Model Style Guide

This guide covers core purpose, summary requirements, format requirements, diagram types, methodological principles, and rating criteria for analytical models in this knowledge base.

---

Executive Summary Requirement

Most Critical 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:

Finding Type	Example
Trajectory/Projection	”…projects uplift increasing from 1.5x to 3-5x by 2030”
Critical Variables	”…identifies X and Y as the key variables determining outcomes”
Risk Magnitude	”…estimates this represents 5-15% of total AI risk”
Uncertainty Assessment	”…finds high variance across scenarios; results depend heavily on [assumption]“
Negative Finding	”…finds no significant effect under current conditions, but this changes if [X]“

Examples

Good summaries:

Topic	Summary
Bioweapons uplift	”This model estimates AI’s contribution to bioweapons risk over time. It projects uplift increasing from 1.5x to 3-5x by 2030, with biosecurity evasion posing the greatest concern.”
Racing dynamics	”This model analyzes competitive pressures among frontier labs. It finds the key variable is whether any single lab can maintain >6 month lead; if not, racing dynamics dominate.”
Lock-in probability	”This model assesses paths to irreversible outcomes. Results are highly uncertain (10-60% range) depending on governance assumptions.”
Sycophancy	”This model maps feedback loops in AI training. It finds no clear evidence of runaway dynamics under current training regimes, but identifies reward hacking as the critical variable to monitor.”

Bad summaries:

Summary	Problem
”Analysis of AI bioweapons risk”	No methodology, no conclusion
”This model examines how racing dynamics affect safety”	No finding at all
”Current LLMs provide 1.3x uplift”	Current state only, no trajectory or implications

Frontmatter Format

---
title: "Racing Dynamics Impact Model"
description: "This model analyzes competitive pressures among frontier labs. It estimates a 60-80% probability that racing dynamics reduce safety investment by 30-50% compared to non-competitive scenarios."
quality: 3
lastEdited: "2025-12-26"
ratings:
  novelty: 4
  rigor: 3
  actionability: 4
  completeness: 3
---

The description field:

• Must state what the model does (methodology/approach)
• Must include key conclusions with quantified estimates where possible
• Should be 1-3 sentences (max ~250 characters for good preview display)
• Is shown in entity cards, backlinks, and search results

Displaying the Summary on the Page

Add a ## Summary section at the top of every model page that references the frontmatter:

{/* Replace with actual entity ID from entities.yaml */}
<DataInfoBox entityId="bioweapons-ai-uplift" ratings={frontmatter.ratings} />

## Summary

{frontmatter.description}

## Overview

[Detailed context and background...]

This way the summary text lives only in frontmatter and is rendered on the page via {frontmatter.description}.

Sync with entities.yaml

For now, model descriptions also need to be updated in src/data/entities.yaml. Keep them in sync manually. (TODO: automate this via build script)

---

Core Purpose: Strategic Prioritization

Most Important Principle

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 Goal

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
• 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:

Element	Question Answered	Example
Magnitude Assessment	How big is this problem?	”This affects 10-30% of total AI risk”
Comparative Importance	How does this rank vs. other risks?	”Less important than misalignment, more than job displacement”
Resource Implications	What does this mean for prioritization?	”Warrants 5-10% of safety resources”
Key Cruxes	What beliefs would change the conclusion?	”If X is true, this becomes top priority”
Actionability	What should actors actually do?	”Labs should implement Y, funders should fund Z”

Common Mistake: Mechanism Without Magnitude

Anti-Pattern

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

Bad example (from a 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)

What’s missing:

• Is sycophancy a top-5 AI risk or a minor concern?
• Should safety orgs prioritize this over alignment research?
• How does this compare to racing dynamics or concentration risks?
• What beliefs would change whether this matters?

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]

---

Part 1: Structure Requirements

Overview Section

Write 2-3 paragraphs of flowing prose (no bullet points). The overview should:

• Explain the model’s central insight in the first paragraph
• Describe why understanding this matters in the second paragraph
• Preview key findings or framework structure in the third paragraph

Required Sections

Section	Purpose	Format
Overview	Central insight and importance	2-3 paragraphs of prose
Conceptual Framework	Visual structure of the model	Mermaid diagram + explanation
Quantitative Analysis	Numbers, estimates, projections	Tables with uncertainty ranges
Scenario Analysis	Probability-weighted futures	3-5 scenarios with probabilities
Limitations	What the model cannot do	Flowing prose, specific caveats
Related Models	Connections to other models	Linked list

Tables

Include tables with 3+ columns and 4+ rows. Tables should add structured information beyond what prose conveys.

Good table characteristics:

• Uncertainty ranges (low/central/high estimates)
• Comparison across multiple dimensions
• Clear headers that explain what each column means
• Sources or confidence levels where relevant

---

Part 2: Diagram Types

Models should include at least one diagram. Choose the type that best represents your model’s structure.

Flowcharts (Process/Causation)

Use for: Showing how things lead to other things, causal chains, decision processes.

Loading diagram...

flowchart TD
    A[Initial Condition] --> B[Intermediate State]
    B --> C[Outcome 1]
    B --> D[Outcome 2]

Network Diagrams (Relationships)

Use for: Showing how multiple factors influence each other, feedback loops, complex interdependencies.

Loading diagram...

State Diagrams (Transitions)

Use for: Showing how systems move between discrete states, regime changes, phase transitions.

Loading diagram...

Interactions Over Time

For showing how different actors or systems interact in sequence, attack chains, or response protocols, use a table rather than a sequence diagram (which has rendering issues):

Step	Actor	Action	Target
1	Actor A	Initial action	System
2	System	Alert triggered	Defender
3	Defender	Response deployed	System
4	System	Action blocked	Actor A

For simple flows, a basic flowchart works well:

Loading diagram...

Entity Relationship Diagrams (Schemas)

Use for: Showing structural relationships between concepts, taxonomies, classification systems.

Loading diagram...

Quadrant Diagrams (2x2 Matrices)

Use for: Classifying items along two dimensions, prioritization frameworks, strategic positioning.

Loading diagram...

Timeline Diagrams

Use for: Showing progression over time, milestone projections, historical development.

Loading diagram...

Subgraph Groupings

Use for: Showing categories of related items, system boundaries, domain separation.

Loading diagram...

---

Part 3: Schema Structures

For models describing structural relationships, use schema-style representations.

When to Use Schemas

• Defining taxonomy of related concepts
• Showing how entities relate to each other
• Describing data structures or classification systems
• Mapping stakeholder relationships

Schema Example: Risk Factor Taxonomy

Loading diagram...

Schema Example: Actor Relationships

Loading diagram...

---

Part 4: Methodological Principles

Distinguish Stocks vs. Flows

Tip

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.

Concept	Stock (Level)	Flow (Rate)
Trust	Current trust level (0-100%)	Trust erosion rate (%/year)
Capability	Current capability score	Capability growth rate
Safety margin	Current margin size	Margin compression rate

Identify Feedback Loops

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

Loading diagram...

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?

Event Type	Historical Base Rate	AI-Specific Adjustment
Major infrastructure failure	~0.5/year globally	Unknown multiplier
Technology-driven job displacement	~2-5%/decade	Potentially 10x faster
Great power conflict	~0.5%/year	Unknown effect

Distinguish Correlation vs. Causation

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

Relationship	Description	Implication
A causes B	Intervening on A changes B	Target A to affect B
B causes A	Intervening on A doesn’t change B	Target B instead
C causes both	A and B are correlated but independent	Target C to affect both
A and B cause each other	Feedback loop	Consider system dynamics

Avoid False Binary Thresholds

Common Problem

Models often imply sharp cutoffs (“if X > 80%, collapse occurs”) when reality involves continuous degradation.

Better approach:

• Use gradient language: “largely past,” “degrading,” “limited risk”
• Acknowledge that most systems degrade continuously
• If using threshold framing, add explicit caveats

Avoid Naive Multiplicative Formulas

Common Problem

Formulas like P(cascade) = P(A) × P(B) × P(C) × P(D) assume independence when factors are often correlated.

Better approach:

1. Acknowledge correlations explicitly in a table
2. Use influence diagrams instead of formulas
3. If using formulas, add caveats about correlation assumptions

Consider Counterfactuals

Good models should address: “Compared to what?”

Comparison	What it reveals
vs. no AI development	Total effect of AI
vs. slower development	Effect of racing
vs. different governance	Effect of policy choices
vs. different actors	Effect of who controls AI

---

Part 5: Rating System

Each model is rated on four dimensions (1-5 scale). These ratings appear in the model’s info box.

Novelty (1-5)

How much does this model add beyond existing frameworks?

Score	Description	Example
1	Restates common knowledge	”AI could be dangerous”
2	Minor variation on existing model	Adding one factor to known framework
3	Useful synthesis or new framing	Combining two existing models in novel way
4	Significant new insight	New mechanism or relationship not previously articulated
5	Paradigm-shifting framework	Fundamentally new way of understanding the problem

Rigor (1-5)

How well-supported and internally consistent is the model?

Score	Description	Characteristics
1	Speculation	No sources, hand-wavy reasoning
2	Plausible	Some logical basis, few sources
3	Well-reasoned	Clear logic, some empirical grounding
4	Strong evidence base	Multiple sources, quantified where possible
5	Rigorous analysis	Comprehensive evidence, sensitivity analysis, peer review

Actionability (1-5)

How useful is this model for decision-making?

Score	Description	Example outputs
1	Abstract only	”Things are complex”
2	General direction	”We should be careful”
3	Specific considerations	”These 3 factors matter most”
4	Concrete recommendations	”Prioritize X intervention over Y because Z”
5	Decision-ready	Clear decision criteria, thresholds, action triggers

Completeness (1-5)

How thoroughly does the model cover its domain?

Score	Description	Missing elements
1	Sketch	Most components missing
2	Partial	Key components missing
3	Adequate	Core model complete, some gaps
4	Comprehensive	Thorough coverage, minor gaps
5	Exhaustive	All relevant factors, edge cases, interactions

How to Assign Ratings

When reviewing a model, ask:

1. Novelty: “Have I seen this idea before? Does it change how I think?”
2. Rigor: “Would I bet money on these claims? What’s the evidence quality?”
3. Actionability: “Could I make a decision based on this? What would I do differently?”
4. Completeness: “What’s missing? Would adding more change the conclusions?”

---

Part 6: Advanced Visualization Ideas

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:

Intervention	Effectiveness	Cost	Feasibility	Time to Impact	Overall
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:

Loading diagram...

---

Part 7: 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, etc.)
• 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

Ratings

• All four ratings assigned (novelty, rigor, actionability, completeness)
• Ratings are justified by content quality
• Ratings are consistent with similar models

Common Issues to Fix

Issue	Fix
Binary threshold language	Use gradient language (“degrading,” “largely past”)
Multiplicative formula without caveat	Add correlation acknowledgment
Missing uncertainty ranges	Add low/central/high estimates
Flowchart for structural relationships	Use entity-relationship or class diagram
No feedback loops shown	Add arrows showing circular dependencies
Ratings don’t match quality	Adjust ratings or improve content