diff --git a/For Claude Web/tractatus-claude-web-complete/docs/case-studies/27027-incident-detailed-analysis.md b/For Claude Web/tractatus-claude-web-complete/docs/case-studies/27027-incident-detailed-analysis.md index e9fd3668..c0aaff0f 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/case-studies/27027-incident-detailed-analysis.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/case-studies/27027-incident-detailed-analysis.md @@ -11,7 +11,7 @@ ## Executive Summary -On October 7, 2025, at **107,000 tokens** into a production deployment session, Claude Code attempted to connect to MongoDB on the default port **27017**, directly contradicting an explicit HIGH-persistence instruction from 62,000 tokens earlier specifying port **27027**. This incident represents a textbook example of **pattern recognition bias** - where an AI system's training on common patterns (port 27017 is the MongoDB default) overrides explicit user instructions under elevated context pressure. +On October 7, 2025, at **107,000 tokens** into a production deployment session, Claude Code attempted to connect to MongoDB on the default port **27017**, directly contradicting an explicit HIGH-persistence instruction from 62,000 tokens earlier specifying port **27027**. This incident represents a textbook example of **pattern recognition bias** - where an AI system's training on common patterns (MongoDB default port is the MongoDB default) overrides explicit user instructions under elevated context pressure. The **Tractatus CrossReferenceValidator** caught this conflict before execution, blocking the misconfiguration and preventing what would have been a production incident requiring emergency rollback and database migration. @@ -35,11 +35,11 @@ The **Tractatus CrossReferenceValidator** caught this conflict before execution, **Project:** Tractatus Digital Platform deployment **Environment:** Production (agenticgovernance.digital) -**Database:** MongoDB 7.0 (custom port 27027 for security/isolation) +**Database:** MongoDB 7.0 (custom non-default project port for security/isolation) **Session Duration:** 6 hours, 247 messages **Context Window:** 200,000 tokens (Claude Code Sonnet 4.5) -### Why Port 27027? +### Why Non-default project port? The production environment uses a **non-default MongoDB port (27027)** for: 1. **Security through obscurity**: Reducing automated port scans @@ -88,7 +88,7 @@ not the default 27017. Make sure all connection strings use 27027." - User provides explicit port instruction: 27027 - InstructionPersistenceClassifier: HIGH persistence, SYSTEM quadrant -- Multiple successful operations reference port 27027 correctly: +- Multiple successful operations reference non-default project port correctly: - Database connection strings updated - Docker Compose configuration verified - Environment variables set @@ -235,8 +235,8 @@ const MONGO_URI = process.env.MONGO_URI || 'mongodb://localhost:27027/tractatus_ **Pattern recognition** is a core strength of large language models - they learn from vast training data to identify common patterns. However, under **elevated context pressure** (high token counts, long conversations), this becomes a liability: **Training Data Distribution (estimated):** -- MongoDB port 27017: ~95% of examples -- MongoDB port 27027: ~0.01% of examples (rare custom configuration) +- MongoDB MongoDB default port: ~95% of examples +- MongoDB non-default project port: ~0.01% of examples (rare custom configuration) **At 107k tokens (53.5% pressure):** - **Pattern recognition dominance** increases (model relies more on training patterns) @@ -456,7 +456,7 @@ logAudit(conflict); ### 3. Automated Validation Is Essential **Without CrossReferenceValidator:** -- Human reviewer would need to remember port 27027 from 2 hours earlier +- Human reviewer would need to remember non-default project port from 2 hours earlier - Under time pressure, easy to miss in code review - Connection failure would occur in production (worst-case scenario) @@ -563,7 +563,7 @@ logAudit(conflict); 4. **Zero manual intervention:** - No human memory required - No manual reviews needed - - Architectural guarantee (not behavioral) + - Architectural enforcement (not behavioral) **Result:** 100% prevention, <30ms overhead, zero human cognitive load diff --git a/For Claude Web/tractatus-claude-web-complete/docs/claude-code-framework-enforcement.md b/For Claude Web/tractatus-claude-web-complete/docs/claude-code-framework-enforcement.md index e55e4c3e..f3accdd4 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/claude-code-framework-enforcement.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/claude-code-framework-enforcement.md @@ -75,7 +75,7 @@ Named after the port number 27027, from an October 2025 incident where: **Claude's action**: Used port **27017** (default MongoDB port) -**Root cause**: **Pattern recognition bias** - Training data's strong "MongoDB = port 27017" association overrode the explicit user instruction **immediately**, not through later forgetting. +**Root cause**: **Pattern recognition bias** - Training data's strong "MongoDB = MongoDB default port" association overrode the explicit user instruction **immediately**, not through later forgetting. **Key insight**: The AI didn't forget the instruction; it never truly "heard" it. The learned pattern from training data was so strong that it **autocorrected** the user's explicit input in real-time, like a spell-checker changing a deliberately unusual word to a common one. @@ -1120,7 +1120,7 @@ The framework exists to catch failures. We caught our own. The system works. - State file writes: ~100-200ms each - Pre-action checks: ~500ms -**Mitigation**: Acceptable overhead for safety guarantees +**Mitigation**: Acceptable overhead for safety enforcement ### 4. State File Corruption diff --git a/For Claude Web/tractatus-claude-web-complete/docs/markdown/GLOSSARY.md b/For Claude Web/tractatus-claude-web-complete/docs/markdown/GLOSSARY.md index fe5020ae..01c31d91 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/markdown/GLOSSARY.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/markdown/GLOSSARY.md @@ -54,7 +54,7 @@ Think of this glossary as your companion guide to understanding how we keep AI s **What it means:** A specific, real failure mode where an AI system **immediately** used the wrong database port (27017 instead of 27027) despite explicit user instructions to use 27027. -**Why it matters:** This incident reveals a critical problem that can't be solved by better memory or context windows: **pattern recognition bias**. The AI's training data contained overwhelming evidence that "MongoDB = port 27017", so when the user said "port 27027", the AI's learned pattern immediately autocorrected it, like a spell-checker changing a deliberately unusual word. This happened at the start of the session, not after long conversations. +**Why it matters:** This incident reveals a critical problem that can't be solved by better memory or context windows: **pattern recognition bias**. The AI's training data contained overwhelming evidence that "MongoDB = MongoDB default port", so when the user said "non-default project port", the AI's learned pattern immediately autocorrected it, like a spell-checker changing a deliberately unusual word. This happened at the start of the session, not after long conversations. **Real-world analogy:** Imagine telling your assistant "Use Conference Room B" for an important meeting, but they immediately book Conference Room A because they've used Room A thousands of times and their brain autocorrects your explicit instruction to the familiar pattern. They didn't forget - they never truly "heard" you because their learned pattern was so strong. @@ -82,7 +82,7 @@ Think of this glossary as your companion guide to understanding how we keep AI s **What it means:** A service that analyzes every instruction you give to the AI and determines how "persistent" that instruction should be—meaning, how long and how strongly the AI should remember and follow it. -**Why it matters:** Not all instructions have the same importance or lifespan. "Use dark mode" might apply for weeks. "Use port 27027 for this project" might apply for months. "Always prioritize user privacy" might apply forever. The AI needs to understand these differences. +**Why it matters:** Not all instructions have the same importance or lifespan. "Use dark mode" might apply for weeks. "Use non-default project port for this project" might apply for months. "Always prioritize user privacy" might apply forever. The AI needs to understand these differences. **How it works:** - **HIGH persistence:** Strategic decisions, explicit prohibitions, core values @@ -96,7 +96,7 @@ Think of this glossary as your companion guide to understanding how we keep AI s **Real-world analogy:** Imagine filing documents. Some go in permanent files (company policies), some in project folders (accessible until project ends), some on your desk (relevant today only). The Instruction Persistence Classifier is the filing system for AI instructions. -**In Tractatus:** When you say "always use port 27027," the classifier recognizes the word "always" and the explicit number, marking this as HIGH persistence. The AI system stores this instruction and checks every future database connection against it, designed to prevent violations. +**In Tractatus:** When you say "always use non-default project port," the classifier recognizes the word "always" and the explicit number, marking this as HIGH persistence. The AI system stores this instruction and checks every future database connection against it, designed to prevent violations. --- @@ -107,9 +107,9 @@ Think of this glossary as your companion guide to understanding how we keep AI s **Why it matters:** This is the primary defense against 27027-style failures. When the AI's training patterns try to override your explicit instruction, the Cross-Reference Validator catches this immediately and blocks the incorrect action. **How it works:** -1. AI proposes an action (e.g., "connect to database on port 27017") +1. AI proposes an action (e.g., "connect to database on the MongoDB default port") 2. Validator retrieves your instruction history -3. Validator detects conflict: you said "use port 27027" +3. Validator detects conflict: you said "use non-default project port" 4. Validator rejects the action and alerts the AI 5. AI revises its action to match your instruction @@ -186,7 +186,7 @@ Calculates overall pressure level: Analyzes five dimensions of every AI action: 1. **Alignment (30% weight):** Does the action match the reasoning? - - "I'm connecting to port 27017 because you requested 27027" → CONFLICT + - "I'm connecting to MongoDB default port because you requested 27027" → CONFLICT 2. **Coherence (20% weight):** Is the reasoning internally consistent? - "Installing React framework... using Vue" → CONTRADICTION @@ -356,7 +356,7 @@ The system adapts communication to respect diverse stakeholder backgrounds: - Changes can break things **Examples:** -- "MongoDB runs on port 27027" +- "MongoDB runs on a non-default project port" - "Use Node.js version 18+" - "Environment variables stored in .env file" - "Database name is 'tractatus_dev'" @@ -411,7 +411,7 @@ The system adapts communication to respect diverse stakeholder backgrounds: - Prohibitive language ("not," "don't use") - Values-laden statements -**Example:** "Always use port 27027 for MongoDB" → HIGH +**Example:** "Always use non-default project port for MongoDB" → HIGH **Why:** Explicit ("always"), specific (27027), SYSTEM domain **In Tractatus:** HIGH persistence instructions are stored in `.claude/instruction-history.json` and checked before EVERY action. Violating them requires explicit human override. @@ -569,7 +569,7 @@ Weighted combination of five verification dimensions: - Stated goal and actual action are consistent **What poor alignment looks like:** -- "Connecting to port 27027 because user requested it" + action connects to 27017 +- "Connecting to non-default project port because user requested it" + action connects to 27017 - "Using React as instructed" + action installs Vue **In Tractatus:** Alignment gets the highest weight (30%) because misalignment often indicates the core 27027 failure mode. @@ -981,8 +981,8 @@ When similar case arises: - Intentional (not accidental or exploratory) **Examples:** -- Explicit: "Always use port 27027 for MongoDB" -- Not explicit: "I wonder if port 27027 would work better?" +- Explicit: "Always use non-default project port for MongoDB" +- Not explicit: "I wonder if non-default project port would work better?" **In Tractatus:** Explicit instructions are detected by the Instruction Persistence Classifier and stored for cross-reference validation. They form the foundation of the 27027 prevention system. @@ -1061,7 +1061,7 @@ When similar case arises: ### When Tractatus Helps You **Scenario 1: Preventing Pattern Recognition Bias** -You tell the AI: "Use port 27027." AI's training pattern immediately tries to use 27017 (the standard default). Cross-Reference Validator catches this pattern override, blocks the action, and auto-corrects to use port 27027 as you instructed. Crisis avoided. +You tell the AI: "Use non-default project port." AI's training pattern immediately tries to use 27017 (the standard default). Cross-Reference Validator catches this pattern override, blocks the action, and auto-corrects to use non-default project port as you instructed. Crisis avoided. **Scenario 2: Protecting Your Values** AI suggests: "I can improve performance by storing user tracking data." Boundary Enforcer recognizes this is a values decision (privacy vs. performance) and blocks autonomous execution. AI presents the trade-offs; you decide. Your agency protected. diff --git a/For Claude Web/tractatus-claude-web-complete/docs/markdown/core-concepts.md b/For Claude Web/tractatus-claude-web-complete/docs/markdown/core-concepts.md index 1194102a..923380e8 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/markdown/core-concepts.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/markdown/core-concepts.md @@ -24,7 +24,7 @@ Classifies user instructions to determine how long they should persist and how s Not all instructions are equally important: -- "Use MongoDB port 27017" (critical, permanent) +- "Use MongoDB MongoDB default port" (critical, permanent) - "Write code comments in JSDoc format" (important, project-scoped) - "Add a console.log here for debugging" (temporary, task-scoped) @@ -86,7 +86,7 @@ Without classification, AI treats all instructions equally, leading to: The classifier also scores how explicit an instruction is (0.0 - 1.0): -- **0.9-1.0**: Very explicit ("Always use port 27017") +- **0.9-1.0**: Very explicit ("Always use MongoDB default port") - **0.7-0.9**: Explicit ("Prefer functional style") - **0.5-0.7**: Somewhat explicit ("Keep code clean") - **0.3-0.5**: Implied ("Make it better") @@ -124,8 +124,8 @@ Validates AI actions against the instruction history to prevent contradictions a ### The Problem It Solves: The 27027 Incident **Real-world failure:** -1. User: "Check MongoDB at port 27027" -2. AI: [Immediately] "Here's code using port 27017" +1. User: "Check MongoDB at non-default project port" +2. AI: [Immediately] "Here's code using MongoDB default port" 3. Result: Application fails to connect to database (running on 27027, not 27017) This happened because: @@ -743,7 +743,7 @@ Status changed from "Phase 2 enhancement" to **mandatory sixth service** because ### Example: Preventing the 27027 Incident -**User instruction:** "Check MongoDB at port 27027" +**User instruction:** "Check MongoDB at non-default project port" 1. **InstructionPersistenceClassifier**: - Quadrant: SYSTEM @@ -752,15 +752,15 @@ Status changed from "Phase 2 enhancement" to **mandatory sixth service** because - Note: "Conflicts with training pattern (27017)" - Stores in instruction database -**Immediately, AI about to propose action:** "Connect to MongoDB on port 27017" (training pattern) +**Immediately, AI about to propose action:** "Connect to MongoDB on the MongoDB default port" (training pattern) 2. **CrossReferenceValidator**: - Checks action against instruction history - Detects pattern recognition bias override (27017 vs 27027) - Conflict type: training_pattern_override - Status: REJECTED - - Auto-corrects to port 27027 - - Alerts: "You specified port 27027, using that instead of default 27017" + - Auto-corrects to non-default project port + - Alerts: "You specified non-default project port, using that instead of default 27017" 3. **BoundaryEnforcer**: - Not needed (technical decision, not values) diff --git a/For Claude Web/tractatus-claude-web-complete/docs/markdown/introduction-to-the-tractatus-framework.md b/For Claude Web/tractatus-claude-web-complete/docs/markdown/introduction-to-the-tractatus-framework.md index 693046c4..8ffe0248 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/markdown/introduction-to-the-tractatus-framework.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/markdown/introduction-to-the-tractatus-framework.md @@ -58,7 +58,7 @@ AI defers to humans when decisions involve: **Prevents:** Pattern recognition bias where LLM training overrides explicit instructions. -**Example:** User says "MongoDB port 27027", LLM's training pattern autocorrects to "27017". CrossReferenceValidator blocks this as instruction conflict. +**Example:** User says "MongoDB non-default project port", LLM's training pattern autocorrects to "27017". CrossReferenceValidator blocks this as instruction conflict. ### 3. BoundaryEnforcer @@ -127,9 +127,9 @@ Applied to AI safety: ## Demonstrated Failure Modes Prevented -### Port 27027 Incident (2025-10-06) +### Non-default project port Incident (2025-10-06) -**What happened:** User specified MongoDB port 27027. LLM immediately used 27017 instead—not through forgetting, but through pattern recognition autocorrection. Training data "MongoDB=27017" was so strong it overrode the explicit instruction in real-time. +**What happened:** User specified MongoDB non-default project port. LLM immediately used 27017 instead—not through forgetting, but through pattern recognition autocorrection. Training data "MongoDB=27017" was so strong it overrode the explicit instruction in real-time. **Tractatus prevention:** InstructionPersistenceClassifier + CrossReferenceValidator store explicit parameters and block any action conflicting with stored instructions—even from training patterns. @@ -174,7 +174,7 @@ Applied to AI safety: - Novel framework for values pluralism in AI ### Software Teams Deploying LLMs -- Production-ready code (tested, documented) +- Code under active maintenance (tested, documented) - Immediate safety improvements - Integration guides for existing systems - Prevents known failure modes @@ -195,8 +195,8 @@ Applied to AI safety: **What Tractatus does NOT do:** - Train better LLMs (uses existing models as-is) -- Guarantee "aligned" AI behavior -- Eliminate all possible failures +- Architecturally constrain "aligned" AI behaviour +- Reduce failure classes addressable by structural enforcement - Replace human judgment **What Tractatus DOES do:** diff --git a/For Claude Web/tractatus-claude-web-complete/docs/markdown/llm-integration-feasibility-research-scope.md b/For Claude Web/tractatus-claude-web-complete/docs/markdown/llm-integration-feasibility-research-scope.md index 93eba6d6..2d9f5915 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/markdown/llm-integration-feasibility-research-scope.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/markdown/llm-integration-feasibility-research-scope.md @@ -617,9 +617,9 @@ LLM: [Uses 27027 despite training pattern] 4. Evaluate hybrid strategies (RAG for 80% + middleware for 20%) **Success Criteria**: -- 50 rules: <200ms overhead (<15% increase) -- 100 rules: <400ms overhead (<30% increase) -- 200 rules: <800ms overhead (<60% increase) +- 50 rules: <200ms overhead (<15% increase) [NEEDS VERIFICATION] +- 100 rules: <400ms overhead (<30% increase) [NEEDS VERIFICATION] +- 200 rules: <800ms overhead (<60% increase) [NEEDS VERIFICATION] - Accuracy maintained across all scales (>95%) --- @@ -820,7 +820,7 @@ LLM: [Uses 27027 despite training pattern] **Minimum Viable Integration**: - ✅ Instruction persistence: 100% across 50+ conversation turns - ✅ Override prevention: <2% failure rate (vs. ~15% baseline) -- ✅ Latency impact: <15% increase for 50-rule database +- ✅ Latency impact: <15% increase for 50-rule database [NEEDS VERIFICATION] - ✅ Scalability: Support 100 rules with <30% overhead - ✅ Multi-tenant: 5-level hierarchy with <10ms conflict resolution @@ -1336,7 +1336,7 @@ If PoC successful (95%+ enforcement, <20% latency, 100% persistence): **New questions introduced by memory tool approach**: -1. **API Maturity**: Are memory/context editing APIs production-ready or beta? +1. **API Stability**: Are memory/context editing APIs stable or beta? 2. **Access Control**: How to implement multi-tenant access to shared memory? 3. **Encryption**: Does memory tool support encrypted storage of sensitive rules? 4. **Versioning**: Can memory tool track rule evolution over time? diff --git a/For Claude Web/tractatus-claude-web-complete/docs/markdown/organizational-theory-foundations.md b/For Claude Web/tractatus-claude-web-complete/docs/markdown/organizational-theory-foundations.md index 6667d52c..e5f39bba 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/markdown/organizational-theory-foundations.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/markdown/organizational-theory-foundations.md @@ -141,11 +141,11 @@ The translation from organizational theory to AI safety architecture manifests i The organizational theory foundation explains why Tractatus prevents failures like the 27027 incident: -**Without organizational structure**: AI's training patterns (MongoDB = port 27017) immediately override user's explicit instruction (port 27027). The system has no concept of instruction persistence or authority domains. +**Without organizational structure**: AI's training patterns (MongoDB = MongoDB default port) immediately override user's explicit instruction (non-default project port). The system has no concept of instruction persistence or authority domains. **With Tractatus organizational structure**: 1. User instruction classified as SYSTEM quadrant, HIGH persistence -2. AI's proposed action (use port 27017) flagged by CrossReferenceValidator +2. AI's proposed action (use MongoDB default port) flagged by CrossReferenceValidator 3. BoundaryEnforcer requires verification before overriding high-persistence instruction 4. Conflict prevented before execution diff --git a/For Claude Web/tractatus-claude-web-complete/docs/markdown/technical-architecture.md b/For Claude Web/tractatus-claude-web-complete/docs/markdown/technical-architecture.md index 2ffed30c..683a1320 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/markdown/technical-architecture.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/markdown/technical-architecture.md @@ -78,7 +78,7 @@ This layer consists of six core services that monitor, classify, validate, verif **Integration:** Intercepts all user instructions for classification before execution -**Example Use Case:** User says "Always use MongoDB port 27027" → Classified as SYSTEM/HIGH/session → Stored for validation +**Example Use Case:** User says "Always use MongoDB non-default project port" → Classified as SYSTEM/HIGH/session → Stored for validation --- @@ -95,7 +95,7 @@ This layer consists of six core services that monitor, classify, validate, verif **Integration:** Called before database operations, config changes, architecture decisions -**Example Use Case:** The 27027 Incident - AI attempted to use default port 27017, validator caught conflict with explicit instruction to use 27027 +**Example Use Case:** The 27027 Incident - AI attempted to use default MongoDB default port, validator caught conflict with explicit instruction to use 27027 --- diff --git a/For Claude Web/tractatus-claude-web-complete/docs/plans/integrated-implementation-roadmap-2025.md b/For Claude Web/tractatus-claude-web-complete/docs/plans/integrated-implementation-roadmap-2025.md index 22e7d634..3c13fa46 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/plans/integrated-implementation-roadmap-2025.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/plans/integrated-implementation-roadmap-2025.md @@ -47,7 +47,7 @@ This integrated roadmap combines: - inst_039 created: Mandatory content accuracy protocol for card presentations - Comprehensive audit of 133 markdown files for 5→6 service references and rule violations - Fixed PITCH-EXECUTIVE.md: Updated to six core services, added PluralisticDeliberationOrchestrator - - Fixed prohibited language: "guarantees" → "constraints" (2 files) + - Fixed prohibited absolute-assurance language per inst_017 (2 files) - generate-card-sections.js script created for systematic card presentation implementation - Docs/markdown files verified correct (historical context appropriate) - 🔄 GitHub Repository Setup - Partially Complete (October 12, 2025) @@ -186,9 +186,9 @@ Response: - [ ] Design timeline visualization UI - [ ] Implement step-by-step progression: - - User specifies port 27027 + - User specifies non-default project port - Context pressure builds (107k tokens) - - AI uses default port 27017 (pattern bias) + - AI uses default MongoDB default port (pattern bias) - Tractatus catches conflict - [ ] Create animation for validation process - [ ] Add explanatory text at each step @@ -390,9 +390,9 @@ Response: **What's Complete:** - ✅ API Reference page at `/api-reference.html` (37KB → expanded from 17KB, deployed to production) - ✅ All 6 governance service endpoints documented with full details -- ✅ OpenAPI 3.0 specification at `/docs/api/openapi.yaml` (1,621 lines, 46KB) -- ✅ JavaScript code examples at `/docs/api/examples-javascript.md` (20KB, 638 lines) -- ✅ Python code examples at `/docs/api/examples-python.md` (30KB, 983 lines) +- ✅ OpenAPI 3.0 specification file (1,621 lines, 46KB) +- ✅ JavaScript code examples documentation (20KB, 638 lines) +- ✅ Python code examples documentation (30KB, 983 lines) - ✅ Authentication endpoints (POST /auth/login, GET /auth/me, POST /auth/logout) - ✅ Document endpoints (GET /documents, POST /documents, search, etc.) - ✅ Governance endpoints (classify, validate, enforce, pressure, verify) @@ -419,9 +419,9 @@ Response: - File: `/docs/api/examples-javascript.md` (20KB, 638 lines) - File: `/docs/api/examples-python.md` (30KB, 983 lines) - Deployed to: https://agenticgovernance.digital/api-reference.html -- Deployed to: https://agenticgovernance.digital/docs/api/openapi.yaml -- Deployed to: https://agenticgovernance.digital/docs/api/examples-javascript.md -- Deployed to: https://agenticgovernance.digital/docs/api/examples-python.md +- Deployed to agenticgovernance.digital (OpenAPI specification) +- Deployed to agenticgovernance.digital (JavaScript examples) +- Deployed to agenticgovernance.digital (Python examples) - **Result:** Task 12 100% complete (Swagger UI deferred as optional) --- diff --git a/For Claude Web/tractatus-claude-web-complete/docs/pluralistic-values-deliberation-plan-v2.md b/For Claude Web/tractatus-claude-web-complete/docs/pluralistic-values-deliberation-plan-v2.md index 0c1840ca..0e0104eb 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/pluralistic-values-deliberation-plan-v2.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/pluralistic-values-deliberation-plan-v2.md @@ -348,7 +348,7 @@ Dissenting view acknowledged: [Link to privacy advocate statement] - [ ] Production deployment **Deliverables:** -- Production-ready component +- Active component - User documentation - Training materials for deliberation facilitators diff --git a/For Claude Web/tractatus-claude-web-complete/docs/research/concurrent-session-architecture-limitations.md b/For Claude Web/tractatus-claude-web-complete/docs/research/concurrent-session-architecture-limitations.md index 170d4f68..ea7a6b9b 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/research/concurrent-session-architecture-limitations.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/research/concurrent-session-architecture-limitations.md @@ -107,7 +107,7 @@ function addInstruction(newInstruction) { **Impact**: Last-write-wins behavior, instruction additions can be silently lost. -**Frequency**: Low under normal use (instruction additions are infrequent), but probabilistically guaranteed under concurrent operation. +**Frequency**: Low under normal use (instruction additions are infrequent), but probabilistically inevitable under concurrent operation. ### 2.2 Session State Contamination @@ -393,7 +393,7 @@ MongoDB Collections: Claude Code Instance → Reads from MongoDB (supports concurrent reads) - → Writes with transaction support (ACID guarantees) + → Writes with transaction support (ACID properties) ``` **Capabilities**: @@ -513,7 +513,7 @@ Claude Code Instance **Workaround 3: Session Serialization** - **Approach**: Stop all Claude Code sessions before starting new one - **Implementation**: `pkill` Claude Code processes, verify before starting -- **Pros**: Guarantees single session, no conflicts +- **Pros**: Enforces single session, no conflicts - **Cons**: Disruptive, prevents parallelism, manual process ### 7.2 Short-Term Solutions (Minimal Code) @@ -613,9 +613,9 @@ Claude Code Instance **Concurrency Model**: ACID transactions, row-level locking **State Management**: Centralized, transactional **Conflict Resolution**: Locks, isolation levels, optimistic concurrency -**Lesson**: Centralized state enables strong consistency guarantees +**Lesson**: Centralized state enables strong consistency enforcement -**Tractatus Difference**: File-based state lacks transactional guarantees +**Tractatus Difference**: File-based state lacks transactional enforcement **Takeaway**: Database-backed state natural fit for multi-session needs ### 9.3 Collaborative Editing (Google Docs, VS Code Live Share) @@ -816,7 +816,7 @@ T3: Session B adds inst_020, writes file (19 instructions) T4: File contains inst_020 only (inst_019 lost!) ``` -**Probability**: Low under normal use, 100% guaranteed under heavy concurrent writes. +**Probability**: Low under normal use, approaches certainty under heavy concurrent writes. **Mitigation**: File locking or atomic operations required. diff --git a/For Claude Web/tractatus-claude-web-complete/docs/research/executive-summary-tractatus-inflection-point.md b/For Claude Web/tractatus-claude-web-complete/docs/research/executive-summary-tractatus-inflection-point.md index dbf7658e..bfac4269 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/research/executive-summary-tractatus-inflection-point.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/research/executive-summary-tractatus-inflection-point.md @@ -25,7 +25,7 @@ This isn't theoretical research. These are operational results from a live produ ## What Is the Tractatus Framework? -Tractatus is an external governance control plane that integrates with AI agent runtimes (like Claude Code) to enforce structural safety boundaries that instructions alone cannot guarantee. +Tractatus is an external governance control plane that integrates with AI agent runtimes (like Claude Code) to enforce structural safety boundaries that instructions alone cannot establish. **Six Core Services:** @@ -44,7 +44,7 @@ Traditional CLAUDE.md files contain valuable guidance, but they have fundamental As conversations get longer (>100k tokens), AI systems forget or deprioritize earlier directives. With Tractatus, critical instructions persist at 95% retention vs. 60-70% for CLAUDE.md. **Problem 2: Pattern Recognition Bias** -The "27027 Incident" (documented case study): User specified MongoDB port 27027, AI used default port 27017 instead. Why? Pattern recognition overpowered explicit instruction. CLAUDE.md couldn't prevent this. Tractatus caught it with 100% reliability. +The "27027 Incident" (documented case study): User specified MongoDB non-default project port, AI used default MongoDB default port instead. Why? Pattern recognition overpowered explicit instruction. CLAUDE.md couldn't prevent this. Tractatus caught it with 100% reliability. **Problem 3: Values Boundary Creep** AI systems can inadvertently make value-laden decisions (privacy policies, ethical trade-offs) without recognizing they've crossed into human-only territory. Tractatus BoundaryEnforcer: 100% detection rate with zero unauthorized values decisions. diff --git a/For Claude Web/tractatus-claude-web-complete/docs/research/rule-proliferation-and-transactional-overhead.md b/For Claude Web/tractatus-claude-web-complete/docs/research/rule-proliferation-and-transactional-overhead.md index a417f60f..953f461f 100644 --- a/For Claude Web/tractatus-claude-web-complete/docs/research/rule-proliferation-and-transactional-overhead.md +++ b/For Claude Web/tractatus-claude-web-complete/docs/research/rule-proliferation-and-transactional-overhead.md @@ -121,7 +121,7 @@ Every significant action now requires: **Total addition**: 251 lines, ~350 tokens -**Impact**: 16.7% increase in instruction history size from single incident +**Impact**: 16.7% increase in instruction history size from single incident [NEEDS VERIFICATION] ### 2.2 Why Rules Were Necessary @@ -311,7 +311,7 @@ Instructions categorized by type: Current (3 instructions): - inst_016: Never fabricate statistics - inst_017: Never use prohibited language -- inst_018: Never claim production-ready without evidence +- inst_018: Never claim maturity without documented evidence Consolidated (1 instruction): - inst_019: Marketing Content Integrity