Models Style Guide

Models Style Guide

This guide defines the standards for analytical models in LongtermWiki. Models should maximize information density while remaining accessible, and help readers make prioritization and strategy decisions.

Prerequisite: All model pages must follow the Common Writing PrinciplesE726Shared writing principles referenced by all domain-specific style guides. Three pillars: epistemic honesty (hedge uncertain claims, use ranges, source confidence levels), language neutrality (avoid...0 — epistemic honesty, language neutrality, and analytical tone. The objectivity rating dimension measures this. For model pages with cost-effectiveness estimates, this is especially critical: always use ranges, show deflators, and include "Why These Numbers Might Be Wrong" sections.

Core Principles

1. Density over brevity — Pack substantive content into every section. A 500-word model with tables and equations beats a 200-word model with bullets.
2. Quantify everything possible — Probabilities, timelines, costs, thresholds. Vague claims waste reader attention.
3. Show structure visually — Tables, diagrams, and equations communicate relationships faster than prose.
4. Paragraphs over bullets — Bullets fragment thinking. Use them only for truly discrete items.
5. Strategic prioritization — Every model should help answer "how important is this and what should we do about it?", not just explain mechanisms.

Core Purpose: Strategic Prioritization

Models exist to help with prioritization and strategy decisions, not just to explain mechanisms. Every model should answer: "How important is this and what should we do about it?"

The knowledge base serves people making strategic decisions about AI safety: researchers deciding what to work on, funders deciding where to allocate resources, policymakers deciding what to regulate, and organizations deciding their focus areas. Models should help them decide what matters most and what to do about it.

Required Strategic Content

Every model must include:

Anti-Pattern: Mechanism Without Magnitude

A model that thoroughly explains how something works but never addresses how important it is fails its core purpose.

Bad example (hypothetical sycophancy model):

"The feedback loop operates through 4 phases over 10 years, with differential equations governing each variable..." (300 lines on mechanism, 0 lines on strategic importance)

Better approach:

"Sycophancy represents approximately 5-15% of near-term AI risk, ranking below core alignment but above most misuse risks. For most safety organizations, this is a secondary priority unless they have specific comparative advantage. The key crux is whether market competition makes sycophancy inevitable — if so, regulatory intervention becomes critical."

Strategic Importance Section Template

Include a section like this in every model:

## Strategic Importance

### Magnitude
- **Share of total AI risk:** [X-Y%]
- **Affected population:** [scope]
- **Timeline:** [when effects materialize]

### Comparative Ranking
| Risk Category | Relative Importance | Reasoning |
|---------------|--------------------:|-----------|
| Core alignment | Higher | [why] |
| This risk | Baseline | - |
| [Other risk] | Lower | [why] |

### Resource Implications
- **Who should work on this:** [actor types]
- **Suggested allocation:** [% of resources]
- **Comparative advantage:** [who is best positioned]

### Key Cruxes
1. If [X], this becomes more important because [Y]
2. If [A], this becomes less important because [B]

Executive Summary Requirement

Every model must have an executive summary that states both what the model does and what it concludes. This summary appears in the description frontmatter and is shown in previews across the site.

The Summary Formula

A good model summary follows this pattern:

"This model [methodology/approach]. It [key finding — trajectory, critical variables, or uncertainty assessment]."

Types of Valid Findings

Summaries should emphasize where things are going and what matters most:

Good summaries:

Bad summaries:

The description field must state what the model does (methodology/approach), include key conclusions with quantified estimates where possible, and be 1-3 sentences (max ~250 characters for good preview display).

Required Sections

Every model should include:

1. Overview (2-3 paragraphs)

State the model's purpose, central question, and key insight. No bullets here — write flowing prose that orients the reader. Explain the central insight in the first paragraph, why it matters in the second, and preview key findings or framework structure in the third.

Bad:

## Overview
- This model looks at X
- Key question: Y
- Main finding: Z

Good:

## Overview

This model analyzes [phenomenon] by decomposing it into [components]. The central question: **[specific question with stakes]?**

The key insight is that [non-obvious conclusion]. This matters because [implication for AI safety/policy].

2. Conceptual Framework

Explain the model's structure. Include at least one of:

• A Mermaid diagram showing relationships
• A mathematical formulation
• A typology table

3. Quantitative Analysis

The heart of the model. Must include:

• Parameter tables with estimates and uncertainty ranges
• Scenario analysis with probability-weighted outcomes
• Sensitivity analysis showing which inputs matter most

4. Strategic Importance

Magnitude, comparative ranking, resource implications, key cruxes, and actionability (see template above).

5. Case Studies or Applications

Concrete examples showing the model applied. Tables comparing cases are ideal.

6. Limitations

Explicit acknowledgment of model weaknesses in flowing prose. Be specific about what the model ignores or gets wrong.

7. Related Models

Links to complementary models in the knowledge base.

Formatting Standards

Tables: Use Extensively

Tables compress information. Use them for parameter estimates with ranges, scenario comparisons, timeline projections, cost/benefit analyses, and threshold indicators.

Minimum table requirements:

• At least 3 columns (simple key-value pairs waste table format)
• At least 4 rows of data
• Header row clearly labeled
• Include units and uncertainty ranges where applicable

Example — Parameter Table:

| Parameter | Best Estimate | Range | Confidence | Source |
|-----------|--------------|-------|------------|--------|
| P(misalignment) | 15% | 5-40% | Low | Expert surveys |
| Time to AGI | 2028 | 2025-2040 | Medium | Metaculus |
| Safety tax | 20% | 10-50% | Medium | Lab estimates |

Example — Scenario Table:

| Scenario | Probability | Outcome | Key Drivers |
|----------|-------------|---------|-------------|
| Coordinated slowdown | 15% | Low risk | International agreement, major incident |
| Competitive race | 45% | Medium-high risk | US-China tension, commercial pressure |
| Unilateral breakout | 25% | Very high risk | Capability surprise, regulatory failure |
| Managed transition | 15% | Low risk | Technical breakthrough in alignment |

Diagrams: Mermaid Required

Every model should include at least one Mermaid diagram. Choose the type that best represents your model's structure.

Flowcharts — causal chains, decision processes, process flows:

flowchart TD
    A[Capability Advance] --> B{Safety Research<br/>Keeps Pace?}
    B -->|Yes| C[Managed Development]
    B -->|No| D[Risk Gap Widens]
    D --> E{Incident Occurs?}
    E -->|Yes| F[Reactive Regulation]
    E -->|No| G[Continued Drift]
    G --> D

Network diagrams — feedback loops, mutual influences, complex interdependencies:

flowchart LR
    A[Factor A] -->|strengthens| B[Factor B]
    B -->|weakens| C[Factor C]
    C -->|influences| A
    B -->|amplifies| D[Factor D]
    D -->|reduces| A

State diagrams — phase transitions, regime changes, reversibility:

stateDiagram-v2
    [*] --> Reversible
    Reversible --> CostlyReversible: Deployment
    CostlyReversible --> PracticallyIrreversible: Lock-in
    PracticallyIrreversible --> AbsolutelyIrreversible: Existential threshold
    AbsolutelyIrreversible --> [*]

Quadrant charts — 2x2 classification, prioritization frameworks:

quadrantChart
    title Risk vs Tractability
    x-axis Low Tractability --> High Tractability
    y-axis Low Risk --> High Risk
    quadrant-1 Urgent Priority
    quadrant-2 Monitor Carefully
    quadrant-3 Opportunistic
    quadrant-4 Core Focus
    Misalignment: [0.3, 0.8]
    Misuse: [0.7, 0.6]
    Accidents: [0.8, 0.4]
    Structural: [0.4, 0.5]

Entity-relationship diagrams — taxonomies, structural relationships between concepts:

erDiagram
    RISK ||--o{ RISK_FACTOR : "has"
    RISK ||--o{ INTERVENTION : "addressed by"
    RISK_FACTOR ||--o{ INDICATOR : "measured by"
    INTERVENTION ||--o{ ACTOR : "implemented by"
    ACTOR ||--o{ RESOURCE : "requires"

Class diagrams — schema-style representations of concept attributes:

classDiagram
    class Risk {
        +name: string
        +severity: 1-5
        +probability: 0-1
        +timeframe: string
    }
    class RiskFactor {
        +name: string
        +type: technical|social|structural
        +measurable: boolean
    }
    class Intervention {
        +name: string
        +effectiveness: 0-1
        +cost: low|medium|high
    }
    Risk "1" --> "*" RiskFactor : influenced by
    Risk "1" --> "*" Intervention : addressed by
    RiskFactor "1" --> "*" Intervention : targeted by

Timeline diagrams — milestone projections, historical development:

timeline
    title Capability Development Timeline
    section Near-term
        2024 : Current state
        2025 : Threshold 1 approached
        2026 : Threshold 1 likely crossed
    section Medium-term
        2027-2028 : Threshold 2 risk
        2029-2030 : Critical period
    section Long-term
        2030+ : High uncertainty

Subgraph groupings — system boundaries, categorized factors:

flowchart TD
    subgraph Technical["Technical Factors"]
        T1[Capability Level]
        T2[Alignment Quality]
    end
    subgraph Social["Social Factors"]
        S1[Racing Dynamics]
        S2[Governance Readiness]
    end
    subgraph Outcomes["Outcomes"]
        O1[Safe Deployment]
        O2[Accident Risk]
    end
    T1 --> O1
    T1 --> O2
    T2 --> O1
    S1 --> O2
    S2 --> O1

For sequences of actor interactions, prefer a table over a sequence diagram (which has rendering issues):

Equations: Show Your Math

Include mathematical formulations where applicable. Use LaTeX:

Inline math for simple expressions: $P(X|Y) = 0.3$

Display math for key equations:

$$
R(t) = R_0 \cdot e^{\alpha t} \cdot (1 + \beta D)
$$

Where:
- $R_0$ = Base reversal cost at deployment
- $\alpha$ = Growth rate (0.1-0.5 per year)
- $t$ = Time since deployment
- $\beta$ = Dependency multiplier
- $D$ = Dependency depth (0 to 1)

Always include variable definitions immediately after the equation, realistic parameter ranges, and intuition for what the equation captures.

Prose: Paragraphs Over Bullets

Bullets should be rare. They're appropriate for truly discrete, unordered items (e.g., list of examples), quick reference lists at end of sections, or items that will be expanded in subsequent sections.

Bad — Bullet brain:

## Racing Dynamics

- Labs compete for capabilities
- Safety work slows deployment
- First-mover advantages exist
- Coordination is difficult
- Racing creates risk

Good — Dense paragraphs:

## Racing Dynamics

Racing dynamics emerge when multiple actors pursue the same capability under competitive pressure. In AI development, labs face a structural tension: safety work requires time and resources that slow deployment, but first-mover advantages in capabilities — talent attraction, data access, revenue, and strategic positioning — create intense pressure to move fast. This produces a classic collective action problem where individually rational choices generate collectively irrational outcomes.

The severity of racing depends on three factors: the perceived magnitude of first-mover advantages, the credibility of competitors' timelines, and the availability of coordination mechanisms. When labs believe winner-take-all dynamics apply, racing pressure intensifies regardless of stated safety commitments.

Methodological Principles

Distinguish Stocks vs. Flows

Stocks are quantities that accumulate (trust level, capability, resources). Flows are rates of change (trust erosion rate, capability growth rate). Models should be clear about which they're describing.

Identify Feedback Loops

Many risks involve feedback loops where effects become causes. Make these explicit.

flowchart LR
    A[Capability Increase] --> B[Deployment Pressure]
    B --> C[Rushed Safety Testing]
    C --> D[Incidents]
    D --> E[Reduced Trust]
    E --> F[Regulatory Response]
    F -->|negative feedback| B
    D -->|positive feedback| A

Positive feedback loops (amplifying): The effect reinforces the cause. Negative feedback loops (stabilizing): The effect counteracts the cause.

Consider Base Rates

Before modeling specific mechanisms, consider: what's the base rate for this type of event?

Distinguish Correlation vs. Causation

When factors co-occur, be explicit about the causal structure:

Avoid False Binary Thresholds

Models often imply sharp cutoffs ("if X > 80%, collapse occurs") when reality involves continuous degradation. Use gradient language ("largely past," "degrading," "limited risk"), acknowledge that most systems degrade continuously, and if using threshold framing, add explicit caveats.

Avoid Naive Multiplicative Formulas

Formulas like P(cascade) = P(A) × P(B) × P(C) × P(D) assume independence when factors are often correlated. Acknowledge correlations explicitly in a table, use influence diagrams instead of formulas, or add caveats about correlation assumptions.

Consider Counterfactuals

Good models should address: "Compared to what?"

Information Density Checklist

Before submitting a model, verify:

• Tables: At least 2 substantive tables (4+ rows, 3+ columns each)
• Diagram: At least 1 Mermaid diagram showing relationships
• Equations: Mathematical formulation where applicable
• Numbers: Probabilities, timelines, or thresholds quantified with ranges
• Scenarios: Multiple scenarios analyzed with probability weights
• Strategic Importance: Magnitude, comparative ranking, resource implications, cruxes
• Paragraphs: Less than 30% of content in bullet points
• Length: Minimum 800 words of substantive content
• Sources: Key claims attributed to sources

Anti-Patterns to Avoid

1. Bullet Lists as Primary Content

Bullets should support paragraphs, not replace them.

2. Vague Qualitative Claims

"Risk is high" → "Risk estimated at 15-30% (median 22%) based on expert surveys"

3. Missing Uncertainty

Always include ranges, not point estimates alone.

4. Tables Without Context

Tables need introductory sentences explaining what they show and key takeaways.

5. Diagrams Without Explanation

Every diagram needs a paragraph explaining what it illustrates and key insights.

6. Orphan Sections

Short sections (< 100 words) should be merged or expanded.

7. Insider Language and False Certainty

"EA organizations should fund this" → Describe specific orgs, use analytical framing. "True Cost: $500K" → "Est. cost: $300K-1M". See Common Writing PrinciplesE726Shared writing principles referenced by all domain-specific style guides. Three pillars: epistemic honesty (hedge uncertain claims, use ranges, source confidence levels), language neutrality (avoid...0.

8. Mechanism Without Magnitude

A model that thoroughly explains how something works but never addresses how important it is fails its core purpose.

Example Model Structure

---
title: [Descriptive Model Name]
description: [This model [methodology]. It [key finding with quantified estimate].]
ratings:
  novelty: [0-10]
  rigor: [0-10]
  actionability: [0-10]
  completeness: [0-10]
---

## Overview
[2-3 paragraphs: purpose, central question, key insight]

## Conceptual Framework
[Structure explanation with diagram]

```mermaid
[diagram here]

Core Model

Mathematical Formulation

$[key equation]$

[Variable definitions and intuition]

Parameter Estimates

Strategic Importance

Magnitude

Comparative Ranking

Resource Implications

Key Cruxes

Analysis

Scenario Analysis

Sensitivity Analysis

[Which parameters matter most and why]

Case Studies

Case 1: [Name]

[Application of model to concrete example]

Implications

[What this model suggests for policy/research/action]

Limitations

[Explicit weaknesses — be specific]

Related Models

• [Link 1] — [relationship]
• [Link 2] — [relationship]

Sources

[Key references]

---

## Advanced Visualization Patterns

### Sensitivity Analysis Tables

Show how conclusions change with different assumptions:

| Parameter | Low Estimate | Central | High Estimate | Conclusion Changes? |
|-----------|--------------|---------|---------------|---------------------|
| Capability growth rate | 10%/yr | 30%/yr | 50%/yr | Yes — timeline shifts 3-5 years |
| Alignment difficulty | Easy | Medium | Hard | Yes — risk estimate changes 2-3x |
| Coordination probability | 10% | 30% | 50% | No — conclusion robust |

### Comparison Matrices

Compare interventions, scenarios, or approaches across multiple dimensions:

| Intervention | Effectiveness | Cost | Feasibility | Time to Impact |
|--------------|---------------|------|-------------|----------------|
| Speed limits | High | Low | Medium | Immediate |
| International treaty | Very High | Medium | Low | 3-5 years |
| Research funding | Medium | Medium | High | 5-10 years |

### Before/After Comparisons

Show how a change affects multiple dimensions:

| Dimension | Before Intervention | After Intervention | Change |
|-----------|---------------------|--------------------|--------|
| Risk level | High (0.7) | Medium (0.4) | -43% |
| Detection time | 2 weeks | 2 days | -86% |
| Recovery cost | \$10B | \$2B | -80% |

### Decision Trees

For models with sequential choices:

```mermaid
flowchart TD
    A{Deploy AI system?} -->|Yes| B{Safety testing complete?}
    A -->|No| C[No immediate risk]
    B -->|Yes| D{Monitoring in place?}
    B -->|No| E[High accident risk]
    D -->|Yes| F[Managed deployment]
    D -->|No| G[Medium accident risk]

Rating Criteria

Models are rated on seven dimensions (0-10 scale, harsh — 7+ is exceptional). See Rating SystemE741This page documents LongtermWiki's content rating system, which combines LLM-graded subscores (focus, novelty, rigor, completeness, objectivity, concreteness, actionability on 0-10 scales) with aut...Quality: 48/100 for the canonical reference.

Focus: Does it answer what the title promises?

Novelty: How original is the framing or analysis?

• 3-4: Useful synthesis or modest extensions
• 5-6: Genuine new framing or connections
• 7+: Novel framework that changes how we think about the problem

Rigor: How well-supported by evidence and logic?

• 3-4: Reasonable extrapolation, some gaps
• 5-6: Most claims sourced with some quantification
• 7+: Strong empirical grounding or formal derivation

Completeness: How thoroughly developed?

• 3-4: Core model sketched, notable gaps
• 5-6: Core model complete, some gaps
• 7+: Comprehensive treatment with edge cases addressed

Objectivity: Epistemic honesty, neutral language, analytical tone?

• 3-4: Estimates without ranges; insider language; one-sided framing
• 5-6: Mostly hedged and neutral; some uncertainty noted
• 7+: All estimates use ranges with caveats; fully accessible; red-teams own conclusions

Concreteness: Specific numbers, examples, recommendations?

Actionability: Does it suggest concrete interventions?

• 3-4: Identifies general areas of concern
• 5-6: Identifies leverage points with some specifics
• 7+: Specific, implementable recommendations with priorities

Review Checklist

Format

• Overview is 2-3 paragraphs of flowing prose (no bullets)
• At least one Mermaid diagram with caption
• Quantitative tables with 3+ columns and uncertainty ranges
• Scenario analysis with probability weights
• Limitations section in prose format
• Related models linked

Diagrams

• Diagram type matches content (flowchart for causation, network for relationships, ER/class for taxonomies)
• Diagram has explanatory caption
• Complex diagrams use subgraphs for grouping
• Color coding is meaningful and explained

Methodology

• No false binary thresholds (or explicitly caveated)
• Multiplicative formulas acknowledge correlations
• Feedback loops identified where relevant
• Stocks vs. flows distinguished
• Base rates considered
• Counterfactual comparisons made

Strategic Content

• Magnitude assessment (share of AI risk, affected population, timeline)
• Comparative ranking against other risks
• Resource implications (who should work on this, suggested allocation)
• Key cruxes identified
• Actionability (what should actors actually do)

Content Consistency

• Every simplifying assumption is explicitly flagged as such (with a pointer to Limitations), never asserted as fact
• Every [0, 1] or other numeric scale has grounded anchors defining what the endpoints mean — write anchors before writing estimates that use the scale
• Rankings with overlapping ranges say "roughly ordered by median" or "suggestive, not definitive" — no definitive ordinal rankings when ranges overlap
• Option value of delay is addressed (additional time lets us learn whether alignment is hard or easy)
• Racing/coordination effects are addressed (does unilateral action just shift activity elsewhere?)
• Recursive dynamics are addressed where relevant (e.g. AI accelerating safety research)
• Effects are distinguished as qualitatively different where appropriate, not just quantitatively shifted

Common Issues to Fix