 DB。
104. Canonical IDs
建議所有實體有穩定 ID:
creator:
artifact:
intent:
context:
decision:
rationale:
operator:
observer:
theory:
counterfactual:
evidence:
105. Graph Relation
核心 relation:
supports
contradicts
derived_from
interprets
causes_candidate
projects_to
observed_in
applies_to
specializes
generalizes
forked_from
merged_from
supersedes
106. Retrieval Layer
AI Query 可以:
semantic retrieval
graph traversal
time filtering
scope filtering
evidence filtering
混合。
107. Search Intent
例如:
為什麼這張角色圖看起來更有危險吸引力?
Runtime 可查:
- artifact decisions;
- Shared-domain states;
- domain theories;
- relevant appeal/tension models;
- observer profile;
- counterfactual evidence。
108. Do Not Let Retrieval Become Truth
Top-1 retrieval:
canonical answer。
AI 還需要:
evidence comparison
scope check
conflict check
109. Theory Applicability Gate
在使用 Theory 前:
domain match?
time scope match?
observer scope match?
medium match?
status active?
counterevidence?
110. Applicability Score
可以:
但不應把 score 當 absolute truth。
111. Conflict-aware Retrieval
若兩個 theory:
contradict
結果中應一起回傳。
112. Evidence-first Response
AI 回答最好:
Known:
...
Strongly supported:
...
Plausible:
...
Alternative:
...
Unknown:
...
113. Operator Discovery Queue
新 pattern:
candidate queue
可排序:
114. Operator Novelty Threshold
避免 AI 無限發明術語。
需要:
minimum recurrence
minimum explanatory gain
minimum distinctness
115. Composite-first Policy
若新候選可以表示為:
優先:
COMPOSITE_OPERATOR
而不是新 primitive。
116. Primitive Promotion Gate
只有:
不可有效分解
具穩定獨立效果
才升:
PRIMITIVE_CANDIDATE
117. Complexity Budget
Knowledge Runtime 也有 complexity cost。
118. Dynamic Canon Layers
建議:
Canonical Core
Domain Canon
Provisional
Experimental
Historical
Deprecated
119. Canonical Core
跨 domain 高穩定,
但仍不是永久 universal。
120. Domain Canon
例如:
character_art
film
UI
各自 canonical operator family。
121. Provisional
有 evidence,
但尚不足升 canonical。
122. Experimental
AI / 人類探索中的新語法。
123. Historical
為重現歷史分析保留。
124. Deprecated
不建議新任務預設使用,
但 lineage 不刪。
125. Theory Snapshot Promotion
正式 production evaluator 應綁:
approved snapshot
而不是直接讀:
latest experimental theory
126. Experimental Theory ≠ Production Default
127. Safety 與 Aesthetic 不應混成一層
沿用既有:
VTEKR 可以保存:
safety constraint
aesthetic effect
audience appeal
但不可互相替代。
128. Character-Specific Expressive Override
Style Core 不應洗掉:
性感
華麗
危險
高飽和
透明材質
等角色特有語義。
所以可保存:
CharacterExpressiveProfile
129. Expressive Profile
{
"identityRef": "...",
"paletteBias": {},
"materialBias": {},
"tensionProfile": {},
"audienceAppealProfile": {},
"styleOverridePolicy": {}
}
130. Style Core ≠ Character Expression
131. Evaluation Pipeline
建議:
132. Explainability Pipeline
再輸出:
what changed
why it matters
which theory applies
what evidence supports it
what remains uncertain
what counterfactual could test it
133. Counterfactual Planning Pipeline
134. Learning Pipeline
135. Theory Update 不應直接在線修改 canonical
生產 runtime:
append evidence
append residual
離線/治理流程:
review
promote
migrate
136. Batch Revalidation
新模型/新 domain 出現時:
revalidate high-impact operators
而非全部盲重算。
137. Revalidation Priority
138. Regression Test
Operator 改版後需測:
known examples
counterexamples
historical cases
style preservation cases
observer cases
139. Knowledge Regression
若新版:
解釋更多
但造成:
大量舊案例錯誤
需阻止 promotion。
140. Rollback
Theory version 可以:
rollback active pointer
但不刪新版失敗版本。
141. Audit Log
所有:
promotion
fork
merge
deprecation
migration
rollback
必須 durable log。
142. Reproducible Research
每篇研究或實驗可引用:
theory snapshot
operator version
observer model
artifact hash
提高重現性。
143. Privacy / Proprietary Works
私人作品可建立:
private semantic record
但:
raw asset
不必公開。
144. Knowledge Derived from Private Asset
若 theory 來自私人資產:
可保存抽象研究結論
但要保留:
source visibility / rights constraint
145. Public Knowledge Export
輸出時可:
strip private asset refs
keep abstract evidence statistics
但不能偽造來源。
146. Artist Archive Ingestion Workflow
建議:
1. register creator
2. ingest works metadata
3. extract observable decisions
4. bind historical context
5. ingest explicit intent evidence
6. infer candidate rationale
7. extract shared-domain profile
8. run counterfactual tests where possible
9. create trajectory
10. human / AI review
147. 不要求一次全部完成
Artist profile 可以:
metadata_only
pending_visual_analysis
partial_evidence
漸進建立。
148. Unknown Intent 是正常狀態
大量歷史藝術:
intent unknown
不影響:
functional analysis
149. Automated Artist Profiling
AI 可以先做:
candidate decisions
candidate style profile
candidate rationale
全部標:
MODEL_INFERENCE
150. Scholar / Human Review
人類可以:
confirm
reject
refine
add evidence
但也不會把 human inference 自動變史實。
151. Cross-AI Analysis
多個 AI 對同作品分析。
可保存:
model-specific observations
再比較:
152. Consensus ≠ Truth
即使十個 AI 都同意:
作者真的這樣想。
但可提升:
shared-domain relation hypothesis
的穩定度。
153. Cross-Observer Agreement
作為 Sharedness Score。
154. Sharedness 隨時間更新
155. Drift Monitor
定期觀察:
operator meaning drift
observer drift
style drift
audience drift
156. Drift ≠ Automatic Mutation
檢測到 drift:
PROPOSE_UPDATE
而不是:
AUTO_REWRITE_CANON
157. Proposed Meta-Operator Family
vusd.meta.operator_discover
vusd.meta.operator_hypothesize
vusd.meta.operator_promote
vusd.meta.operator_split
vusd.meta.operator_merge
vusd.meta.operator_deprecate
vusd.meta.scope_rebind
vusd.meta.evidence_append
vusd.meta.observer_recalibrate
vusd.meta.snapshot_create
158. Meta-Operator Authority
但只在知識治理域。
159. Security / Governance Boundary
Provider 或外部模型不得:
直接提交 canonical theory mutation
只能:
submit proposal / evidence
160. Untrusted External Evidence
外部 AI / web / user source:
untrusted-by-default
需:
provenance
verification
161. Source Quality 不應全域固定
不同研究問題需要:
不同 source hierarchy
例如作者意圖最重作者來源,
functional effect 可能更重實驗。
162. Evidence Weight by Claim Type
所以:
不是固定:
academic paper > everything
163. Claim Type
AUTHOR_INTENT
HISTORICAL_FACT
FUNCTIONAL_EFFECT
OBSERVER_RESPONSE
EVOLUTIONARY_RATIONALE
164. Claim-aware Evidence Resolver
API:
resolve_evidence(claim_type, claim)
165. Data Separation
推薦物理層:
raw sources
observations
semantic records
canonical theory
experimental theory
audit
分區。
166. Backup / Export
至少支援:
JSONL
graph export
markdown synthesis
snapshot bundle
方便跨 AI 使用。
167. Canonical Markdown Export
每個 theory 可導出:
Definition
Scope
Evidence
Counterevidence
Examples
Counterexamples
Version History
方便人讀。
168. Machine-native Export
同時導出:
JSON / graph
供 AI runtime。
169. Human-readable ≠ Machine-readable
兩者都要。
170. API-level Invariants
VTEKR-I1
VTEKR-I2
VTEKR-I3
VTEKR-I4
VTEKR-I5
VTEKR-I6
VTEKR-I7
VTEKR-I8
VTEKR-I9
VTEKR-I10
VTEKR-I11
VTEKR-I12
171. MVP 範圍
VTEKR 第一個 MVP 不需要完整藝術史。
建議只做:
Character Art Domain MVP
原因:
EveAtelier 已有角色圖實驗
Style Control 已建立
Appeal / Exposure / Tension 已有 domain models
RVGR 可直接使用
172. MVP Core Objects
只實作:
Artifact
VisualDecision
SharedOperator
SharedState
Rationale
Evidence
Observer
Counterfactual
Theory
173. MVP Operator Set
先固定約:
salience
contrast
direction
reciprocity
distance
reveal_conceal
repetition
symmetry
dominance
uncertainty
tension.directional
tension.relational
174. MVP 不聲稱完整
README 必須寫:
Bootstrap operator set.
Not a final visual ontology.
175. MVP Query 1
analyze_artifact(artifact)
輸出:
decisions
shared state
rationale candidates
uncertainty
176. MVP Query 2
compare_artifacts(A, B)
輸出:
ΔDecision
ΔSharedState
possible meaning changes
177. MVP Query 3
predict_counterfactual(A, intervention)
178. MVP Query 4
record_observed_counterfactual(A, B)
校準:
179. MVP Query 5
query_style_boundary(style)
180. MVP Query 6
get_visual_rationale(decision)
帶:
scope
evidence
alternatives
181. MVP Mutation
只允許:
append evidence
append observation
propose rationale
propose operator
Canonical promotion 暫時人工。
182. Phase 1
VTEKR-1 — Canonical Schema
建立:
IDs
objects
revision
evidence relation
183. Phase 2
VTEKR-2 — Bootstrap Operator Registry
12–20 個 operator。
184. Phase 3
VTEKR-3 — Character Art Adapter
接:
Exposure
Tension
Audience Appeal
Style Control
185. Phase 4
VTEKR-4 — Counterfactual Store
記:
prediction
variant
observed delta
186. Phase 5
VTEKR-5 — AADS Query Bridge
讓 AADS:
why / target shared state
可查。
187. Phase 6
VTEKR-6 — RVGR Observation Bridge
生成中回傳:
shared-domain drift
188. Phase 7
VTEKR-7 — Operator Proposal Queue
Residual-driven candidate。
189. Phase 8
VTEKR-8 — Governance / Promotion
加入:
review
promotion
fork
merge
deprecate
190. Phase 9
VTEKR-9 — Artist Trajectory
開始畫家建檔。
191. Phase 10
VTEKR-10 — Cross-domain Expansion
再擴:
film
UI
painting
animation
192. MVP 驗證問題
V1
同一 artifact 多次分析是否能穩定抽取主要 decision?
V2
不同 AI 是否在 selected shared operators 上有可測 agreement?
V3
Counterfactual prediction 是否優於純 caption-based prediction?
V4
加入 Shared-domain diagnosis 後,RVGR repair 是否更精準?
V5
Style-only transformation 是否降低 semantic drift?
V6
Operator Proposal 是否能避免無限術語爆炸?
V7
Theory version migration 是否可重放?
193. 成功標準
第一階段不要求:
AI 理解所有藝術
只要求:
- 資料層不混淆;
- evidence lineage 可追;
- counterfactual 可記;
- operator 有版本;
- observer 有版本;
- theory 可以 competing;
- unknown 可保存;
- AADS / RVGR 可查;
- canonical mutation 有 gate;
- schema 可擴充。
194. 非目標
v0.1 不做:
Universal Beauty Score
Global Artist Ranking
Final Art Ontology
Automatic Art-history Truth Engine
Automatic Canon Mutation
195. 最大風險一:Theory Hallucination
AI 很會:
合理解釋
但:
合理
不等於:
有 evidence
防護:
evidence class
intent status
alternative explanation
UNKNOWN
196. 最大風險二:Ontology Explosion
AI 發明太多 operator。
防護:
composite-first
novelty threshold
complexity budget
promotion gate
197. 最大風險三:Ontology Fossilization
反過來 canonical 太穩,
新藝術無法進入。
防護:
experimental edge
candidate operator
residual monitor
open extension namespace
198. 最大風險四:Observer Collapse
把:
human
AI
culture
era
壓成一個 universal observer。
防護:
ObserverRecord
ProjectionRecord
time scope
199. 最大風險五:Historical Hallucination
把現代術語投射到歷史作者。
防護:
anachronism flag
source lineage
historical vs modern interpretation
200. 最大風險六:Aesthetic Default Convergence
Runtime 的「最佳實務」最後變成:
所有圖都同一種好看
防護:
style boundary
artist trajectory
rare operator preservation
domain canon
observer plurality
201. 最大風險七:AI Theory Self-lock
AI 自己提出 theory,
自己 promotion,
自己用 theory 評估,
最後形成封閉回音室。
防護:
202. 最低治理架構
Proposal
↓
Evidence
↓
Independent Evaluation
↓
Counterfactual Test
↓
Review
↓
Promotion
203. 可選 Multi-model Review
不同模型:
proposer
critic
counterfactual evaluator
角色分開。
但不是 v0.1 必須。
204. Human Authority
某些:
aesthetic preference
historical interpretation
canonical research claim
仍適合 human review。
205. Human ≠ Always Correct
但 human review 也:
有 provenance
有 disagreement
而不是不可追蹤的 final truth。
206. 最終架構
207. 與整個 EveAtelier 的最終關係
完整:
Human / AI Intent
↓
AADS
↓
VTEKR
rationale / theory
↓
RABCL
↓
Operator Plan
↓
MRMIC / NVCL / RVGR
↓
Provider Execution
↓
Artifact
↓
Observer
↓
VTEKR
evidence / residual
↓
Theory Proposal / Update
這形成:
Visual Intelligence Learning Loop
208. 這不是讓 AI 永遠服從舊理論
反而是:
讓 AI 能夠知道現在使用的是哪一套理論、它的證據在哪、它的適用域在哪,以及什麼時候應該懷疑它。
因此:
209. 最終命題
一套 AI-native 美術知識系統不應只是:
更多畫風
更多 artist embeddings
更多 rules
更多 prompt templates
它真正需要的是:
210. 結論
VUSD 五篇理論提出了一個核心問題:
AI 是否能從「看見作品」走向「理解作品為什麼這樣存在」?
VTEKR 則回答:
如果要把這種理解變成真正能被系統使用的能力,我們需要什麼資料結構與治理機制?
答案不是建立一本更厚的美術百科。
而是建立一個:
有來源
有觀察
有推論
有證據
有反例
有 Observer
有版本
有時間
有反事實
有競爭理論
有 Operator Lifecycle
的動態知識執行層。
最重要的不變量仍然是:
所以 VTEKR 必須同時具備兩個看似矛盾的能力:
足夠穩定,讓 AI 今天能工作;
足夠開放,讓明天的新美術不必服從今天的分類。
最終:
才是 VUSD 從理論走向 AI-native Visual Intelligence Runtime 的合理工程形態。
End of VUSD Technical Whitepaper 01 — v0.1