Skip to content

The Framework of "Authoritative Reference Sources"

Providing authoritative reference sources for AI decisions is the foundation of reliability in AI-driven development.

Positioning of This Document

This document provides the sole physical foundation that makes the worldview presented in the Vision (01-vision) viable. Where the Vision explains "why authoritative reference sources are necessary for AI-driven development," this document answers that question with a logical and technical framework. This architecture is not invented from scratch — it restructures the standards humanity has accumulated (RFCs, W3C, legislation) as design infrastructure for the AI era.

Meta Information
What this chapter establishesFour characteristics of reference sources (authority, structure, versioning, machine-readability) and MCP design principles
What this chapter does NOT coverMCP server implementation details (→implementation guides), team domain knowledge (→Skills)
Dependencies01-vision (definition of AI constraints)
Common misuseApplying MCP where RAG would suffice. Identifying when structured access is truly needed is critical

About This Document

In AI-driven development, the quality of AI output is determined by "what it references." This document systematically organizes the "authoritative reference sources" that AI should use as the basis for its decisions, clarifying why they are necessary, what characteristics they should have, and how they should be connected.

By converting authoritative information sources such as RFCs, W3C specifications, and legislation into MCP format, AI output can have verifiable evidence. This enables a development environment where the question "Is what the AI says really true?" can always be answered by pointing to the original source.

Note: This document primarily covers "external authoritative information sources." For team domain knowledge and best practices, see Skills. For the distinction between MCP and Skills, see vs-mcp.md.

Why Does AI Output Vary?

1.1 The Nature of Probabilistic Generation

Large Language Models (LLMs) are fundamentally probabilistic text generation systems. The diagram below illustrates how this process works from input to output:

CharacteristicDescriptionImpact on AI
Statistical Pattern LearningPredicts "likely next tokens" from co-occurrence patterns in training dataOutputs "plausible" rather than "correct"
Non-deterministic SamplingSame input can produce different outputsConsistency guarantees are difficult
Context DependencySubtle differences in prompts change outputReproducibility issues

1.2 The Four Fundamental Limitations of AI

Understanding these four core categories of limitations is essential for working with AI effectively:

1.2.1 Accuracy Limitations (Hallucination Problem)

AI doesn't "know" — it "generates."

User: What is the meaning of status code 1006 for Close frames
      as defined in Section 5.5.1 of RFC 6455?

AI Possibility A: "1006 indicates unexpected disconnection" (correct)
AI Possibility B: "1006 indicates a protocol error" (wrong - that's 1002)
AI Possibility C: "Section 5.5.1 contains the definition of 1006" (wrong - 7.4.1 is correct)

Hallucination Generation Mechanisms

The following table explains the primary causes that lead to hallucinated outputs:

CauseDescriptionExample
Sparse Training DataRare information is undertrainedDetails of minor RFCs
Confusion with Similar PatternsConfuses similar conceptsClose code 1002 vs 1006
Overconfident CompletionFills in unknown parts "plausibly"Generating non-existent section numbers
Context ContaminationTreats misinformation in conversation as truthAmplifies user misunderstandings

1.2.2 Recency Limitations

AI knowledge is fixed at a point in time, making it unable to reflect real-world changes that occur after training:

Specific Impacts

The following examples demonstrate how recency limitations manifest in practical scenarios:

CategoryProblemExample
New RFCsUnaware of existenceDoes not know RFC 9562 (UUIDv7)
Legal AmendmentsAnswers based on old lawAnswers based on pre-amendment privacy law
Deprecation/UpdatesTreats old specifications as currentReferences RFC 2616 as the HTTP/1.1 standard
SecurityUnaware of known vulnerabilitiesDoes not know CVEs discovered after cutoff

1.2.3 Authority Limitations

AI output is "one interpretation" and not an official opinion. When multiple interpretations of a specification are possible, AI may select one without guarantee of correctness:

Problem Structure:

RFC 6455 Original Text

    Multiple Interpretation Possibilities
    ├── Interpretation A (Strict)
    ├── Interpretation B (Lenient)
    └── Interpretation C (Context-dependent)

AI Output

    Outputs one interpretation "plausibly"

    No guarantee that it is the correct interpretation

Situations Where Lack of Authority Is Problematic

These scenarios illustrate where the absence of official authority creates real risks:

SituationRiskRequired Response
Specification ImplementationNon-compliant implementationVerify RFC original text
Legal DecisionsCompliance violationsVerify legal text
SecurityOverlooking vulnerabilitiesCheck official advisories
Contracts/SLAsDisputes from misinterpretationVerify contract text

1.2.4 Accountability Limitations

AI output has no subject of accountability. This differs fundamentally from traditional information sources where responsibility can be traced:

ProblemDescriptionResult
Source OpacityUnknown basis for generationCannot verify
Revision UntraceabilityUnknown when information is fromCannot audit
Error AttributionUnclear who bears responsibilityDifficult risk management

These accountability issues create significant challenges for risk management and compliance.

What Are "Authoritative Reference Sources"?

2.1 Definition

Authoritative Reference Sources are information sources that satisfy the following characteristics, each contributing to overall reliability:

2.2 The Five Characteristics

2.2.1 Authoritativeness

The information source has official decision-making authority or expertise in its domain. Different types of authority carry different weights:

Type of AuthorityDescriptionExamples
Institutional AuthorityOfficial bodies established by law or treatyIETF, W3C, ISO, National Governments
De facto AuthorityEntities recognized as de facto standards in the industryOWASP, Ecma International
Academic AuthorityAcademic communities with peer review processesIEEE, ACM
Technical AuthorityDevelopers/maintainers of technologyOSS Projects, Vendors

The hierarchy of authority is shown in the following diagram:

2.2.2 Immutability & Versioning

Once published, content either does not change or when it does, clear version management is applied. These patterns ensure you know the precise source of information:

PatternDescriptionExample
Complete ImmutabilityNever changed once publishedRFC (except Errata)
Versioned ChangesNew version replaces old versionISO Standards, W3C Recommendations
Explicit DeprecationOld versions explicitly deprecatedRFC obsoletes/updates

Here is how the RFC versioning model exemplifies this principle:

RFC Immutability Model:

RFC 2616 (HTTP/1.1, 1999)
    ↓ obsoleted by
RFC 7230-7235 (2014)
    ↓ obsoleted by
RFC 9110-9114 (2022)

→ Each RFC is unchanged after publication
→ New RFCs "replace" old RFCs
→ Clear which point in time the specification is from

2.2.3 Structuredness

Information is systematically organized so specific information can be precisely referenced. Structured sources allow for unambiguous citations:

Structuring ElementDescriptionBenefit for AI
Hierarchical StructureClear hierarchy of chapters/sections/paragraphsPrecise reference to specific locations
IdentifiersUnique section/article numbersUnambiguous citations
Cross-referencesExplicit links to other documents/sectionsTracking related information
IndexTerm index, requirement listsEfficient searching

2.2.4 Verifiability

AI output can be confirmed against the original source.

Verifiability directly connects to the "Third Axis of Knowledge Transformation: Verification (Spec-to-Test)" defined in the Vision. Beyond merely referencing specifications, converting them into verifiable tests forms a quality closed loop for AI outputs.

The following sequence diagram shows how verification works in practice:

Elements Ensuring Verifiability

ElementDescriptionImplementation
Persistent URIReference won't disappearDOI, RFC number, Legal article number
Version SpecificationClarify which version was referencedRFC 9110, ISO 27001:2022
Section SpecificationClarify which part was referencedSection 7.4.1
Original Text CitationShow the referenced wordingMUST/SHOULD/MAY original text

2.2.5 Accessibility

Provided in a format that AI can access programmatically. Different levels of accessibility present different opportunities and challenges:

LevelDescriptionExample
Structured APIAccessible in machine-readable formatRFC XML, e-Gov API
HTML/PDFPublished on web but requires parsingW3C specs, most ISO
Paid/RestrictedAccess has constraintsSome ISO standards

The diagram below illustrates the accessibility spectrum:

2.3 Criteria for Evaluating "Authoritative Reference Sources"

Use this decision flowchart to evaluate whether an information source qualifies as authoritative:

Hierarchical Structure of Reference Sources

3.1 The Four-Layer Model

Reference sources are organized into a four-level hierarchy, where each level has different compliance requirements. The compliance obligations follow RFC 2119 terminology.

LevelCompliance ObligationRisk of Deviation
Level 1 International Standards & RegulationsMUST (mandatory)Interoperability issues, legal problems
Level 2 Industry Standards & De factoSHOULD (recommended)Compatibility issues within industry
Level 3 Organization/Project RulesLOCAL (org-mandatory)Team-level inconsistency
Level 4 Best PracticesMAY (optional)Missed quality improvement opportunities

3.2 Level Details

Level 1: International Standards & Regulations (MUST Comply)

Highest authority reference sources. Violations cause interoperability issues or legal problems. The following table shows examples from different domains:

CategoryReference5 Characteristics RatingMCP Status
Communication ProtocolsIETF RFC◎◎◎◎◎✅ rfcxml-mcp
Web StandardsW3C / WHATWG◎◎◎◎○✅ w3c-mcp
International StandardsISO◎◎◎○△⚡ Partial (pdf-spec-mcp)
Japanese Lawse-Gov◎◎◎◎◎✅ hourei-mcp
EU RegulationsEUR-Lex◎◎◎◎○📋 Planned

IETF RFCs exemplify the highest standard for all five characteristics:

Authoritativeness: ◎ Official publication by IETF, WG consensus
Immutability:      ◎ No changes after publication, managed via obsoletes/updates
Structuredness:    ◎ Section numbers, clear MUST/SHOULD/MAY definitions
Verifiability:     ◎ Uniquely identified by RFC number and section number
Accessibility:     ◎ Published in RFC XML format, free access

Level 2: Industry Standards & De facto (SHOULD Comply)

Widely adopted standards. Non-compliance causes compatibility issues within the industry. Key examples include:

CategoryReferenceCharacteristicsDeliveryMCP Value
API DesignOpenAPI SpecDe facto standard for REST APIsMCPHigh
SecurityOWASPWeb security best practicesMCPHigh
AuthenticationOAuth 2.0 / OIDCDe facto standard for authorizationMCPHigh
MessagingAsyncAPIAsync API specificationMCPMedium

Level 3: Organization/Project Rules (Local Compliance)

Rules that should be unified within teams/projects. Since external API access is not required, Skills-based systematization is the primary delivery method.

TypeCharacteristicsManagement MethodDeliveryMCP Value
Coding StandardsProject-specific stylesMarkdown / Linter configsSkillsLow
ADRArchitecture decision recordsGit-managed MarkdownSkillsLow
CLAUDE.mdClaude-specific instructionsProject root placementDirect

Level 3 knowledge is self-contained within the organization, making Skills (Markdown-based domain knowledge systematization) more appropriate than MCP (dynamic access to external APIs). Tool enforcement such as linter configs functions as guardrails (inviolable constraints).

Recommendations based on experience. Apply as appropriate to the situation. By formalizing team-agreed practices as Skills, AI can reference them consistently.

TypeSourceApplication JudgmentDeliveryMCP Value
Design PrinciplesSOLID, DRY, KISSSituationalSkillsLow
Design PatternsGoF, POSAWhen matching the problemSkillsLow
Clean CodeRobert C. MartinWithin team agreementSkillsLow

Level 4 has the nature of "no single correct answer," making it unsuitable for MCP. Formalizing adopted principles as Skills and providing them as AI judgment criteria is the effective approach.

AI Decision Flow

4.1 Decision Algorithm Based on Reference Sources

When implementing a feature or making a decision, AI should follow this hierarchical decision flow:

4.2 Output Templates

When providing answers, structure your response according to whether authoritative sources are available. Here is the template for cases where a reference source is found:

markdown
## Answer

Status code 1006 for WebSocket Close frames indicates "Abnormal Closure."

### Evidence

- **Source**: RFC 6455, Section 7.4.1
- **Original Text**: "1006 is a reserved value and MUST NOT be set as a status code
  in a Close control frame by an endpoint. It is designated for use in
  applications expecting a status code to indicate that the connection
  was closed abnormally"
- **Requirement Level**: MUST NOT (must not be set in implementation)

### Notes

This code is for applications to detect abnormal termination and
cannot be sent in actual Close frames.

When no authoritative reference source is available, use this template to transparently acknowledge the limitation:

markdown
## Answer

I could not identify an authoritative reference source for this matter.

### Information Sources Checked

- RFC 6455: No relevant description found
- W3C WebSocket API: No relevant description found

### Speculation

As a general implementation practice, there is a tendency to~, but
this is not defined by any standard.

### Recommendation

If precise specifications are needed, I recommend checking~.

Design Requirements for Reference Source MCPs

5.0 Separation of MCP and Skills

MCP and Skills are both means to achieve "authoritative reference sources", but they serve different purposes:

AspectMCPSkills
TargetExternal authoritative information sourcesDomain knowledge & best practices
ExamplesRFC, Laws, W3C standardsDesign principles, Coding standards
CharacteristicsDynamic access, via APIStatic reference, Markdown format
UpdatesDependent on external systemsTeam-driven updates

For details, see skills/vs-mcp.md.

5.1 Required Functions

MCPs should provide these core functions to enable comprehensive access to reference sources:

FunctionDescriptionExample
SearchKeyword search within specifications"WebSocket close frame"
Structure RetrievalChapter/section hierarchyTable of contents for RFC 6455
Requirements ExtractionExtract MUST/SHOULD/MAYList of normative requirements
Term DefinitionsGet definitions of technical termsDefinition of "Origin"
Reference RelationshipsDependencies on other specificationsRFC 6455 → RFC 2616
Checklist GenerationGenerate implementation verification itemsClient implementation checklist
ValidationCheck if implementation complies with specStatement validation

5.2 RFC MCP Tool Design

The following TypeScript interface illustrates the structure of an RFC MCP implementation:

typescript
interface RfcMcpTools {
	// Search & Retrieval
	searchRfc(keyword: string): RfcSummary[];
	getRfcStructure(rfcNumber: number): Section[];
	getRfcSection(rfcNumber: number, section: string): Content;

	// Requirements Extraction
	getRequirements(rfcNumber: number, level?: RequirementLevel): Requirement[];
	getDefinitions(rfcNumber: number, term?: string): Definition[];

	// Relationships
	getDependencies(rfcNumber: number): Dependency[];
	getRelatedSections(rfcNumber: number, section: string): RelatedSection[];

	// Implementation Support
	generateChecklist(rfcNumber: number, role: 'client' | 'server'): Checklist;
	validateStatement(rfcNumber: number, statement: string): ValidationResult;
}

Concrete Example — Electronic Signature Act × RFC 3161

6.2 Verification Workflow via MCP Integration

Resolving Reference Source Conflicts

7.1 Conflict Resolution Rules

  1. Higher levels take priority - Laws > Industry standards > Organization rules
  2. Newer versions take priority - RFC 9110 > RFC 7230 (obsolete)
  3. More specific specifications take priority - WebSocket RFC > General TCP specifications
  4. When contradictions exist, defer to human judgment

7.2 Example: HTTP Specification References

❌ RFC 2616 (HTTP/1.1 - obsolete)
✅ RFC 9110 (HTTP Semantics - current)
✅ RFC 9111 (HTTP Caching - current)

List of Built Reference Source MCPs

MCPTargetMain FunctionsStatusRepository
rfcxml-mcpIETF RFCStructure retrieval, requirements extraction, checklist generation✅ ProductionGitHub
w3c-mcpW3C/WHATWG/IETF Web StandardsWebIDL, CSS, HTML elements✅ ProductionGitHub
hourei-mcpJapanese Laws (e-Gov)Law search, article retrieval✅ ProductionGitHub
pdf-spec-mcpPDF Spec (ISO 32000)Structure retrieval, requirements extraction, version comparison✅ ProductionGitHub

Future Expansion Candidates

High Priority

CandidateTargetValueStatus
OpenAPI MCPOpenAPI SpecAPI design standards compliance📋 Planned
OWASP MCPOWASP Top 10 etc.Security requirements checking📋 Planned
OAuth MCPOAuth 2.0 / OIDCAuthentication flow implementation support📋 Planned

Medium Priority

CandidateTargetValueStatus
ISO MCPISO StandardsInternational standards reference📋 Planned
BIM/IFC MCPbuildingSMART IFCBuilding information model📋 Planned
HL7 FHIR MCPHL7 FHIRHealthcare information exchange📋 Planned

Summary

Core Messages

  1. AI output is inherently variable - Probabilistic generation, training data constraints, lack of authority
  2. "Authoritative reference sources" are needed - Authoritativeness, immutability, structuredness, verifiability, accessibility
  3. Organize hierarchically - International standards > Industry standards > Organization rules > Best practices
  4. Connect to external information sources via MCP - Provide RFC, laws, W3C standards in a format AI can reference
  5. Systematize domain knowledge with Skills - Share team know-how in reusable formats
  6. Always cite evidence - Show source, section, and original text

The Value of "Authoritative Reference Sources"

By providing authoritative reference sources for AI decisions, output reliability and verifiability are supported.

The "authoritative reference sources" presented in this document form the physical foundation of the "verifiable design philosophy" articulated in the Vision. The responsibility shift model (design-time, execution-time, structural constraints) and the two-layer verification structure (guardrails and evaluation pipelines) defined in the Vision are built upon the characteristics and hierarchical structure systematized here.

Released under the MIT License.