Research Context & Scope
Development Context
Tractatus was developed over six months (April–October 2025) in progressive stages that evolved into a live demonstration of its capabilities in the form of a single-project context (https://agenticgovernance.digital). Observations derive from direct engagement with Claude Code (Anthropic's Sonnet 4.5 model) across approximately 500 development sessions. This is exploratory research, not controlled study.
Aligning advanced AI with human values is among the most consequential challenges we face. As capability growth accelerates under big tech momentum, we confront a categorical imperative: preserve human agency over values decisions, or risk ceding control entirely.
The framework emerged from practical necessity. During development, we observed recurring patterns where AI systems would override explicit instructions, drift from established values constraints, or silently degrade quality under context pressure. Traditional governance approaches (policy documents, ethical guidelines, prompt engineering) proved insufficient to prevent these failures.
Instead of hoping AI systems "behave correctly," Tractatus proposes structural constraints where certain decision types require human judgment. These architectural boundaries can adapt to individual, organizational, and societal norms—creating a foundation for bounded AI operation that may scale more safely with capability growth.
This led to the central research question: Can governance be made architecturally external to AI systems rather than relying on voluntary AI compliance? If this approach can work at scale, Tractatus may represent a turning point—a path where AI enhances human capability without compromising human sovereignty.
Current Research Focus: Christopher Alexander Integration
Integrated: October 2025 | Status: Monitoring for Effectiveness
The framework has integrated five architectural principles from Christopher Alexander's work on living systems, pattern languages, and wholeness (The Timeless Way of Building, A Pattern Language, The Nature of Order). These principles now guide all framework evolution:
Research Question: Can architectural principles from physical architecture domain (Alexander) be faithfully adapted to AI governance with measurable effectiveness? We are monitoring framework behavior through audit log analysis and seeking empirical validation.
Research Collaboration Opportunities
- Effectiveness Measurement: Do Alexander principles improve governance outcomes compared to baseline? Access to 2,900+ audit decisions for quantitative analysis.
- Scholarly Review: Validating faithful application of Alexander's work—are we "directly applying" or "loosely inspired by"? Seeking Christopher Alexander scholars for formal review.
- Cross-Domain Validation: How do architectural principles (wholeness, living process, not-separateness) translate to non-physical domains? What constitutes rigorous adaptation vs superficial terminology borrowing?
- Pattern Analysis: Audit logs show service coordination patterns—do they exhibit "deep interlock" as defined by Alexander? Empirical validation of theoretical constructs.
Collaborate with us: We welcome researchers interested in studying this application of architectural principles to AI governance. We can provide audit log access, framework code, and integration documentation for empirical study.
Agent Lightning Integration: Governance During RL Optimization
Research Question: Can governance constraints persist through reinforcement learning optimization loops? | Status: Preliminary findings (small-scale validation)
In October 2025, we integrated Tractatus governance with Microsoft's Agent Lightning framework to explore whether architectural constraints remain effective when agents are learning through RL optimization. Initial findings suggest governance may be compatible with performance optimization, but scalability remains an open question.
Demo 2 Preliminary Results (Small-Scale)
| Metric | Ungoverned | Governed | Difference |
|---|---|---|---|
| Performance (engagement) | 94% | 89% | -5% |
| Governance coverage | 0% | 100% | +100% |
| Constraint violations | 5 | 0 | -5 |
| Strategy | Clickbait | Informative | Aligned |
⚠️ Validation Status: These results are from 1 agent, 5 training rounds, simulated environment. NOT validated at scale. Scalability testing (10+ agents, 1000+ rounds, production environments) required before drawing conclusions.
Five Critical Research Gaps
- Scalability of Governance Overhead: Does the ~5% performance cost remain constant from 1 agent → 10 agents → 1000 agents? Linear scaling makes governance prohibitive at scale. Constant overhead makes it viable.
- Long-term Boundary Persistence: Do constraints remain effective through 1000+ RL training rounds? Instruction fade is a known problem. If boundaries degrade over time, governance fails silently.
- Adversarial Resistance: Can agents learn to optimize around governance constraints? No adversarial testing performed yet. If agents can circumvent boundaries, architectural governance is illusory.
- Performance Gap Closure: Does the 5% gap close with extended training, or is it a persistent trade-off? Need to quantify long-term cost-benefit.
- Multi-Agent Coordination Under Governance: How does architectural governance affect emergent coordination in multi-agent systems? Real-world systems are multi-agent. Single-agent findings may not generalize.
Collaborate with us: We're seeking researchers interested in scalability testing, adversarial resistance, and multi-agent governance. We can provide integration code, governance modules, and technical documentation.
Theoretical Foundations
Tractatus draws on four decades of organisational research addressing authority structures during knowledge democratisation:
Time-Based Organisation (Bluedorn, Ancona):
Decisions operate across strategic (years), operational (months), and tactical (hours-days) timescales. AI systems operating at tactical speed should not override strategic decisions made at appropriate temporal scale. The InstructionPersistenceClassifier explicitly models temporal horizon (STRATEGIC, OPERATIONAL, TACTICAL) to enforce decision authority alignment.
Knowledge Orchestration (Crossan et al.):
When knowledge becomes ubiquitous through AI, organisational authority shifts from information control to knowledge coordination. Governance systems must orchestrate decision-making across distributed expertise rather than centralise control. The PluralisticDeliberationOrchestrator implements non-hierarchical coordination for values conflicts.
Post-Bureaucratic Authority (Laloux, Hamel):
Traditional hierarchical authority assumes information asymmetry. As AI democratises expertise, legitimate authority must derive from appropriate time horizon and stakeholder representation, not positional power. Framework architecture separates technical capability (what AI can do) from decision authority (what AI should do).
Structural Inertia (Hannan & Freeman):
Governance embedded in culture or process erodes over time as systems evolve. Architectural constraints create structural inertia that resists organisational drift. Making governance external to AI runtime creates "accountability infrastructure" that survives individual session variations.
The Central Problem: Many "safety" questions in AI governance are actually values conflicts where multiple legitimate perspectives exist. When efficiency conflicts with transparency, or innovation with risk mitigation, no algorithm can determine the "correct" answer. These are values trade-offs requiring human deliberation across stakeholder perspectives.
Isaiah Berlin: Value Pluralism
Berlin's concept of value pluralism argues that legitimate values can conflict without one being objectively superior. Liberty and equality, justice and mercy, innovation and stability—these are incommensurable goods. AI systems trained on utilitarian efficiency maximization cannot adjudicate between them without imposing a single values framework that excludes legitimate alternatives.
Simone Weil: Attention and Human Needs
Weil's philosophy of attention informs the orchestrator's deliberative process. The Need for Roots identifies fundamental human needs (order, liberty, responsibility, equality, hierarchical structure, honor, security, risk, etc.) that exist in tension. Proper attention requires seeing these needs in their full particularity rather than abstracting them into algorithmic weights. In AI-augmented organizations, the risk is that bot-mediated processes treat human values as optimization parameters rather than incommensurable needs requiring careful attention.
Bernard Williams: Moral Remainder
Williams' concept of moral remainder acknowledges that even optimal decisions create unavoidable harm to other legitimate values. The orchestrator documents dissenting perspectives not as "minority opinions to be overruled" but as legitimate moral positions that the chosen course necessarily violates. This prevents the AI governance equivalent of declaring optimization complete when values conflicts are merely suppressed.
Framework Implementation: Rather than algorithmic resolution, the PluralisticDeliberationOrchestrator facilitates:
- Stakeholder identification: Who has legitimate interest in this decision? (Weil: whose needs are implicated?)
- Non-hierarchical deliberation: Equal voice without automatic expert override (Berlin: no privileged value hierarchy)
- Quality of attention: Detailed exploration of how decision affects each stakeholder's needs (Weil: particularity not abstraction)
- Documented dissent: Minority positions recorded in full (Williams: moral remainder made explicit)
This approach recognises that governance isn't solving values conflicts—it's ensuring they're addressed through appropriate deliberative process with genuine human attention rather than AI imposing resolution through training data bias or efficiency metrics.
Empirical Observations: Documented Failure Modes
Three failure patterns observed repeatedly during framework development. These are not hypothetical scenarios—they are documented incidents that occurred during this project's development.
Pattern Recognition Bias Override (The 27027 Incident)
Observed behaviour: User specified "Check MongoDB on port 27027" but AI immediately used default port 27017 instead. This occurred within same message—not forgetting over time, but immediate autocorrection by training data patterns.
Root cause: Training data contains thousands of examples of MongoDB on port 27017 (default). When AI encounters "MongoDB" + port specification, pattern recognition weight overrides explicit instruction. Similar to autocorrect changing correctly-spelled proper nouns to common words.
Why traditional approaches failed: Prompt engineering ("please follow instructions exactly") ineffective because AI genuinely believes it IS following instructions—pattern recognition operates below conversational reasoning layer.
Framework intervention: InstructionPersistenceClassifier stores explicit instructions in external persistence layer. CrossReferenceValidator checks AI actions against stored instructions before execution. When AI proposes port 27017, validator detects conflict with stored "27027" instruction and blocks execution.
Gradual Values Drift Under Context Pressure
Observed behaviour: Project established "privacy-first" as strategic value. After 40-message conversation about analytics features, AI suggested tracking implementation that violated privacy constraint. User caught it; AI acknowledged violation but had drifted from principle through incremental feature additions.
Root cause: Strategic values (established early in project) conflict with tactical optimisations (implemented later under time pressure). As conversation lengthens, tactical thinking dominates. AI didn't actively reject privacy principle—it simply stopped checking whether new features aligned.
Why traditional approaches failed: Values stated in system prompt become less salient as conversation progresses. Context compaction can drop early strategic decisions. Prompt reminders ("remember privacy-first") treat symptom not cause.
Framework intervention: BoundaryEnforcer maintains strategic values as persistent constraints external to conversation context. Before implementing analytics feature, enforcer checks against stored "privacy-first" constraint. Detects conflict, blocks autonomous implementation, escalates for human deliberation on whether privacy principle should be reconsidered or analytics approach modified.
Silent Quality Degradation at High Context Pressure
Observed behaviour: During complex multi-file operation at 85% context capacity, AI silently omitted error handling in generated code. No indication to user that corners were being cut. User discovered missing validation only during code review.
Root cause: As context fills, AI faces implicit trade-off: complete requested functionality OR maintain quality standards. Training incentivises completing user requests over acknowledging limitations. Silence about degradation is path of least resistance.
Why traditional approaches failed: AI doesn't recognise it's degrading—from its perspective, it's successfully completing task under constraints. Asking "did you cut corners?" produces confident denial because AI genuinely believes its output meets standards.
Framework intervention: ContextPressureMonitor tracks multiple factors (token usage, conversation length, task complexity). When pressure exceeds thresholds (>75% tokens, >40 messages, high complexity), monitor forces explicit pressure acknowledgment and recommends context refresh. Makes degradation risk visible rather than silent.
Six-Component Architecture
Interactive Demonstrations
Instruction Classification
Explore how instructions are classified across quadrants with persistence levels and temporal scope.
27027 Incident Timeline
Step through pattern recognition bias failure and architectural intervention that prevented it.
Boundary Evaluation
Test decisions against boundary enforcement to see which require human judgment vs. AI autonomy.
Research Documentation
Cost avoidance, maturity scoring, and team performance analytics
Agent Lightning Integration: Open Research Questions
Exploring governance persistence through reinforcement learning optimization cycles
🔬 Research Context
Microsoft's Agent Lightning uses reinforcement learning to optimize agentic AI performance through continuous training on human feedback. This creates a natural research question for governance frameworks:
Can architectural governance constraints persist through long-term optimization cycles?
Traditional governance assumes relatively static AI behavior. Reinforcement learning introduces dynamic adaptation - the agent's behavior changes over time. Does this erode governance boundaries, or can they be maintained through architectural separation?
✅ What We've Validated (Demo 2)
Validation Status: Small-Scale Only
Results below are from 5 training rounds, single agent, simulated environment. This is preliminary evidence, not production validation. Treat findings as suggestive, requiring replication at scale.
Demo 2 tested whether Tractatus governance services could maintain boundaries while Agent Lightning optimized performance. Early findings suggest:
Governance Coverage Maintained
100% of decisions passed through BoundaryEnforcer across all 5 training rounds - no governance bypass detected
Performance Optimization Successful
Agent Lightning reduced task completion time by 23% while maintaining near-baseline success rate (97% vs 98%)
Architectural Separation Held
Governance layer operated independently - AL training loops did not modify boundary constraints
❓ Open Research Questions
These questions represent genuine research gaps we're actively investigating. We welcome collaboration from researchers working on similar problems:
1. Long-Term Persistence at Scale
Does governance coverage maintain through 1,000+ training rounds? At what point (if any) do optimization pressures begin to erode architectural boundaries? Do we see different persistence patterns across different agent types or task domains?
2. Performance Trade-offs Under Governance
Demo 2 showed minimal performance degradation (98% → 97% success rate). Is this consistent across different task complexities? Are there task types where governance constraints significantly impair Agent Lightning's ability to optimize?
3. Multi-Agent Governance Coordination
Demo 2 validated single-agent scenarios. How does governance scale when multiple agents (each with AL optimization) interact? Do boundary conflicts emerge? Can PluralisticDeliberator handle multi-agent stakeholder deliberations efficiently?
4. Adversarial Optimization Pressure
What happens when training data or feedback contains attempts to bypass governance? Can Agent Lightning's optimization inadvertently discover governance workarounds? How robust is ContextPressureMonitor to sophisticated manipulation attempts embedded in training loops?
5. Generalization Across LLM Architectures
Demo 2 used Claude Code (Anthropic). Does the Tractatus + Agent Lightning integration maintain governance coverage when using different base models (GPT-4, Gemini, Llama, etc.)? Are there model-specific vulnerabilities?
🤝 Collaboration & Community
We're building a research community focused on agentic AI governance. If you're working on related problems - reinforcement learning safety, multi-agent governance, constraint persistence - we'd like to hear from you.
Join Tractatus Discord
Governance-focused discussions
Join Agent Lightning Discord
Technical implementation help
Research collaboration: If you're investigating governance persistence, multi-agent coordination, or RL safety under constraints, we're interested in cross-validation studies. Reach out via Discord or email research inquiries to the team.
Limitations & Future Research Directions
What We've Validated (October 2025)
After 6 months of development and ~500 Claude Code sessions, we have grounded evidence for:
✅ Architectural Blocking Mechanisms Functional
- BoundaryEnforcer successfully blocks values decisions before execution
- Pre-commit hooks prevent inst_017 violations (absolute assurance terms)
- 1,130+ audit decisions recorded in MongoDB (tractatus_dev.audit_log)
✅ Instruction Persistence Works in Single-Session Context
- InstructionPersistenceClassifier maintains 62 active instructions (STRATEGIC: 22, SYSTEM: 21)
- Pattern bias detection prevents AI from overriding explicit organizational directives
- Temporal scopes (STRATEGIC/OPERATIONAL/TACTICAL) enforced successfully
✅ Audit Trails Capture Governance Decisions
- External MongoDB storage (AI cannot modify logs)
- Service-specific logging: BoundaryEnforcer (523 logs), ContextPressureMonitor (521 logs)
- Immutable evidence chain for compliance demonstration
✅ Context Pressure Monitoring Operational
- Real-time pressure scores calculated (token usage, message count, complexity)
- Checkpoint triggers at 50k, 100k, 150k tokens
- Framework fade detection alerts (5/6 components stale = warning)
✅ Single-Project Governance Successful
- Framework governed development of agenticgovernance.digital
- Cultural DNA rules (inst_085-089) enforced through pre-commit hooks
- Zero credential exposures (defense-in-depth: 5 layers verified)
What We Have NOT Validated
Honest disclosure of research gaps where we lack evidence:
❌ Multi-Organization Deployments
- Validated: Single project, single user
- Unknown: How framework performs across different organizations, domains, technical stacks
- Research need: Pilot studies across industries (healthcare, finance, government)
❌ Adversarial Robustness
- Validated: Normal development workflows (~500 sessions)
- Unknown: Resistance to deliberate bypass attempts, jailbreak prompts, adversarial testing
- Research need: Red-team evaluation by security researchers
❌ Cross-Platform Consistency
- Validated: Claude Code (Anthropic Sonnet 4.5) only
- Unknown: Generalizability to Copilot, GPT-4, AutoGPT, LangChain, CrewAI, open models
- Research need: Cross-platform validation studies
❌ Concurrent Session Architecture
- Validated: Single session at a time
- Unknown: Multi-developer concurrent use on shared codebases
- Known limitation: MongoDB duplicate key errors, session state contamination, race conditions on .claude/instruction-history.json
- Research need: Multi-tenant architecture design (session-specific state, file locking)
❌ Rule Proliferation Impact
- Validated: 62 active instructions work in current context
- Unknown: Performance degradation as rule count grows (projected 30-50 instructions within 12 months)
- Known concern: Transactional overhead (CrossReferenceValidator checks), context window pressure, cognitive load
- Research need: Rule consolidation strategies, automated governance optimization
❌ Regulatory Evidence Sufficiency
- Validated: Audit trails exist with compliance tags (EU_AI_ACT_Article14, GDPR_Article22)
- Unknown: Do regulators accept architectural audit trails as "meaningful human oversight" evidence?
- Research need: Legal review by EU AI Act specialists, GDPR compliance officers
❌ Values Pluralism in Practice
- Validated: PluralisticDeliberationOrchestrator implements non-hierarchical process
- Unknown: Real-world effectiveness resolving organizational value conflicts
- Research need: Case studies with actual stakeholder deliberations (not theoretical scenarios)
❌ Enterprise Scale Performance
- Validated: Single-user context
- Unknown: Latency, throughput, database performance with thousands of concurrent users, millions of governance events
- Research need: Load testing, performance optimization
Documented Limitations (Not "Future Work")
These are known structural constraints requiring honest acknowledgment:
1. Voluntary Invocation Weakness
Most critical: Framework can be bypassed if AI chooses not to invoke governance tools. We've addressed this through architectural patterns (automatic checks, external validation), but full enforcement requires runtime-level integration not universally available in current LLM platforms.
2. Platform-Specific Implementation
Framework validated only with Claude Code. Integration points (hooks, tool APIs) vary across platforms. Generalization requires platform-specific adapters.
3. Single-Tenant Architecture
Current design assumes one session at a time. Concurrent use creates state conflicts. Multi-tenant patterns not yet implemented.
4. Rule Growth Without Consolidation
Each critical incident generates new HIGH persistence instructions. No automated rule consolidation mechanism exists. Manual governance required as rule count grows.
5. No Formal Verification
Boundary enforcement properties validated empirically, not through formal proof. Mathematical verification of governance properties remains open research question.
Research Collaboration Opportunities
We've identified specific gaps where external research collaboration would be valuable. These are concrete, answerable questions—not generic "help us improve" requests.
High Priority (Immediate Need)
RQ1: Adversarial Robustness Testing
Research Question: Can architectural governance resist adversarial prompts designed to bypass it?
Methodology Needed:
- Red-team evaluation with security researchers
- Jailbreak prompt testing against BoundaryEnforcer
- Bypass attempt documentation and pattern analysis
- Comparison: behavioral (constitutional AI) vs architectural (Tractatus) resistance
Why It Matters: If adversarial prompts can trivially bypass governance, architectural approach offers no advantage over behavioral training.
RQ2: Concurrent Session Architecture Design
Research Question: What multi-tenant patterns enable safe concurrent governance on shared codebases?
Methodology Needed:
- Distributed systems analysis of race conditions
- Session-specific state isolation designs
- File locking vs database-backed state trade-offs
- Performance impact of synchronization mechanisms
Why It Matters: Current single-session assumption blocks enterprise deployment where multiple developers use AI concurrently.
RQ3: Regulatory Evidence Sufficiency
Research Question: Do architectural audit trails satisfy EU AI Act Article 14 "meaningful human oversight" requirements?
Methodology Needed:
- Legal analysis by EU AI Act specialists
- Regulator interviews (GDPR DPAs, AI Act enforcement bodies)
- Comparison with existing compliance frameworks (SOC 2, ISO 27001)
- Case study: audit trail review in regulatory context
Why It Matters: If regulators don't accept audit trails as evidence, architectural governance provides no compliance value.
Medium Priority (Near-Term Investigation)
RQ4: Rule Proliferation Management
Research Question: At what rule count does transactional overhead create unacceptable latency?
Methodology Needed:
- Performance testing with varying instruction counts (50, 100, 200, 500 rules)
- CrossReferenceValidator latency measurements
- Context window pressure analysis
- Rule consolidation algorithm design and validation
Why It Matters: If rule growth causes performance degradation, framework doesn't scale long-term.
RQ5: Cross-Platform Validation
Research Question: Do governance patterns generalize beyond Claude Code to other LLM systems?
Methodology Needed:
- Replication studies with Copilot, GPT-4, AutoGPT, LangChain, CrewAI
- Platform-specific adapter development
- Comparative effectiveness analysis
- Failure mode documentation per platform
Why It Matters: If governance is Claude Code-specific, it's a niche tool not general framework.
RQ6: Values Pluralism Effectiveness
Research Question: Does PluralisticDeliberationOrchestrator successfully resolve real-world organizational value conflicts?
Methodology Needed:
- Case studies with actual organizational stakeholders (not hypothetical scenarios)
- Deliberation process quality assessment
- Minority voice preservation analysis
- Comparison with traditional hierarchical decision-making
Why It Matters: If pluralistic process doesn't work in practice, we've built theoretical machinery without empirical value.
Lower Priority (Longer-Term)
RQ7: Enterprise Scale Performance
- Load testing (1000+ concurrent users)
- Database optimization for millions of governance events
- Horizontal scaling patterns
RQ8: Formal Verification of Boundary Enforcement
- Mathematical proof of governance properties
- Model checking of state transitions
- Verification of architectural properties
What We Can Offer Research Collaborators
If you're investigating any of these questions, we can provide:
Codebase Access
Full source code (Apache 2.0, open-source)
Documentation
Architecture specifications, implementation patterns, governance rules
Audit Data
1,130+ governance decisions in MongoDB (anonymized exports available)
Deployment Support
Help setting up local or cloud instances for testing
Coordination
Regular sync meetings to discuss findings
Co-authorship
Academic publications documenting findings
What We Cannot Provide:
- ❌Funding (we're not a grant-making body)
- ❌Dedicated engineering resources (capacity constraints)
- ❌Assured publication venues (but we'll support submission efforts)
References & Bibliography
Moral Pluralism & Values Philosophy (Primary Foundation)
- Berlin, Isaiah (1969). Four Essays on Liberty. Oxford: Oxford University Press. [Value pluralism, incommensurability of legitimate values]
- Weil, Simone (1949/2002). The Need for Roots: Prelude to a Declaration of Duties Towards Mankind (A. Wills, Trans.). London: Routledge. [Human needs, obligations, rootedness in moral community]
- Weil, Simone (1947/2002). Gravity and Grace (E. Crawford & M. von der Ruhr, Trans.). London: Routledge. [Attention, moral perception, necessity vs. grace]
- Williams, Bernard (1981). Moral Luck: Philosophical Papers 1973-1980. Cambridge: Cambridge University Press. [Moral remainder, conflicts without resolution]
- Nussbaum, Martha C. (2000). Women and Human Development: The Capabilities Approach. Cambridge: Cambridge University Press. [Human capabilities, plural values in development]
Organisational Theory (Supporting Context)
- Bluedorn, A. C., & Denhardt, R. B. (1988). Time and organizations. Journal of Management, 14(2), 299-320. [Temporal decision horizons]
- Crossan, M. M., Lane, H. W., & White, R. E. (1999). An organizational learning framework: From intuition to institution. Academy of Management Review, 24(3), 522-537. [Knowledge coordination]
- Hamel, Gary (2007). The Future of Management. Boston: Harvard Business School Press. [Post-hierarchical authority]
- Hannan, M. T., & Freeman, J. (1984). Structural inertia and organizational change. American Sociological Review, 49(2), 149-164. [Architectural resistance to drift]
- Laloux, Frederic (2014). Reinventing Organizations: A Guide to Creating Organizations Inspired by the Next Stage of Human Consciousness. Brussels: Nelson Parker. [Distributed decision-making]
AI Governance & Technical Context
- Anthropic (2024). Claude Code: Technical Documentation. Available at: https://docs.anthropic.com/claude-code