c984ebfd7d
Created new /architecture.html page with generic architecture diagram that emphasizes Tractatus works with any agent runtime (not just Claude Code). Revised tone to reflect early-stage research status, including limitations section and call for industry collaboration, per user feedback on avoiding overconfident claims. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>
65 lines
3.9 KiB
Text
65 lines
3.9 KiB
Text
graph TB
|
|
subgraph "Agent Runtime Layer (Any LLM Agent System)"
|
|
AGENT["<b>Agentic AI Runtime</b><br/>LangChain • AutoGPT • CrewAI<br/>Claude Code • Custom Agents<br/>Multi-Agent Systems<br/><i>Tool Use • Planning • Execution</i>"]
|
|
end
|
|
|
|
subgraph "Tractatus Governance Layer (External Enforcement)"
|
|
BE["<b>BoundaryEnforcer</b><br/>Blocks values decisions<br/>• Privacy policies<br/>• Ethical trade-offs<br/>• Strategic direction<br/>• User agency violations<br/><b style='color:#ef4444'>⚠ Cannot be bypassed by prompting</b>"]
|
|
IPC["<b>InstructionPersistenceClassifier</b><br/>Classifies & stores instructions<br/>• Quadrant (STR/OPS/TAC/SYS)<br/>• Persistence (HIGH/MED/LOW)<br/>• Temporal scope<br/><b style='color:#ef4444'>⚠ External to AI memory</b>"]
|
|
CRV["<b>CrossReferenceValidator</b><br/>Prevents pattern bias override<br/>• Checks instruction history<br/>• Detects conflicts (27027)<br/>• Blocks contradictions<br/><b style='color:#ef4444'>⚠ Independent verification</b>"]
|
|
CPM["<b>ContextPressureMonitor</b><br/>Detects degraded conditions<br/>• Token budget tracking<br/>• Error accumulation<br/>• Checkpoint reporting<br/><b style='color:#ef4444'>⚠ Objective metrics, not self-reported</b>"]
|
|
MV["<b>MetacognitiveVerifier</b><br/>Validates complex operations<br/>• >3 files or >5 steps<br/>• Architecture changes<br/>• Confidence scoring<br/><b style='color:#ef4444'>⚠ Structural pause-and-verify</b>"]
|
|
PDO["<b>PluralisticDeliberationOrchestrator</b><br/>Facilitates values deliberation<br/>• Multi-stakeholder engagement<br/>• Moral framework mapping<br/>• Precedent documentation<br/><b style='color:#ef4444'>⚠ Human judgment required</b>"]
|
|
end
|
|
|
|
subgraph "Persistent Storage Layer (Immutable Audit Trail)"
|
|
GR["<b>governance_rules</b><br/>• rule_id (STR-001...)<br/>• quadrant<br/>• persistence level<br/>• enforced_by<br/>• violation_action<br/>• active status"]
|
|
AL["<b>audit_logs</b><br/>• timestamp<br/>• service (which enforcer)<br/>• action (BLOCK/WARN)<br/>• instruction<br/>• rule_violated<br/>• session_id"]
|
|
SS["<b>session_state</b><br/>• session_id<br/>• token_count<br/>• message_count<br/>• pressure_level<br/>• last_checkpoint<br/>• framework_active"]
|
|
IH["<b>instruction_history</b><br/>• instruction_id<br/>• content<br/>• classification<br/>• persistence<br/>• created_at<br/>• active status"]
|
|
end
|
|
|
|
subgraph "Human Approval Workflows"
|
|
HA["<b>Human Oversight</b><br/>Values Decisions<br/>Strategic Changes<br/>Boundary Violations<br/><i>Final authority on incommensurable values</i>"]
|
|
end
|
|
|
|
%% Data Flow - Agent to Governance
|
|
AGENT -->|"All actions pass through governance checks"| BE
|
|
AGENT --> IPC
|
|
AGENT --> CRV
|
|
AGENT --> CPM
|
|
AGENT --> MV
|
|
AGENT --> PDO
|
|
|
|
%% Governance to Storage
|
|
BE --> GR
|
|
BE --> AL
|
|
IPC --> GR
|
|
IPC --> IH
|
|
CRV --> IH
|
|
CRV --> AL
|
|
CPM --> SS
|
|
CPM --> AL
|
|
MV --> AL
|
|
PDO --> AL
|
|
|
|
%% Human Approval Flow
|
|
BE -->|"Boundary violation"| HA
|
|
PDO -->|"Values conflict"| HA
|
|
HA -->|"Approval/Rejection"| BE
|
|
|
|
%% Styling
|
|
classDef agent fill:#dbeafe,stroke:#3b82f6,stroke-width:3px
|
|
classDef governance fill:#f0fdf4,stroke:#10b981,stroke-width:3px
|
|
classDef persistence fill:#fef9c3,stroke:#eab308,stroke-width:2px
|
|
classDef human fill:#fce7f3,stroke:#ec4899,stroke-width:3px
|
|
|
|
class AGENT agent
|
|
class BE,IPC,CRV,CPM,MV,PDO governance
|
|
class GR,AL,SS,IH persistence
|
|
class HA human
|
|
|
|
%% Key Insight Box
|
|
NOTE["<b style='color:#dc2626; font-size:14px'>🔒 KEY JAILBREAK DEFENSE</b><br/><span style='font-size:12px'>Governance layer operates OUTSIDE agent runtime<br/>Cannot be overridden by adversarial prompts<br/>Structural boundaries, not behavioral training<br/>Immutable audit trail independent of AI</span>"]
|
|
|
|
class NOTE governance
|