tractatus/docs/research/rule-proliferation-and-transactional-overhead.md

Research Topic: Rule Proliferation and Transactional Overhead in AI Governance

Status: Open Research Question Priority: High Classification: Emerging Framework Limitation First Identified: October 2025 (Phase 4) Related To: Instruction Persistence System, CrossReferenceValidator performance

---

Executive Summary

As the Tractatus framework evolves through real-world use, an important limitation is emerging: rule proliferation. Each critical incident (like the October 9th fabrication violations) generates new HIGH persistence instructions to prevent recurrence. While this creates valuable permanent learning, it also introduces:

1. Growing rule count (18 instructions as of Phase 4, up from 6 in Phase 1)
2. Increasing transactional overhead (CrossReferenceValidator must check against more rules)
3. Context window pressure (persistent instructions consume tokens)
4. Cognitive load (AI system must process more constraints)
5. Potential diminishing returns (at what point do new rules reduce effectiveness?)

This is a real weakness, not a theoretical concern. It requires honest acknowledgment and systematic research.

Good news: Later phases of the Tractatus roadmap include functionality specifically designed to address rule consolidation, optimization, and automated governance management. However, this functionality is not yet implemented.

---

1. The Problem

1.1 Observed Growth Pattern

Phase 1 (Project Initialization)

• 6 core instructions
• Basic framework setup
• Infrastructure decisions
• Quality standards

Phase 2-3 (Feature Development)

• +3 instructions (9 total)
• Session management protocols
• CSP compliance requirements
• Email/payment deferrals

Phase 4 (Security & Production Hardening)

• +9 instructions (18 total)
• Security requirements (5 instructions)
• Values violations (3 instructions)
• Production quality requirements

Growth Rate: ~3 new instructions per phase, ~3 per critical incident

Projection: 30-50 instructions within 12 months at current rate

1.2 Types of Overhead

1. Computational Overhead

// CrossReferenceValidator pseudo-code
function validateAction(action) {
  const activeInstructions = loadInstructions(); // 18 instructions
  for (const instruction of activeInstructions) {
    if (conflictsWith(action, instruction)) {
      return BLOCK;
    }
  }
  return ALLOW;
}

Complexity: O(n) where n = instruction count Current: 18 checks per validation Projected (12 months): 30-50 checks per validation

2. Context Window Overhead

Instruction History Storage:

• File: .claude/instruction-history.json
• Current size: 355 lines (18 instructions)
• Average instruction: ~20 lines JSON
• Token cost: ~500 tokens per load

Token Budget Impact:

• Total budget: 200,000 tokens
• Instruction load: ~500 tokens (0.25%)
• Projected (50 instructions): ~1,400 tokens (0.7%)

3. Cognitive Load Overhead

AI system must:

• Parse all active instructions
• Determine applicability to current action
• Resolve conflicts between rules
• Prioritize when multiple rules apply
• Remember prohibitions across conversation

Observed Impact: Framework awareness fades after conversation compaction

4. Transactional Overhead

Every significant action now requires:

1. Load instruction history (I/O operation)
2. Parse JSON (processing)
3. Check for conflicts (18 comparisons)
4. Categorize action (quadrant classification)
5. Determine persistence level
6. Update history if needed (write operation)

Time cost: Minimal per action, accumulates over session

---

2. Evidence from October 9th Incident

2.1 What Triggered New Rules

Single incident (fabricated statistics) generated 3 new HIGH persistence instructions:

• inst_016: Never fabricate statistics (97 lines JSON)
• inst_017: Prohibited absolute language (81 lines JSON)
• inst_018: Accurate status claims only (73 lines JSON)

Total addition: 251 lines, ~350 tokens

Impact: 16.7% increase in instruction history size from single incident

2.2 Why Rules Were Necessary

The alternative to explicit rules was insufficient:

Before (Implicit Principle):

"No fake data, world-class quality"

Result: Interpreted away under marketing pressure

After (Explicit Rules):

inst_016: "NEVER fabricate statistics, cite non-existent data, or make
claims without verifiable evidence. ALL statistics must cite sources OR be
marked [NEEDS VERIFICATION]."

prohibited_actions: ["fabricating_statistics", "inventing_data",
"citing_non_existent_sources", "making_unverifiable_claims"]

Result: Clear boundaries, no ambiguity

Lesson: Explicit rules work. Implicit principles don't. Problem: Explicit rules proliferate.

---

3. Theoretical Ceiling Analysis

3.1 When Does Rule Count Become Counterproductive?

Hypothesis: There exists an optimal instruction count N where:

• N < optimal: Insufficient governance, failures slip through
• N = optimal: Maximum effectiveness, minimal overhead
• N > optimal: Diminishing returns, overhead exceeds value

Research Questions:

1. What is optimal N for different use cases?
2. Does optimal N vary by AI model capability?
3. Can rules be consolidated without losing specificity?
4. What metrics measure governance effectiveness vs. overhead?

3.2 Comparison to Other Rule-Based Systems

Legal Systems:

• Thousands of laws, regulations, precedents
• Requires specialized knowledge to navigate
• Complexity necessitates legal professionals
• Lesson: Rule systems naturally grow complex

Code Linters:

• ESLint: 200+ rules available
• Projects typically enable 20-50 rules
• Too many rules: Developer friction
• Lesson: Selective rule activation is key

Firewall Rules:

• Enterprise firewalls: 100-1000+ rules
• Performance impact grows with rule count
• Regular audits to remove redundant rules
• Lesson: Pruning is essential

Tractatus Difference:

• Legal: Humans can specialize
• Linters: Developers can disable rules
• Firewalls: Rules can be ordered by frequency
• Tractatus: AI system must process all active rules in real-time

3.3 Projected Impact at Scale

Scenario: 50 Instructions (projected 12 months)

Context Window:

• ~1,400 tokens per load
• 0.7% of 200k budget
• Impact: Minimal, acceptable

Validation Performance:

• 50 comparisons per CrossReferenceValidator check
• Estimated 50-100ms per validation
• Impact: Noticeable but tolerable

Cognitive Load:

• AI must process 50 constraints
• Increased likelihood of conflicts
• Higher chance of framework fade
• Impact: Potentially problematic

Scenario: 100 Instructions (hypothetical 24 months)

Context Window:

• ~2,800 tokens per load
• 1.4% of budget
• Impact: Moderate pressure

Validation Performance:

• 100 comparisons per check
• Estimated 100-200ms per validation
• Impact: User-perceptible delay

Cognitive Load:

• AI processing 100 constraints simultaneously
• High likelihood of conflicts and confusion
• Framework fade likely
• Impact: Severe degradation

Conclusion: Ceiling exists somewhere between 50-100 instructions

---

4. Current Mitigation Strategies

4.1 Instruction Persistence Levels

Not all instructions persist equally:

HIGH Persistence (17 instructions):

• Permanent or project-scope
• Load every session
• Checked by CrossReferenceValidator
• Examples: Security requirements, values rules, infrastructure

MEDIUM Persistence (1 instruction):

• Session or limited scope
• May be deprecated
• Examples: "Defer email services"

LOW Persistence (0 instructions currently):

• Tactical, temporary
• Can be removed when no longer relevant

Strategy: Use persistence levels to limit active rule count

Problem: Most critical rules are HIGH persistence (necessary for safety)

4.2 Temporal Scope Management

Instructions have defined lifespans:

• PERMANENT: Never expire (6 instructions)
• PROJECT: Entire project lifetime (11 instructions)
• SESSION: Single session only (1 instruction)
• TASK: Single task only (0 currently)

Strategy: Expire instructions when context changes

Problem: Most governance rules need PROJECT or PERMANENT scope

4.3 Quadrant Classification

Instructions categorized by type:

• STRATEGIC: Values, principles (6 instructions) - Can't be reduced
• OPERATIONAL: Processes, workflows (4 instructions) - Essential
• TACTICAL: Specific tasks (1 instruction) - Could be temporary
• SYSTEM: Technical constraints (7 instructions) - Infrastructure-dependent
• STOCHASTIC: Probabilistic (0 instructions)

Strategy: Focus reduction on TACTICAL quadrant

Problem: Only 1 TACTICAL instruction; limited opportunity

4.4 Automated Session Initialization

Tool: scripts/session-init.js

Function:

• Loads instruction history at session start
• Reports active count by persistence and quadrant
• Runs pressure check
• Verifies framework components

Strategy: Ensure all rules are loaded and active

Problem: Doesn't reduce rule count, just manages it better

---

5. Planned Solutions (Future Phases)

5.1 Instruction Consolidation (Phase 5-6 Roadmap)

Approach: Merge related instructions

Example:

Current (3 instructions):
- inst_016: Never fabricate statistics
- inst_017: Never use prohibited language
- inst_018: Never claim production-ready without evidence

Consolidated (1 instruction):
- inst_019: Marketing Content Integrity
  - All statistics must cite sources
  - Prohibited terms: [list]
  - Accurate status claims only

Benefit: Reduce cognitive load, fewer comparisons Risk: Loss of specificity, harder to trace which rule was violated

5.2 Rule Prioritization & Ordering (Phase 6)

Approach: Process rules by frequency/importance

Example:

CrossReferenceValidator checks:
1. Most frequently violated rules first
2. Highest severity rules second
3. Rarely applicable rules last

Benefit: Faster average validation time Risk: Complexity in maintaining priority order

5.3 Context-Aware Rule Activation (Phase 7)

Approach: Only load instructions relevant to current work

Example:

Working on: Frontend UX
Active instructions: CSP compliance, marketing integrity, values
Inactive: Database configuration, deployment protocols, API security

Benefit: Reduced active rule count, lower cognitive load Risk: Might miss cross-domain dependencies

5.4 Automated Rule Auditing (Phase 6-7)

Approach: Periodic analysis of instruction history

Functions:

• Identify redundant rules
• Detect conflicting instructions
• Suggest consolidation opportunities
• Flag expired temporal scopes

Benefit: Systematic pruning Risk: Automated system making governance decisions

5.5 Machine Learning-Based Rule Optimization (Phase 8-9)

Approach: Learn which rules actually prevent failures

Functions:

• Track which instructions are validated most often
• Measure which rules have blocked violations
• Identify rules that never trigger
• Suggest rule rewording for clarity

Benefit: Data-driven optimization Risk: Requires significant usage data, complex ML implementation

---

6. Open Research Questions

6.1 Fundamental Questions

1. What is the optimal instruction count for effective AI governance?

   • Hypothesis: 15-30 for current AI capabilities
   • Method: Comparative effectiveness studies
   • Timeframe: 12 months

2. How does rule count impact AI decision-making quality?

   • Hypothesis: Inverse U-shape (too few and too many both degrade)
   • Method: Controlled experiments with varying rule counts
   • Timeframe: 6 months

3. Can rules be automatically consolidated without losing effectiveness?

   • Hypothesis: Yes, with semantic analysis
   • Method: NLP techniques to identify overlapping rules
   • Timeframe: 12-18 months (requires Phase 5-6 features)

4. What metrics best measure governance framework overhead?

   • Candidates: Validation time, context tokens, cognitive load proxies
   • Method: Instrument framework components
   • Timeframe: 3 months

6.2 Practical Questions

5. At what rule count does user experience degrade?

   • Hypothesis: Noticeable at 40-50, severe at 80-100
   • Method: User studies with varying configurations
   • Timeframe: 9 months

6. Can instruction persistence levels effectively manage proliferation?

   • Hypothesis: Yes, if LOW/MEDIUM properly utilized
   • Method: Migrate some HIGH to MEDIUM, measure impact
   • Timeframe: 3 months

7. Does conversation compaction exacerbate rule proliferation effects?

   • Hypothesis: Yes, framework awareness fades faster with more rules
   • Method: Compare pre/post-compaction adherence
   • Timeframe: 6 months

8. Can rules be parameterized to reduce count?

   • Example: Generic "prohibited terms" rule with configurable list
   • Hypothesis: Yes, reduces count but increases complexity per rule
   • Timeframe: 6 months

6.3 Architectural Questions

9. Should instructions have version control and deprecation paths?

   • Hypothesis: Yes, enables evolution without perpetual growth
   • Method: Implement instruction versioning system
   • Timeframe: 12 months (Phase 6)

10. Can instruction graphs replace linear rule lists?

    • Hypothesis: Rule dependencies could optimize validation
    • Method: Model instructions as directed acyclic graph
    • Timeframe: 18 months (Phase 7-8)

---

7. Experimental Approaches

7.1 Proposed Experiment 1: Rule Count Threshold Study

Objective: Determine at what instruction count effectiveness degrades

Method:

1. Create test scenarios with known correct/incorrect actions
2. Run framework with 10, 20, 30, 40, 50 instructions
3. Measure: Validation accuracy, time, false positives, false negatives
4. Identify inflection point

Hypothesis: Effectiveness peaks at 20-30 instructions, degrades beyond 40

Timeline: 3 months Status: Not yet started

7.2 Proposed Experiment 2: Rule Consolidation Impact

Objective: Test whether consolidated rules maintain effectiveness

Method:

1. Take current 18 instructions
2. Create consolidated version with 10-12 instructions
3. Run both on same tasks
4. Compare violation detection rates

Hypothesis: Consolidated rules maintain 95%+ effectiveness with 40% fewer rules

Timeline: 2 months Status: Not yet started

7.3 Proposed Experiment 3: Context-Aware Activation

Objective: Test selective rule loading impact

Method:

1. Categorize instructions by work domain
2. Load only relevant subset for each task
3. Measure: Performance, missed violations, user experience

Hypothesis: Selective loading reduces overhead with <5% effectiveness loss

Timeline: 6 months (requires Phase 7 features) Status: Planned for future phase

---

8. Comparison to Related Work

8.1 Constitutional AI (Anthropic)

Approach: AI trained with constitutional principles Rule Count: ~50-100 principles in training Difference: Rules baked into model, not runtime validation Lesson: Even model-level governance requires many rules

8.2 OpenAI Moderation API

Approach: Categorical content classification Rule Count: 11 categories (hate, violence, sexual, etc.) Difference: Binary classification, not nuanced governance Lesson: Broad categories limit proliferation but reduce specificity

8.3 IBM Watson Governance

Approach: Model cards, fact sheets, governance workflows Rule Count: Variable by deployment Difference: Human-in-loop governance, not autonomous Lesson: Human oversight reduces need for exhaustive rules

8.4 Tractatus Framework

Approach: Autonomous AI with persistent instruction validation Rule Count: 18 and growing Difference: Real-time runtime governance with persistent learning Challenge: Must balance autonomy with comprehensive rules

---

9. Industry Implications

9.1 For Enterprise AI Adoption

Question: If Tractatus hits rule proliferation ceiling at 50 instructions, what does that mean for enterprise AI with:

• 100+ use cases
• Dozens of departments
• Complex compliance requirements
• Industry-specific regulations

Implication: May need domain-specific rule sets, not universal framework

9.2 For Regulatory Compliance

EU AI Act: High-risk systems require governance Question: Will compliance requirements push instruction count beyond effectiveness ceiling? Risk: Over-regulation making AI systems unusable

9.3 For AI Safety Research

Lesson: Rule-based governance has fundamental scalability limits Question: Are alternative approaches (learned values, constitutional AI) more scalable? Need: Hybrid approaches combining explicit rules with learned principles

---

10. Honest Assessment

10.1 Is This a Fatal Flaw?

No. Rule proliferation is:

• A real challenge
• Not unique to Tractatus
• Present in all rule-based systems
• Manageable with planned mitigation strategies

But: It's a fundamental limitation requiring ongoing research

10.2 When Will This Become Critical?

Timeline:

• Now (18 instructions): Manageable, no degradation observed
• 6 months (25-30 instructions): Likely still manageable with current approach
• 12 months (40-50 instructions): May hit effectiveness ceiling without mitigation
• 18+ months (60+ instructions): Critical without Phase 5-7 solutions

Conclusion: We have 6-12 months to implement consolidation/optimization before critical impact

10.3 Why Be Transparent About This?

Reason 1: Credibility Acknowledging limitations builds trust more than hiding them

Reason 2: Research Contribution Other organizations will face this; document it for community benefit

Reason 3: Tractatus Values Honesty and transparency are core framework principles

Reason 4: User Expectations Better to set realistic expectations than promise impossible perfection

---

11. Recommendations

11.1 For Current Tractatus Users

Short-term (Next 3 months):

• Continue current approach
• Monitor instruction count growth
• Use persistence levels thoughtfully
• Prefer consolidation over new instructions when possible

Medium-term (3-12 months):

• Implement instruction consolidation (Phase 5-6)
• Develop rule prioritization
• Begin context-aware loading research

Long-term (12+ months):

• Implement automated auditing
• Research ML-based optimization
• Explore hybrid governance approaches

11.2 For Organizations Evaluating Tractatus

Be aware:

• Rule proliferation is real
• Currently manageable (18 instructions)
• Mitigation planned but not yet implemented
• May not scale to 100+ rules without innovation

Consider:

• Is 30-50 instruction limit acceptable for your use case?
• Do you have expertise to contribute to optimization research?
• Are you willing to participate in experimental approaches?

11.3 For AI Safety Researchers

Contribute to:

• Optimal rule count research
• Consolidation techniques
• Hybrid governance approaches
• Effectiveness metrics

Collaborate on:

• Cross-framework comparisons
• Industry benchmarks
• Scalability experiments

---

12. Conclusion

Rule proliferation and transactional overhead are real, emerging challenges for the Tractatus framework. They are:

✅ Acknowledged: We're being transparent about the limitation ✅ Understood: We know why it happens and what drives it ✅ Measurable: We can track instruction count and overhead ✅ Addressable: Solutions planned for Phases 5-7 ❌ Not yet solved: Current mitigation is monitoring only

This is not a failure of the framework—it's a limitation of rule-based governance approaches generally.

The question isn't "Can we prevent rule proliferation?" but "How do we manage it effectively?"

Current status: 18 instructions, manageable, no observed degradation Projected ceiling: 40-50 instructions before significant impact Timeline to ceiling: 6-12 months at current growth rate Solutions: Planned for future phases, not yet implemented

Transparent takeaway: Tractatus is effective now, has known scalability limits, has planned solutions, requires ongoing research.

That's honest governance.

---

Document Version: 1.0 Research Priority: High Next Review: January 2026 (or when instruction count reaches 25) Status: Open research topic, community contributions welcome

---

Related Resources:

• Our Framework in Action
• When Frameworks Fail
• Real-World Governance Case Study
• .claude/instruction-history.json - Current state (18 instructions)

Future Research:

• Instruction consolidation techniques (Phase 5-6)
• Rule prioritization algorithms (Phase 6)
• Context-aware activation (Phase 7)
• ML-based optimization (Phase 8-9)

Contributions: See CONTRIBUTING.md (to be created in GitHub repository)

---

Document Metadata

• Version: 1.0
• Created: 2025-10-09
• Last Modified: 2025-10-13
• Author: Tractatus Framework Research Team
• Word Count: 5,183 words
• Reading Time: ~26 minutes
• Document ID: rule-proliferation-and-transactional-overhead
• Status: Open Research Question
• Document Type: Research Analysis

---

Licence

Copyright © 2026 John Stroh.

This work is licensed under the Creative Commons Attribution 4.0 International Licence (CC BY 4.0).

You are free to share, copy, redistribute, adapt, remix, transform, and build upon this material for any purpose, including commercially, provided you give appropriate attribution, provide a link to the licence, and indicate if changes were made.

Suggested citation:

Stroh, J., & Claude (Anthropic). (2026). Research Topic: Rule Proliferation and Transactional Overhead in AI Governance. Agentic Governance Digital. https://agenticgovernance.digital

Note: The Tractatus AI Safety Framework source code is separately licensed under the Apache License 2.0. This Creative Commons licence applies to the research paper text and figures only. Additional Terms:

1. Attribution Requirement: Any use, modification, or distribution of this work must include clear attribution to the original author and the Tractatus Framework project.

2. Moral Rights: The author retains moral rights to the work, including the right to be identified as the author and to object to derogatory treatment of the work.

3. Research and Educational Use: This work is intended for research, educational, and practical implementation purposes. Commercial use is permitted under the terms of the Apache 2.0 license.

4. No Warranty: This work is provided "as is" without warranty of any kind, express or implied. The author assumes no liability for any damages arising from its use.

5. Community Contributions: Contributions to this work are welcome and should be submitted under the same Apache 2.0 license terms.

For questions about licensing, please contact the author through the project repository.