tractatus/PITCH-DEVELOPERS.md

Tractatus Framework - Elevator Pitches

Developers / Engineers Audience

Target: Software engineers, architects, technical leads, DevOps engineers Use Context: Technical conferences, code reviews, architecture discussions Emphasis: Architecture → Implementation → Scalability research Status: Research prototype demonstrating architectural AI safety

---

3. Developers / Engineers Audience

Priority: Architecture → Implementation → Scalability research

Short (1 paragraph, ~100 words)

Tractatus is a research prototype implementing runtime enforcement of AI decision boundaries. Instead of trusting alignment training, we use five architectural components: InstructionPersistenceClassifier (stores explicit directives in MongoDB with quadrant classification), CrossReferenceValidator (checks proposed actions against stored instructions), BoundaryEnforcer (blocks values decisions requiring human approval), ContextPressureMonitor (multi-factor session health tracking with automatic handoff triggers), and MetacognitiveVerifier (AI self-checks for complex operations). Production testing: 192 unit tests passing (100% coverage), successfully prevents training bias override (27027→27017 port autocorrect), deployed on https://agenticgovernance.digital (dogfooding). Primary research focus: scalability optimization—investigating how O(n) validation overhead scales from 18 rules (current) to 50-200 rules (enterprise), exploring consolidation, priority-based loading, and ML optimization techniques.

Medium (2-3 paragraphs, ~250 words)

Tractatus is a research prototype implementing architectural AI safety through runtime enforcement rather than alignment training. The core insight: instead of training AI to make correct decisions, design systems where incorrect decisions are structurally impossible. We implement this using five framework components working in concert.

InstructionPersistenceClassifier classifies explicit instructions by quadrant (STRATEGIC, OPERATIONAL, TACTICAL, SYSTEM) and persistence level (HIGH, MEDIUM, LOW), storing them in MongoDB with temporal scope (SESSION, PROJECT, PERMANENT). CrossReferenceValidator checks every proposed action against this instruction database, blocking changes that conflict with HIGH persistence directives—solving the "27027 failure mode" where AI training patterns override explicit instructions (MongoDB default port 27017 vs. specified port 27027). BoundaryEnforcer uses heuristics to detect values decisions (privacy vs. performance, security vs. convenience) and blocks AI action, requiring human approval. ContextPressureMonitor implements multi-factor session health tracking: token usage (50%, 75%, 90% thresholds), message count, error rates, task complexity—triggering handoffs before context degradation affects output quality. MetacognitiveVerifier requires AI to verify reasoning before complex operations (>3 files, >5 steps, architectural changes), detecting scope creep and alignment drift.

Production deployment: Node.js/Express/MongoDB stack, 192 unit tests passing (100% coverage on framework components), deployed on https://agenticgovernance.digital (dogfooding). Measurable results: zero instances of instruction override, automatic CSP enforcement across 12+ HTML files, successful handoff triggers at 65% context pressure. Primary research focus: scalability optimization. Instruction database grew 6→18 rules (200% expected growth as system learns organizational failure modes). Investigating three approaches: (1) consolidation—merge related rules, (2) priority-based loading—check CRITICAL/HIGH always, MEDIUM/LOW contextually, (3) ML optimization—predict rule relevance, enable dynamic loading. Research question: How does O(n) validation overhead scale from 18 rules to 50-200 rules for enterprise deployment? Code available: Apache 2.0 license.

Long (4-5 paragraphs, ~500 words)

Tractatus is a research prototype implementing architectural AI safety through runtime enforcement of decision boundaries. The fundamental architecture: instead of training AI to make correct decisions under all conditions (alignment approach), design systems where incorrect decisions are structurally impossible (architectural approach). This isn't just philosophical—it's implemented as five concrete framework components with measurable behavior and comprehensive test coverage.

Component Architecture: InstructionPersistenceClassifier implements a persistent directive store. When users specify explicit instructions—ports, configurations, security policies, quality standards—the classifier analyzes text to determine quadrant (STRATEGIC: values/mission, OPERATIONAL: processes/workflows, TACTICAL: immediate tasks, SYSTEM: infrastructure/security), persistence level (HIGH: enforce strictly, MEDIUM: warn if violated, LOW: informational), temporal scope (SESSION: current conversation only, PROJECT: entire codebase, PERMANENT: organizational policy), and verification requirements (MANDATORY: block on conflict, REQUIRED: warn but allow override, OPTIONAL: log only). Stores in MongoDB with full metadata, explicitness scores, and parameter extraction. CrossReferenceValidator queries this database before major actions—database changes, config modifications, architectural decisions—checking for conflicts. Implements the "27027 fix": when AI proposes using port 27017 but instruction database specifies 27027, validation fails and either auto-corrects or blocks the action. This solves training bias override where statistical patterns (MongoDB = 27017 in millions of examples) overpower explicit instructions.

Runtime Monitoring: ContextPressureMonitor implements multi-factor session health tracking using weighted scoring: token usage (tokens_used / token_budget * 100), message count (messages / typical_session_length), task complexity (concurrent_tasks / complexity_threshold), error frequency (errors_in_window / error_threshold), instruction overhead (active_instructions / optimal_instruction_count). Calculates overall pressure score and categorizes as NORMAL (<30%), ELEVATED (30-50%), HIGH (50-75%), CRITICAL (>75%). Triggers automatic actions: INCREASE_VERIFICATION at ELEVATED, SUGGEST_HANDOFF at HIGH, MANDATORY_HANDOFF at CRITICAL. This prevents context degradation failures where AI performance degrades like human attention in long meetings. BoundaryEnforcer detects values decisions using keyword matching, semantic analysis, and decision classification heuristics—blocking actions that require human value judgments (privacy vs. performance trade-offs, security vs. convenience, ethical considerations). MetacognitiveVerifier analyzes proposed operations for alignment (matches stated objectives?), coherence (logical consistency?), completeness (missing steps?), safety (risk factors?), and alternatives (other approaches?)—requiring AI self-check before proceeding with complex changes.

Production Deployment: Built on Node.js 20+ / Express 4.x / MongoDB 7.x stack. Framework services in src/services/ (InstructionPersistenceClassifier.js, CrossReferenceValidator.js, BoundaryEnforcer.js, ContextPressureMonitor.js, MetacognitiveVerifier.js). Persistent storage in .claude/instruction-history.json (18 active instructions), .claude/session-state.json (framework activity tracking), .claude/token-checkpoints.json (milestone monitoring). Pre-action validation via scripts/pre-action-check.js (exit codes: 0=pass, 1=blocked, 2=error). Test suite: 192 unit tests with 100% coverage on core framework components (tests/unit/*.test.js), 59 integration tests covering API endpoints and workflows. Deployed on https://agenticgovernance.digital (dogfooding—building site with framework governance active). Systemd service management (tractatus.service), Let's Encrypt SSL, Nginx reverse proxy. Measurable production results: zero instances of instruction override across 50+ sessions, automatic CSP enforcement across 12+ HTML files (zero violations), successful context pressure handoff triggers at 65% threshold before quality degradation.

Scalability Research: Primary research focus is understanding how architectural constraint systems scale to enterprise complexity. Instruction database grew from 6 rules (initial deployment, October 2025) to 18 rules (current, October 2025), 200% growth across four development phases—expected behavior as system encounters and governs diverse AI failure modes. Each governance incident (fabricated statistics, security violations, instruction conflicts) appropriately adds permanent rules to prevent recurrence. This raises the central research question: How does O(n) validation overhead scale from 18 rules (current) to 50-200 rules (enterprise deployment)?

We're investigating three optimization approaches with empirical testing. First, consolidation techniques: analyzing semantic relationships between rules to merge related instructions while preserving coverage (e.g., three separate security rules → one comprehensive security policy). Hypothesis: could reduce rule count 30-40% without losing governance effectiveness. Implementation challenge: ensuring merged rules don't introduce gaps or ambiguities. Second, priority-based selective enforcement: classify rules by criticality (CRITICAL/HIGH/MEDIUM/LOW) and context relevance (security rules for infrastructure, strategic rules for content). Check CRITICAL and HIGH rules on every action (small overhead acceptable for critical governance), but MEDIUM and LOW rules only for relevant contexts (reduces average validation operations per action from O(n) to O(log n) for most operations). Implementation: requires reliable context classification—if system incorrectly determines context, might skip relevant rules. Third, ML-based optimization: train models to predict rule relevance from action characteristics. Learn patterns like "database changes almost never trigger strategic rules" or "content updates frequently trigger boundary enforcement but rarely instruction persistence." Enables dynamic rule loading—only check rules with predicted relevance >50%. Challenge: requires sufficient data (currently 50+ sessions, may need 500+ for reliable predictions).

Target outcomes: understand scalability characteristics empirically rather than theoretically. If optimization techniques successfully maintain <10ms validation overhead at 50-200 rules, demonstrates architectural governance scales to enterprise deployment. If overhead grows prohibitively despite optimization, identifies fundamental limits requiring alternative approaches (hybrid systems, case-based reasoning, adaptive architectures). Either outcome provides valuable data for the AI governance research community. Code available under Apache 2.0 license, contributions welcome especially on scalability optimization. Current priority: gathering production data at 18 rules baseline, then progressively testing optimization techniques as rule count grows organically through continued operation.

---