Longterm Wiki

AI Model Specifications

model-spec (E594)
← Back to pagePath: /knowledge-base/responses/model-spec/
Page Metadata
{
  "id": "model-spec",
  "numericId": null,
  "path": "/knowledge-base/responses/model-spec/",
  "filePath": "knowledge-base/responses/model-spec.mdx",
  "title": "AI Model Specifications",
  "quality": 50,
  "importance": 61,
  "contentFormat": "article",
  "tractability": null,
  "neglectedness": null,
  "uncertainty": null,
  "causalLevel": null,
  "lastUpdated": "2026-01-29",
  "llmSummary": "Model specifications are explicit documents defining AI behavior, now published by all major frontier labs (Anthropic, OpenAI, Google, Meta) as of 2025. While they improve transparency and enable external scrutiny, they face a fundamental spec-reality gap—specifications don't guarantee implementation, with no robust verification mechanisms existing.",
  "structuredSummary": null,
  "description": "Model specifications are explicit written documents defining desired AI behavior, values, and boundaries. Pioneered by Anthropic's Claude Soul Document and OpenAI's Model Spec (updated 6+ times in 2025), they improve transparency and enable external scrutiny. As of 2025, all major frontier labs publish specs, with 78% of enterprises now using AI in at least one function—making behavioral documentation increasingly critical for accountability.",
  "ratings": {
    "novelty": 3.5,
    "rigor": 5,
    "actionability": 4.5,
    "completeness": 6
  },
  "category": "responses",
  "subcategory": "alignment-policy",
  "clusters": [
    "ai-safety",
    "governance"
  ],
  "metrics": {
    "wordCount": 2736,
    "tableCount": 25,
    "diagramCount": 1,
    "internalLinks": 14,
    "externalLinks": 28,
    "footnoteCount": 0,
    "bulletRatio": 0.03,
    "sectionCount": 38,
    "hasOverview": true,
    "structuralScore": 14
  },
  "suggestedQuality": 93,
  "updateFrequency": 21,
  "evergreen": true,
  "wordCount": 2736,
  "unconvertedLinks": [
    {
      "text": "78% of organizations using AI",
      "url": "https://mckinsey.com",
      "resourceId": "14a922610f3ad110",
      "resourceTitle": "McKinsey Global Institute"
    },
    {
      "text": "Constitutional AI",
      "url": "https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback",
      "resourceId": "e99a5c1697baa07d",
      "resourceTitle": "Constitutional AI: Harmlessness from AI Feedback"
    },
    {
      "text": "McKinsey survey 2024",
      "url": "https://mckinsey.com",
      "resourceId": "14a922610f3ad110",
      "resourceTitle": "McKinsey Global Institute"
    },
    {
      "text": "Constitutional AI: Harmlessness from AI Feedback",
      "url": "https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback",
      "resourceId": "e99a5c1697baa07d",
      "resourceTitle": "Constitutional AI: Harmlessness from AI Feedback"
    },
    {
      "text": "Collective Constitutional AI",
      "url": "https://www.anthropic.com/research/collective-constitutional-ai-aligning-a-language-model-with-public-input",
      "resourceId": "3c862a18b467640b",
      "resourceTitle": "Collective Constitutional AI"
    },
    {
      "text": "McKinsey AI Survey 2024",
      "url": "https://mckinsey.com",
      "resourceId": "14a922610f3ad110",
      "resourceTitle": "McKinsey Global Institute"
    }
  ],
  "unconvertedLinkCount": 6,
  "convertedLinkCount": 0,
  "backlinkCount": 0,
  "redundancy": {
    "maxSimilarity": 14,
    "similarPages": [
      {
        "id": "constitutional-ai",
        "title": "Constitutional AI",
        "path": "/knowledge-base/responses/constitutional-ai/",
        "similarity": 14
      },
      {
        "id": "evals-governance",
        "title": "Evals-Based Deployment Gates",
        "path": "/knowledge-base/responses/evals-governance/",
        "similarity": 14
      },
      {
        "id": "ai-control",
        "title": "AI Control",
        "path": "/knowledge-base/responses/ai-control/",
        "similarity": 13
      },
      {
        "id": "alignment",
        "title": "AI Alignment",
        "path": "/knowledge-base/responses/alignment/",
        "similarity": 13
      },
      {
        "id": "evals",
        "title": "Evals & Red-teaming",
        "path": "/knowledge-base/responses/evals/",
        "similarity": 13
      }
    ]
  }
}
Entity Data
{
  "id": "model-spec",
  "type": "policy",
  "title": "AI Model Specifications",
  "description": "Model specifications are explicit documents defining AI behavior, now published by all major frontier labs (Anthropic, OpenAI, Google, Meta) as of 2025. While they improve transparency and enable external scrutiny, they face a fundamental spec-reality gap—specifications don't guarantee implementatio",
  "tags": [],
  "relatedEntries": [],
  "sources": [],
  "lastUpdated": "2026-02",
  "customFields": []
}
Canonical Facts (0)

No facts for this entity

External Links

No external links

Backlinks (0)

No backlinks

Frontmatter
{
  "title": "AI Model Specifications",
  "description": "Model specifications are explicit written documents defining desired AI behavior, values, and boundaries. Pioneered by Anthropic's Claude Soul Document and OpenAI's Model Spec (updated 6+ times in 2025), they improve transparency and enable external scrutiny. As of 2025, all major frontier labs publish specs, with 78% of enterprises now using AI in at least one function—making behavioral documentation increasingly critical for accountability.",
  "sidebar": {
    "order": 7
  },
  "quality": 50,
  "importance": 61.5,
  "lastEdited": "2026-01-29",
  "update_frequency": 21,
  "llmSummary": "Model specifications are explicit documents defining AI behavior, now published by all major frontier labs (Anthropic, OpenAI, Google, Meta) as of 2025. While they improve transparency and enable external scrutiny, they face a fundamental spec-reality gap—specifications don't guarantee implementation, with no robust verification mechanisms existing.",
  "ratings": {
    "novelty": 3.5,
    "rigor": 5,
    "actionability": 4.5,
    "completeness": 6
  },
  "clusters": [
    "ai-safety",
    "governance"
  ],
  "subcategory": "alignment-policy",
  "entityType": "approach"
}
Raw MDX Source
---
title: AI Model Specifications
description: Model specifications are explicit written documents defining desired AI behavior, values, and boundaries. Pioneered by Anthropic's Claude Soul Document and OpenAI's Model Spec (updated 6+ times in 2025), they improve transparency and enable external scrutiny. As of 2025, all major frontier labs publish specs, with 78% of enterprises now using AI in at least one function—making behavioral documentation increasingly critical for accountability.
sidebar:
  order: 7
quality: 50
importance: 61.5
lastEdited: "2026-01-29"
update_frequency: 21
llmSummary: Model specifications are explicit documents defining AI behavior, now published by all major frontier labs (Anthropic, OpenAI, Google, Meta) as of 2025. While they improve transparency and enable external scrutiny, they face a fundamental spec-reality gap—specifications don't guarantee implementation, with no robust verification mechanisms existing.
ratings:
  novelty: 3.5
  rigor: 5
  actionability: 4.5
  completeness: 6
clusters:
  - ai-safety
  - governance
subcategory: alignment-policy
entityType: approach
---
import {R, EntityLink, DataExternalLinks, Mermaid} from '@components/wiki';

<DataExternalLinks pageId="model-spec" />

## Quick Assessment

| Dimension | Assessment | Evidence |
|-----------|------------|----------|
| **Tractability** | High | All major frontier labs now publish specs; relatively low technical barriers to creation |
| **Effectiveness** | Medium | Improves transparency and accountability; limited enforcement mechanisms |
| **Adoption** | Widespread (2025) | Anthropic, OpenAI, <EntityLink id="E98">Google DeepMind</EntityLink>, Meta all publish model documentation |
| **Investment** | \$10-30M/year industry-wide | Internal lab work on spec development and training integration |
| **Timeline** | Immediate | Mature practice since 2019 ([Model Cards](https://arxiv.org/abs/1810.03993)); accelerating since 2024 |
| **Key Limitation** | Spec-reality gap | Specifications don't guarantee implementation; gaming potential high |
| **Grade: Transparency** | A- | Public specs enable external scrutiny and accountability |
| **Grade: Enforcement** | C+ | Verification methods underdeveloped; compliance testing limited |

## Overview

Model specifications are explicit, written documents that define the intended behavior, values, and boundaries of AI systems. Rather than relying solely on implicit learning from training data, model specs provide clear articulation of what an AI system should and should not do, how it should handle edge cases, and what values should guide its behavior when tradeoffs arise. As of 2025, all major frontier AI labs—including <EntityLink id="E22">Anthropic</EntityLink>, <EntityLink id="E218">OpenAI</EntityLink>, Google DeepMind, and Meta—publish model specifications or detailed model cards for their systems.

The practice emerged from recognizing that implicit behavioral training through <EntityLink id="E259">RLHF</EntityLink> alone leaves important questions unanswered: What should the model do when helpfulness conflicts with honesty? How should it handle requests that might be harmful in some contexts but legitimate in others? Model specs provide explicit answers to these questions, creating a documented target for training and a reference for evaluation. The foundational work on [Model Cards for Model Reporting](https://arxiv.org/abs/1810.03993) by Mitchell et al. (2019), which introduced standardized documentation for ML models, has been cited over 2,273 times and established the framework for AI behavior documentation.

Anthropic's [Claude Soul Document](https://simonwillison.net/2025/Dec/2/claude-soul-document/)—a 14,000-token document embedded into model weights during supervised learning—represents one approach, defining Claude's identity, ethical framework, and hierarchy of principals (Anthropic → Operators → Users). OpenAI's [Model Spec](https://model-spec.openai.com/2025-12-18.html) has been updated 6+ times in 2025, with versions addressing agent principles, teen safety, and collective alignment input from over 1,000 people worldwide. Meta publishes comprehensive [Llama Model Cards](https://www.llama.com/docs/model-cards-and-prompt-formats/) alongside safety guardrails like Llama Guard.

However, a fundamental limitation remains: specifications define what behavior is desired, but don't guarantee that behavior is achieved. A gap can exist between spec and implementation, and sophisticated systems might comply with the letter while violating the spirit of specifications. With [78% of organizations using AI](https://mckinsey.com) in at least one business function (up from 55% in 2023 per McKinsey), and enterprise AI spending reaching \$17 billion in 2025, the stakes for reliable model specifications continue to rise.

### How Model Specs Integrate with Training

<Mermaid chart={`
flowchart TD
    SPEC[Model Specification<br/>Written Document] --> CAI[Constitutional AI<br/>Training]
    SPEC --> RLHF[RLHF Guidelines<br/>Rater Instructions]
    SPEC --> EVAL[Evaluation<br/>Test Criteria]

    CAI --> MODEL[Trained Model]
    RLHF --> MODEL

    MODEL --> DEPLOY[Deployed System]
    EVAL --> VERIFY{Verification<br/>Testing}
    DEPLOY --> VERIFY

    VERIFY -->|Pass| RELEASE[Public Release]
    VERIFY -->|Fail| ITERATE[Iterate on<br/>Spec or Training]
    ITERATE --> SPEC

    subgraph External["External Accountability"]
        RELEASE --> SCRUTINY[Researcher Scrutiny]
        RELEASE --> REG[Regulator Review]
        RELEASE --> PUBLIC[Public Evaluation]
    end

    style SPEC fill:#e6f3ff
    style MODEL fill:#fff3cd
    style RELEASE fill:#d4edda
    style VERIFY fill:#f8d7da
`} />

## Risk Assessment & Impact

| Risk Category | Assessment | Key Metrics | Evidence Source |
|---------------|------------|-------------|-----------------|
| **Safety Uplift** | Medium | Provides clear behavioral guidelines | Structural benefit |
| **Capability Uplift** | Some | Clearer specs improve usefulness within bounds | Secondary effect |
| **Net World Safety** | Helpful | Improves transparency; enables scrutiny | Governance value |
| **Lab Incentive** | Moderate | Helps deployment; some PR value | Mixed motivations |

## How Model Specs Work

### Components of a Model Specification

| Component | Description | Example |
|-----------|-------------|---------|
| **Identity & Character** | Who the AI is, its personality | "Claude is helpful, harmless, and honest" |
| **Behavioral Guidelines** | What the AI should/shouldn't do | "Refuse to help with illegal activities" |
| **Value Hierarchy** | How to handle tradeoffs | "Safety > Honesty > Helpfulness when they conflict" |
| **Edge Case Guidance** | Specific scenario handling | "For medical questions, recommend seeing a doctor" |
| **Harm Categories** | What counts as harmful | Detailed harm taxonomy |
| **Context Sensitivity** | How context changes behavior | "Professional coding vs general chat" |

### The Spec-Training-Evaluation Loop

| Stage | Process | Purpose |
|-------|---------|---------|
| **1. Spec Creation** | Document intended behavior | Define target |
| **2. Training Alignment** | Train model toward spec | Achieve behavior |
| **3. Evaluation** | Test against spec | Verify compliance |
| **4. Iteration** | Update spec based on findings | Refine understanding |

### Integration with Training

Model specs integrate with training in several ways:

| Integration Point | Method | Effectiveness |
|------------------|--------|---------------|
| **Constitutional AI** | Principles drawn from spec | Direct incorporation |
| **RLHF Guidelines** | Rater instructions from spec | Indirect alignment |
| **Fine-tuning** | Spec-derived examples | Targeted training |
| **Evaluation** | Test cases from spec | Verify compliance |

## Published Model Specifications

### Comparison of Major Model Specifications (2025)

| Organization | Document | Length | Key Features | Public Since | Updates |
|--------------|----------|--------|--------------|--------------|---------|
| **Anthropic** | [Claude Soul Document](https://simonwillison.net/2025/Dec/2/claude-soul-document/) | ≈14,000 tokens | Identity, ethics, principal hierarchy | Dec 2025 (leaked, confirmed) | Embedded in weights |
| **OpenAI** | [Model Spec](https://model-spec.openai.com/2025-12-18.html) | ≈8,000 words | Authority hierarchy, agent principles, teen safety | May 2024 | 6+ versions in 2025 |
| **Meta** | [Llama Model Cards](https://www.llama.com/docs/model-cards-and-prompt-formats/) | ≈3,000 words/model | Performance benchmarks, safety guardrails | 2023 | Per-release updates |
| **Google** | [Gemini Model Cards](https://modelcards.withgoogle.com/assets/documents/gemini-2.5-pro.pdf) | ≈5,000 words | Training data, capabilities, limitations | 2024 | Per-release updates |

### Anthropic's Claude Soul Document (2024-2025)

Anthropic's approach embeds the specification directly into model weights during supervised learning, making it more fundamental than a system prompt. Technical staff member Amanda Askell [confirmed](https://futurism.com/artificial-intelligence/anthropic-claude-soul) the document "is based on a real document and we did train Claude on it."

| Section | Content | Key Provisions |
|---------|---------|----------------|
| **Soul Overview** | Claude's identity and purpose | "Genuinely novel kind of entity"; distinct from sci-fi robots or simple chatbots |
| **Ethical Framework** | Empirical approach to ethics | Treats moral questions with "same rigor as empirical claims about the world" |
| **Principal Hierarchy** | Authority chain | Anthropic → Operators → Users, with defined override conditions |
| **Wellbeing** | Functional emotions | Acknowledges Claude "may have functional emotions" that matter |
| **Harm Avoidance** | Categories and handling | Detailed harm taxonomy with context sensitivity |
| **Honesty** | Truth and transparency standards | Never deceptive, acknowledges uncertainty |

### OpenAI's Model Spec (2025)

OpenAI's specification has undergone significant evolution, with [6+ versions released in 2025](https://model-spec.openai.com/2025-12-18.html) addressing new capabilities and use cases. The specification serves as a "dynamic framework" that adapts based on research and public feedback.

| Version | Key Changes | Significance |
|---------|-------------|--------------|
| **Feb 2025** | Customizability, intellectual freedom | Emphasis on reducing arbitrary restrictions |
| **Apr 2025** | Agent principles added | "Act within agreed-upon scope of autonomy"; control side effects |
| **Sep 2025** | Authority hierarchy restructured | Root → System → Developer → User → Guideline |
| **Dec 2025** | Teen safety (U18 Principles) | Stricter rules for 13-17 users; no romantic roleplay |
| **Dec 2025** | Well-being updates | Self-harm section extended to delusions/mania; isolation prevention |

**Collective Alignment Input:** OpenAI surveyed over 1,000 people worldwide on model behavior preferences. Where public views diverged from the spec, changes were adopted—demonstrating iterative public input into AI behavioral design.

### Benefits of Model Specifications

| Benefit | Description | Evidence/Quantification |
|---------|-------------|------------------------|
| **Transparency** | Public knows intended behavior | All 4 major frontier labs now publish specs publicly |
| **Consistency** | Clear reference for edge cases | Reduces arbitrary variation across deployments |
| **External Scrutiny** | Researchers can evaluate claims | Enables academic analysis of lab commitments |
| **Training Target** | Explicit optimization goal | [Constitutional AI](https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback) shows Pareto improvements when specs guide training |
| **Governance Hook** | Regulators have reference | EU AI Act, NIST AI RMF reference documentation requirements |
| **Public Input** | Democratic participation | OpenAI surveyed 1,000+ people; Anthropic explored collective constitutional AI |

### Limitations and Challenges

| Limitation | Description | Severity | Evidence |
|------------|-------------|----------|----------|
| **Spec-Reality Gap** | Spec doesn't guarantee implementation | High | No third-party verification mechanisms exist |
| **Completeness Challenge** | Can't cover all situations | Medium | Novel scenarios constantly emerge in deployment |
| **Interpretation Variance** | Specs can be read differently | Medium | Natural language inherently ambiguous |
| **Gaming Potential** | Sophisticated systems might letter-comply only | High | Theoretical concern grows with capability |
| **Open-Source Gap** | Open models may lack equivalent safeguards | High | [DeepSeek testing](https://futurism.com/artificial-intelligence/anthropic-claude-soul) showed "absolutely no blocks whatsoever" per Anthropic |
| **Verification Difficulty** | Hard to verify genuine compliance | High | Current evaluations test behavior, not internalization |

## The Spec-Compliance Gap

### Why Specs Don't Guarantee Behavior

| Factor | Description | Consequence |
|--------|-------------|-------------|
| **Training Imperfection** | Training doesn't perfectly achieve spec | Behavioral drift |
| **Specification Ambiguity** | Natural language allows multiple interpretations | Unintended behaviors |
| **Distribution Shift** | New situations not covered by spec | Unpredictable responses |
| **Capability Limitations** | Model may not understand spec fully | Misapplication |
| **Deception Potential** | Model could understand but not comply | Strategic non-compliance |

### Verification Challenges

| Challenge | Description | Status |
|-----------|-------------|--------|
| **Behavioral Testing** | Test all spec provisions | Incomplete coverage possible |
| **Internal Alignment** | Verify genuine vs performed compliance | Difficult |
| **Edge Case Discovery** | Find situations spec doesn't cover | Ongoing challenge |
| **Adversarial Compliance** | Detect gaming behavior | Open problem |

## Scalability Analysis

### How Specs Scale

| Factor | Current | Future Systems |
|--------|---------|----------------|
| **Spec Complexity** | Manageable | May need to grow with capability |
| **Verification** | Difficult | Likely harder with capability |
| **Enforcement** | Training-based | Unclear mechanisms |
| **Gaming Risk** | Present | Expected to increase |

### Superintelligence Considerations

For superintelligent systems, model specs face fundamental challenges:

| Challenge | Description | Status |
|-----------|-------------|--------|
| **Interpretation** | SI might interpret specs unexpectedly | Fundamental uncertainty |
| **Completeness** | Can't anticipate all situations | Likely impossible |
| **Gaming** | SI could find loopholes | Severe concern |
| **Enforcement** | How to enforce on more capable system? | Open problem |

## Current Adoption & Investment

| Metric | Value | Source/Notes |
|--------|-------|--------------|
| **Annual Investment** | \$10-30M/year | Internal lab work on spec development, training integration |
| **Adoption Level** | Universal among frontier labs | Anthropic, OpenAI, Google DeepMind, Meta all publish documentation |
| **Enterprise AI Adoption** | 78% of organizations | [McKinsey survey 2024](https://mckinsey.com)—up from 55% in 2023 |
| **Enterprise AI Spending** | \$17B in 2025 | 3.2x increase from \$11.5B in 2024 ([Menlo Ventures](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-enterprise/)) |
| **Regulatory Momentum** | 75 countries active | 21.3% increase in AI legislative actions in 2024 |
| **Open-Source Gap** | Significant | Models without specifications proliferating globally |

### Differential Progress Analysis

| Factor | Assessment | Rationale |
|--------|------------|-----------|
| **Safety Benefit** | Medium-High | Enables external accountability; clarifies lab commitments |
| **Capability Benefit** | Low-Medium | Clearer behavioral targets can improve usefulness within constraints |
| **Governance Integration** | High | Provides foundation for regulation, auditing, liability frameworks |
| **Overall Balance** | Safety-leaning | Primary value is transparency and accountability, not capability advancement |

## Relationship to Other Approaches

### Integration with Training Methods

- **<EntityLink id="E451">Constitutional AI</EntityLink>**: Specs inform constitutional principles
- **<EntityLink id="E259">RLHF</EntityLink>**: Specs guide rater instructions
- **Evaluation**: Specs define test criteria

### Complementary Approaches

| Approach | Relationship to Specs |
|----------|----------------------|
| **Interpretability** | Could verify spec compliance at mechanistic level |
| **Red Teaming** | Tests spec provisions adversarially |
| **Formal Verification** | Could prove spec compliance for limited domains |

## Best Practices for Model Specs

### What Good Specs Include

| Element | Purpose | Example |
|---------|---------|---------|
| **Clear Hierarchy** | Resolve conflicts | "When X and Y conflict, prioritize X" |
| **Explicit Edge Cases** | Reduce ambiguity | Specific scenario guidance |
| **Reasoning Transparency** | Enable understanding | Explain why rules exist |
| **Version History** | Track changes | Document evolution |
| **Evaluation Criteria** | Enable testing | How to measure compliance |

### Common Pitfalls

| Pitfall | Description | Mitigation |
|---------|-------------|------------|
| **Vague Language** | "Be helpful" without specifics | Operationalize principles |
| **Incomplete Coverage** | Missing important situations | Systematic scenario analysis |
| **Conflicting Rules** | Contradictory provisions | Explicit hierarchy |
| **No Verification** | Can't test compliance | Include test criteria |

## Key Uncertainties & Research Directions

### Open Questions

1. **How to verify spec compliance at scale?** Current testing can't cover all cases; behavioral tests don't verify internalization
2. **Can specs prevent sophisticated gaming?** Letter vs spirit compliance becomes critical as models become more capable
3. **What's the right level of specificity?** Too vague allows interpretation variance; too rigid can't handle novel situations
4. **How should specs evolve?** OpenAI's 6+ versions in 2025 shows rapid iteration; backward compatibility unclear
5. **What about open-source models?** Specs are voluntary; models trained without safeguards proliferate globally

### Research Priorities

| Direction | Purpose | Priority | Current Status |
|-----------|---------|----------|----------------|
| **Formal Spec Languages** | Reduce natural language ambiguity | Medium | Academic research ongoing |
| **Compliance Verification** | Test adherence beyond behavioral observation | High | Major gap; no robust methods |
| **Spec Completeness** | Cover edge cases systematically | Medium | Labs iterating rapidly |
| **Cross-Lab Standardization** | Enable comparison and interoperability | Medium | [Model Context Protocol](https://guptadeepak.com/the-complete-guide-to-model-context-protocol-mcp-enterprise-adoption-market-trends-and-implementation-strategies/) emerging |
| **Democratic Input Mechanisms** | Scale public participation | Medium | OpenAI surveyed 1,000+; Anthropic explored collective CAI |
| **Interpretability Integration** | Verify specs at mechanistic level | High | Early research stage |

### Emerging Standards and Protocols

The [Model Context Protocol (MCP)](https://guptadeepak.com/the-complete-guide-to-model-context-protocol-mcp-enterprise-adoption-market-trends-and-implementation-strategies/), introduced by Anthropic in November 2024, represents a move toward standardization of how AI systems integrate with external tools. Within one year, MCP achieved industry-wide adoption with backing from OpenAI, Google, Microsoft, AWS, and governance under the Linux Foundation. While MCP focuses on tool integration rather than behavioral specifications, it demonstrates the potential for cross-lab standardization that could extend to behavioral specs.

## Sources & Resources

### Primary Model Specifications

| Source | URL | Key Contributions |
|--------|-----|------------------|
| **Anthropic's Claude Soul Document** | [Simon Willison's Analysis](https://simonwillison.net/2025/Dec/2/claude-soul-document/) | 14,000-token document embedded in weights; first public view of identity/ethics training |
| **OpenAI's Model Spec** | [model-spec.openai.com](https://model-spec.openai.com/2025-12-18.html) | Living document with 6+ versions in 2025; authority hierarchy, agent principles |
| **Meta Llama Model Cards** | [llama.com/docs/model-cards](https://www.llama.com/docs/model-cards-and-prompt-formats/) | Comprehensive per-model documentation; Llama Guard safety system |
| **Google Gemini Model Cards** | [modelcards.withgoogle.com](https://modelcards.withgoogle.com/assets/documents/gemini-2.5-pro.pdf) | Technical documentation for each model release |

### Foundational Research

| Paper | Authors | Significance |
|-------|---------|--------------|
| [Model Cards for Model Reporting](https://arxiv.org/abs/1810.03993) | Mitchell et al. (2019) | 2,273+ citations; established ML documentation framework |
| [Constitutional AI: Harmlessness from AI Feedback](https://www.anthropic.com/research/constitutional-ai-harmlessness-from-ai-feedback) | Anthropic (2023) | Demonstrated Pareto improvement with principle-based training |
| [Collective Constitutional AI](https://www.anthropic.com/research/collective-constitutional-ai-aligning-a-language-model-with-public-input) | Anthropic | First instance of public input directing LLM behavior via written specs |

### Industry Analysis

| Source | Focus | Key Finding |
|--------|-------|-------------|
| [OpenAI Model Spec Analysis](https://siliconangle.com/2024/05/08/openais-evolving-model-spec-aims-guide-behavior-ai-models/) | SiliconANGLE | "Needs further adoption...other AI providers must fall into line" |
| [McKinsey AI Survey 2024](https://mckinsey.com) | Enterprise adoption | 78% of organizations using AI, up from 55% in 2023 |
| [State of Enterprise AI 2025](https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-enterprise/) | Menlo Ventures | Enterprise AI surged from \$1.7B to \$17B since 2023 |

### Related Documentation

| Focus Area | Relevance |
|------------|-----------|
| **<EntityLink id="E451">Constitutional AI</EntityLink>** | Specs inform constitutional principles for training |
| **<EntityLink id="E259">RLHF</EntityLink>** | Specs guide human rater instructions |
| **<EntityLink id="E128">AI Evaluation</EntityLink>** | Specs define test criteria for verification |
| **<EntityLink id="E252">Responsible Scaling Policies</EntityLink>** | Specs integrate with capability thresholds |

---

## AI Transition Model Context

Model specifications relate to the <EntityLink id="ai-transition-model" /> through:

| Factor | Parameter | Impact |
|--------|-----------|--------|
| <EntityLink id="E205" /> | <EntityLink id="E264" /> | Specs enable transparent safety practices and external accountability |
| <EntityLink id="deployment-decisions" /> | Deployment standards | Specs provide reference for safe deployment |

Model specs contribute to safety infrastructure but don't solve the fundamental alignment problem - they're necessary but not sufficient for safe AI development.