Omphalos / AUSI Runtime v1.0 Roadmap Technical Whitepaper v0.1

從模型原生搜尋到可替換、可驗證、可學習的 AI Search Method Runtime

Codename: Omphalos
Technical Name: AUSI Runtime — AI-Native Unified Search Intelligence Runtime
Document Type: Canonical Technical Roadmap Whitepaper
Version: v0.1
Date: 2026-09-03
Status: Canonical Engineering Roadmap toward v1.0
Reference Repository: kakon77777-commits/ai-web-research
Recommended Future Repository Name: omphalos

0. 文件目的

本文件不是功能清單，也不是市場 Roadmap。

它的目的，是固定 Omphalos / AUSI Runtime 從目前狀態推進到 v1.0 的工程邊界、版本順序、驗收條件與 release discipline，避免未來因新 Provider、新搜尋法、新 Domain Pack 或新研究題目不斷插入，導致 Runtime 永遠停留在 0.x。

本文件從 2026-09-03 起應視為：

Omphalos v1.0 的 Canonical Engineering Roadmap。

若後續沒有正式修訂本文件，任何新功能、Provider、Domain Pack、搜尋方法或研究想法，都不應自動改變 v1.0 主路線。

1. v1.0 的正式定義

Omphalos v1.0 不代表：

• 已實作人類歷史上所有搜尋方法；
• 已接入所有搜尋 API；
• 已完成所有 Domain Pack；
• 已完成經濟研究平台；
• 已完成氣象研究平台；
• 已完成所有 Patent Intelligence 能力；
• 已經達到 AGI 等級自主研究。

v1.0 的正式定義是：

第一個真正可長期使用、Provider 可替換、Policy-aware、Evidence-grounded、Auditable、Replayable、Gap-directed 且具有基礎 Experience Learning 的 AI-native Search Method Runtime。

2. v1.0 核心閉環

Task
↓
Search State
↓
Search Strategy / Planner
↓
Search Method
↓
Provider Binding
↓
Provider State / Routing
↓
Policy / Authorization
↓
Execution
↓
Provider Observation
↓
Acquisition / Evidence
↓
Gap Analysis
↓
Continue / Switch Method / Switch Provider / Stop
↓
Search Receipt
↓
Experience
↓
Future Planner Prior

形式化：

3. 永久架構不變量

從現在到 v1.0，以下規則不得破壞：

Method != Provider
Provider != Surface
Provider != API
Spec != Implementation
CapabilityMatch != BindingExists
Planning != Authorization
UNKNOWN != ALLOW
Retrieved != Verified
Citation != Support
QuoteMatch != SemanticSupport
NotFound != False
Saturation != CompleteRecall
SearchReceipt != ChainOfThought
Learning != SelfAuthorization
ProviderObservation != VerifiedEvidence
Access != Acquisition != Usage
EvidenceQuality != UsageRights

最核心的工程原則：

4. Scope Discipline

從本文件生效後採用：

Release-Driven Development

而不是：

Idea-Driven Development

新的研究想法、搜尋法、Provider 與 Domain 能力，預設先進：

Backlog
Search Method Corpus
Provider Candidate Registry
Experimental Domain Pack

只有符合目前 milestone 的項目才進 Runtime 主線。

5. 當前基礎

截至本 Roadmap 建立時，Omphalos 已具備或已開始具備：

SearchMethodSpec
Search Method Registry
ProviderSpec
Provider Registry
ProviderTopology
MethodBinding
Search Graph AST
Plan validation
ExecutionRuntime
TrustedExecutionRuntime
Policy Registry
UsageEnvelope
ProviderObservation
AcquiredAsset
CandidateEvidence
EvidenceAnchor
VerificationResult
Evidence Ledger
GapProjection
Search Receipt
Brave Search provider
Crossref provider
EPO OPS provider
Grok Web provider
Grok X provider
Gemini Google Search provider
Patent Domain contracts
Source lineage / reverse tracing

因此 v1.0 Roadmap 是收斂與完成，不是從零開始。

6. Version Strategy

v0.2 Provider State & Dynamic Routing
v0.3 Search Method Corpus + Core Method Set
v0.4 Autonomous Search Planner
v0.5 Stopping / Coverage / Saturation
v0.6 Evidence / Provenance Closure
v0.7 Search Receipt / Experience Learning
v0.8 Evaluation / Benchmark Suite
v0.9 API Freeze / Hardening / Release Candidate
v1.0 Final Release Gate

原則上每一版完成前，不開始下一版核心工作。

7. v0.2 — Provider State & Dynamic Routing

Goal

讓 Omphalos 證明：

同一 Search Method 可以依 runtime state 動態切換不同 Provider，而不改變 Method identity。

ProviderState

新增 typed contract：

provider_ref
surface_id
available
healthy
credential_available
quota_remaining
quota_reset_at
estimated_cost
estimated_latency
policy_freshness
runtime_capabilities
model_available
last_checked_at

BindingSelector / ProviderRouter

輸入：

SearchMethod
SearchTask
SearchState
ProviderState[]
RoutingPolicy

輸出：

MethodBinding

Routing Examples

Google-native grounding → Gemini
X / social discourse    → Grok X
General neutral Web     → Brave
Academic / NPL          → Crossref
Patent machine data     → EPO OPS

Provider Degradation

若：

quota exhausted
credential missing
API down
policy stale
model unavailable

Runtime 應重新選 Binding，而不是讓 Search Method 消失。

必須證明

method.lexical_search
↓
Brave unavailable
↓
Grok Web selected
↓
Grok unavailable
↓
Gemini selected

Method identity 不變。

Exit Gate

• ProviderState typed contract
• health / availability state
• credential presence state
• quota / cost / latency state
• deterministic routing
• provider preference
• fallback
• same-method provider substitution test
• routing decision enters Receipt
• no credential leakage

8. v0.3 — Search Method Corpus + Core Method Set

Goal

建立真正的：

Search Method Space

而不是只增加 Provider。

Search Method Corpus

每個方法至少記：

method_id
canonical_name
aliases
history
domain
goal
representation
direction
interaction_mode
input_contract
output_contract
strengths
failure_modes
preconditions
provider_requirements
composition_rules
references
implementation_status
validation_status

Method Lifecycle

DOCUMENTED
EXPERIMENTAL
EXECUTABLE
VALIDATED
DEPRECATED

長期 Corpus

Boolean Search
Exact Search
Phrase Search
Faceted Search
Classification Search
Citation Chasing
Backward Search
Forward Search
Snowballing
Berrypicking
Exploratory Search
Systematic Review Search
Counter-Evidence Search
Prior-Art Search
Identity Search
Entity Search
Temporal Search
Version Search
Graph Search
Cross-Language Search
Federated Search
Adversarial Search

v1.0 不要求全部 executable。

Core Executable Method Set Target

至少：

lexical_search
exact_search / phrase_search
query_divergence
identity_search
entity_search
classification_search
backward_citation
forward_citation
counter_evidence_search
temporal_version_search
relation / family resolve
fetch_document
extract_candidate_evidence

Method Coverage

新增：

MethodCoverage

與 ProviderCoverage 分開。

Exit Gate

• Search Method Corpus schema
• Method lifecycle
• documented/executable distinction
• Core Method Set
• MethodCoverage metric
• provider requirements
• composition metadata
• no provider-branded method IDs

9. v0.4 — Autonomous Search Planner v1

Goal

從 deterministic baseline 進入：

Task / Evidence / Gap / ProviderState-aware autonomous planning

Planner Input

SearchTask
SearchState
EvidenceState
GapState
ProviderState
MethodRegistrySnapshot
ProviderRegistrySnapshot
Budget

Planner Output

SearchPlan

包含：

ActionNode
BranchNode
JoinNode
LoopNode
StopNode

AI Proposes, Runtime Validates

AI / planner proposes plan
↓
PlanValidator
↓
PolicyEvaluator
↓
ExecutionRuntime

AI 不得越過 validator / policy。

Planner 能力

至少：

branching
parallel methods
join
bounded loops
fallback
provider substitution
method substitution
gap-directed continuation
budget-aware planning

Exit Gate

• autonomous plan proposal
• validator remains authoritative
• multi-method plan
• branch / join
• bounded loop
• fallback
• method substitution
• provider substitution
• budget awareness
• gap-directed replanning
• planner metadata enters Receipt
• no persisted planner CoT

10. v0.5 — Stopping / Coverage / Saturation

Goal

回答：

AI 到底什麼時候應該停止搜尋？

Contracts

StopCondition
SearchBudget
CoverageState
SaturationState
UncertaintyState

Canonical Stop Reasons

NO_MATERIAL_GAP_REMAINS
COVERAGE_TARGET_MET
MARGINAL_GAIN_BELOW_THRESHOLD
SATURATION_REACHED
BUDGET_EXHAUSTED
PROVIDER_UNAVAILABLE
POLICY_BLOCKED
HUMAN_REVIEW_REQUIRED
TIME_LIMIT_REACHED

Saturation

只表示：

在目前 method/provider/budget 下，新增資訊的 marginal gain 已下降。

不得解讀為完整 recall。

Coverage Axes

Method Coverage
Provider Coverage
Source Coverage
Evidence Coverage
Jurisdiction Coverage
Language Coverage
Temporal Coverage
Domain-specific Coverage

Exit Gate

• typed StopCondition
• typed SearchBudget
• CoverageState
• SaturationState
• UncertaintyState
• marginal-gain stop logic
• human-review stop
• budget stop
• Receipt stop reason
• Saturation != CompleteRecall tests

11. v0.6 — Evidence / Provenance Closure

Goal

收斂現有 evidence / source-lineage 元件，形成穩定 v1 Evidence Boundary。

Canonical Pipeline

Discovery Candidate
↓
Fetched Source
↓
AcquiredAsset
↓
CandidateEvidence
↓
EvidenceAnchor
↓
Verification
↓
Claim Link
↓
Corroboration / Contradiction
↓
Evidence Ledger

禁止捷徑

下列都不能直接成為 Verified Evidence：

Brave rank
Grok citation
Gemini grounding citation
Crossref rank
EPO search rank

Model-Native Boundary

Gemini grounded answer != Verified Evidence
Grok synthesis != Verified Evidence

只使用可驗證 source references 作 discovery / acquisition frontier。

Provenance Identity

至少追：

source identity
work identity
version identity
content hash
retrieval time
publication time
origin relation
source family
independent root

Exit Gate

• stable Candidate → Evidence pipeline
• provenance identity
• source-family / origin tracing
• independent source roots
• anchor verification
• semantic support separate
• contradiction support
• version identity
• append-only Evidence Ledger
• policy metadata travels downstream

12. v0.7 — Search Receipt & Experience Runtime

Goal

把 Search Receipt 從 audit artifact 提升成：

Search Experience Dataset

Search Receipt 必須記

task
epoch
action
method
provider
surface
binding
provider state
routing decision
policy result
query/input refs
observation
result count
latency
cost
candidate gain
evidence gain
gap reduction
failure
stop reason

禁止保存

hidden chain-of-thought
private reasoning
provider hidden reasoning
raw credentials
secret tokens

Experience Metrics

MethodSuccess(method, task_class)
ProviderSuccess(provider, method)
GapResolution(method, gap_type)
EvidenceYield
CostEfficiency
LatencyEfficiency
ProviderFailureRate
MethodFailureRate

Planner Prior

歷史 experience 可以形成 planner prior，但不能形成 authorization。

Exit Gate

• complete action receipts
• final SearchReceipt
• replay metadata
• failure receipts
• provider substitution history
• method substitution history
• gap reduction metrics
• cost / latency metrics
• planner prior interface
• Learning != Authorization test

13. v0.8 — Evaluation / Benchmark Suite

Goal

Omphalos 必須能回答：

比直接讓模型搜尋有什麼改善？

Provider Substitution Benchmark

比較：

Brave-only
Grok Web-only
Gemini-only
Hybrid routing

Metrics：

Verified Evidence Yield
Independent Source Coverage
Gap Reduction
Cost
Latency
Failure Rate

Method Diversity Benchmark

比較：

lexical only

vs.

lexical
+ exact
+ classification
+ citation
+ counter evidence

Planner Benchmark

fixed plan
vs
gap-directed adaptive plan

Provenance Benchmark

raw source mentions
vs
independent source roots

Replay Benchmark

同一：

Task
Registry Snapshot
Provider Fixture
Policy Snapshot

應能 deterministic replay。

Exit Gate

• provider benchmark
• method benchmark
• planner benchmark
• provenance benchmark
• replay benchmark
• machine-readable outputs
• baseline comparisons
• reproducibility artifacts
• benchmark docs

14. v0.9 — API Freeze / Hardening / RC

Goal

停止新增大功能。

只做：

stability
migration
security
packaging
documentation
release validation

Freeze Candidate Interfaces

至少：

SearchMethodSpec v1
ProviderSpec v1
ProviderState v1
MethodBinding v1
SearchTask v1
SearchState v1
SearchPlan v1
AuthorizedAction v1
ProviderObservation v1
Evidence interfaces v1
Gap interfaces v1
SearchReceipt v1

Error Taxonomy

統一：

policy rejection
credential failure
provider failure
binding failure
validation failure
quota failure
timeout
rate limit
malformed response
evidence verification failure
replay mismatch

Packaging

Python package 暫時保留：

ai_web_research

避免 Big-Bang rename。

可考慮新增：

import omphalos

作 stable facade。

Documentation

至少：

Quickstart
Architecture
Provider authoring
Method authoring
Policy authoring
Domain Pack authoring
Evidence model
Search Receipt
Security / credentials
Benchmarking
Migration

Exit Gate

• public API freeze
• schema/version freeze
• migration rules
• error taxonomy
• clean install
• fresh environment
• CLI/API quickstart
• example workflows
• security scan
• credential leak scan
• package reproducibility
• Release Candidate tag

15. v1.0 — Final Release Gate

v1.0 不以功能很多判定。

必須通過九個 Gate。

G1 — Method Gate

• Search Method 是 first-class；
• 有 versioned Method Registry；
• 有 Core Method Set；
• Method 與 Provider 分離。

G2 — Provider Gate

至少有四種 topology 實例：

MODEL_NATIVE
PROVIDER_NEUTRAL
DOMAIN_SPECIFIC
LOCAL_PRIVATE

G3 — Replaceability Gate

必須證明：

same Method
→ provider A
→ provider A unavailable
→ provider B

Method identity 不變。

G4 — Planner Gate

AI 可以：

select
compose
branch
fallback
switch method
switch provider
replan

Runtime validator 仍具最後權威。

G5 — Policy Gate

必須證明：

UNKNOWN != ALLOW
DENY blocks execution
credentials never persist
rights metadata travels downstream

G6 — Evidence Gate

必須證明：

Retrieved != Verified
Citation != Support
QuoteMatch != SemanticSupport
ProviderGrounding != VerifiedEvidence

G7 — Gap / Stop Gate

Runtime 必須能：

identify gaps
continue search
switch methods
switch providers
stop with explicit reason

G8 — Receipt Gate

每次 research execution 必須：

auditable
replayable
versioned
non-CoT
credential-safe

G9 — Evaluation Gate

至少有：

provider substitution benchmark
method diversity benchmark
planner benchmark
provenance benchmark
replay benchmark

16. v1.0 Reference Workflows

v1.0 不能只靠 unit tests。

至少需要四條完整 E2E。

16.1 General Web Research

Task
↓
Planner
↓
Brave / Grok / Gemini routing
↓
multiple methods
↓
evidence
↓
counter evidence
↓
gap analysis
↓
stop

16.2 X / Current Discourse Research

Grok X
+
Grok Web / Brave
↓
candidate sources
↓
fetch / provenance verification
↓
independent source analysis

16.3 Academic / NPL Research

Crossref
+
general Web
+
identity
+
citation
+
counter-evidence

16.4 Patent Prior-Art

EPO
+
Crossref NPL
+
classification
+
family
+
priority
+
claims
+
coverage gaps

Patent Intelligence 作為 v1.0 第一個正式 Domain Pack reference implementation。

17. 不屬於 v1.0 Blocker

以下項目重要，但不得阻塞 Omphalos v1.0。

Economic Research Platform

應作為：

Omphalos-compatible Economic Domain Pack

獨立演進。

Meteorological Research Platform

應作為：

Omphalos-compatible Meteorological Domain Pack

獨立演進。

所有人類已知搜尋方法全部 executable

不屬於 v1.0 blocker。

v1.0 只要求：

Search Method Corpus exists
+
Core Executable Method Set exists

18. Domain Pack Boundary

Omphalos Runtime 負責：

search methods
planning
providers
policy
execution
evidence
gaps
receipts
experience

Domain Pack 負責：

domain concepts
domain methods
domain identifiers
domain validation
domain forecasts
domain benchmarks

19. Backlog Policy

任何新功能先分類：

A. Runtime Blocker

直接進目前 milestone。

B. Runtime Enhancement

排到 v1.x。

C. Search Method Corpus

先 documented，不一定 executable。

D. Provider Candidate

先記 Provider Profile。

E. Domain Pack

獨立版本管理。

F. Research Experiment

不得直接污染 stable runtime contract。

20. Version Discipline

每一 milestone 必須：

Design
↓
TDD
↓
Fresh verification
↓
Source diff
↓
PR
↓
Review
↓
Merge
↓
Version checkpoint

禁止：

implementation half done
↓
version number changed
↓
old plan forgotten

21. Branch / PR Discipline

建議：

integration/omphalos-v0.2-provider-state
integration/omphalos-v0.3-method-corpus
integration/omphalos-v0.4-planner
...

每個 milestone 原則上一個 review boundary。

避免無限疊一個巨大 PR。

22. Verification Discipline

任何「完成」宣稱前至少執行：

pytest
compileall
import smoke
source diff
branch compare
secret scan
artifact verification

v0.9 / v1.0 再增加：

clean install
reproducibility
benchmark replay

23. Credential Discipline

所有 Provider credentials：

xAI API key
Gemini API key
Brave API key
EPO OAuth
future economic API keys
future meteorological tokens

只能存在：

runtime secret store
environment
execution context
credential manager

不能出現在：

GitHub source
Search Receipt
ProviderObservation
Evidence object
logs
fixtures
release bundle

24. Provider / API Change Discipline

若 API 改版：

Provider Adapter Version
↓
Binding Version
↓
ProviderState

不應直接修改 Search Method identity，除非 Method semantics 本身真的改變。

25. Search Method Governance

新 Search Method 進 Executable Registry 前，至少需要：

canonical definition
input contract
output contract
failure modes
provider requirements
test fixture
at least one implementation
method-level validation

26. Domain Governance

Domain-specific Method 不應硬塞進 core。

例如：

patent.family_resolve
meteorology.station_bias_correction
economics.vintage_revision_analysis

應保持 Domain Pack identity。

27. v1.0 Public Identity

建議正式對外描述：

Omphalos / AUSI Runtime is a provider-replaceable, policy-aware, evidence-grounded, auditable AI-native Search Method Runtime.

中文：

Omphalos / AUSI Runtime 是一個 Provider 可替換、具 Policy 邊界、以 Evidence 為基礎、可審計的 AI 原生搜尋方法 Runtime。

28. v1.0 不應宣稱

solves all search
complete recall
all human search methods implemented
all evidence verified
AGI research system
fully autonomous science
perfect forecasting
complete patent recall

29. Release Artifact

v1.0 Final Release 至少產生：

source tree
canonical docs
API reference
examples
benchmark artifacts
release manifest
dependency lock / reproducibility metadata
FINAL ZIP
SHA-256
Git tag
GitHub Release

30. Final v1.0 Verification Procedure

1. source-diff verification
2. final HEAD verification
3. clean-environment installation
4. full Runtime test suite
5. Reference Workflow verification
6. Benchmark replay
7. credential / secret scan
8. policy fixture verification
9. reproducible package build
10. FINAL ZIP
11. SHA-256
12. release tag

31. Canonical Roadmap Summary

CURRENT
│
├─ Grok Web / X
├─ Gemini Google Search
├─ Brave
├─ Crossref
├─ EPO
├─ Policy
├─ Evidence
└─ Search Receipt foundations
│
▼
v0.2
ProviderState + Dynamic Routing
│
▼
v0.3
Search Method Corpus + Core Method Set
│
▼
v0.4
Autonomous Search Planner
│
▼
v0.5
Stopping + Coverage + Saturation
│
▼
v0.6
Evidence + Provenance Closure
│
▼
v0.7
Receipt + Experience Learning
│
▼
v0.8
Benchmarks + Evaluation
│
▼
v0.9
API Freeze + Hardening + RC
│
▼
v1.0
Provider-replaceable
Policy-aware
Evidence-grounded
Auditable
Replayable
Gap-directed
AI-native Search Method Runtime

32. Local AI Canonical Instruction

給任何後續接手 Omphalos 的 AI：

不要為了新 API 偏離 v1.0 Roadmap。

不要把 Provider 新增誤認成 Search Intelligence 增長。

不要為了完成 Omphalos v1.0，先去完成整個經濟研究院、氣象研究院或所有 Domain Pack。

不要因為某個新搜尋法很有趣，就跳過目前 milestone。

新方法先進 Search Method Corpus。

新 Provider 先進 Provider Candidate / Binding。

新 Domain 能力進 Domain Pack。

主 Runtime 只按照本 Roadmap 往 v1.0 推進。

33. 最終原則

Omphalos 的 v1.0 不是 API 數量競賽，也不是功能數量競賽。

它要證明的是：

可以形成一個穩定、可擴張的 Runtime。

最終仍回到：

而 v1.0 的真正完成標準是：

Omphalos 已經能把「搜尋方法」本身當成 AI 可以自主組合與執行的第一級計算／認知物件。

End of Canonical Roadmap