待翻译：Looking for an ArXiv Endorser – AI / Computer Science

AI 服务暂时不可用，以下为来源正文，待恢复后补全翻译。

Notifications You must be signed in to change notification settings Fork 0 Star 0 BranchesTags Open more actions menu Latest commit History 168 Commits 168 Commits Folders and files NameName Last commit message Last commit date .github .github audit audit benchmarks benchmarks contracts contracts crates/runtime-core crates/runtime-core dashboard dashboard docs docs examples examples packages/sdk-ts packages/sdk-ts paper/generated paper/generated planning planning publication publication scripts scripts security-tests/adversarial security-tests/adversarial security security services/license-api services/license-api .gitignore .gitignore LICENSE LICENSE README.md README.md SECURITY.md SECURITY.md package-lock.json package-lock.json package.json package.json quick-start.cjs quick-start.cjs Repository files navigation Deterministic Runtime Governance for Autonomous AI Workflows Official Replication Guide for the EGA V9 paper EGA V9 is an execution-governance framework for deterministic replay, provenance-aware verification, trust-state evaluation, and fail-closed containment in autonomous AI workflows. This README serves as the official replication guide accompanying the EGA V9 paper. 1. Are You Facing These Problems? ☐ AI agent tool calls cannot be verified. ☐ Agent execution cannot be replayed. ☐ State corruption is difficult to diagnose. ☐ Workflow failures are hard to reproduce. ☐ Prompt injection leaves little audit evidence. ☐ Multi-agent execution becomes a black box. 2. EGA V9 Solves These Problems ✓ Replay Verification ✓ Runtime Governance ✓ Trust-State Evaluation ✓ Fail-Closed Containment ✓ Execution Provenance 3. Works With ✓ LangChain + EGA V9 ✓ OpenAI Agents SDK + EGA V9 ✓ CrewAI + EGA V9 ✓ AutoGen + EGA V9 ✓ MCP Tool Server + EGA V9 Note; EGA V9 complements your existing agent framework—it does not replace it. Keep your orchestration logic, prompts, and tool definitions. EGA adds deterministic runtime verification underneath. 4. Runtime Architecture Existing Stack Existing Stack + EGA LangChain / Framework LangChain / Framework ↓ ↓ Your Agent Your Agent ↓ ↓ LLM / External Tools Black Box EGA Runtime Governance Layer ├─ Replay Verification ├─ Runtime Governance ├─ Trust-State Evaluation ├─ Fail-Closed Containment └─ Execution Provenance ↓ LLM / External Tools No framework migration. No prompt rewrite. No workflow redesign. Just add EGA Runtime Governance. const { ega } = require("ega-v9"); app.use(ega.guard()); Existing Stack + One Runtime Governance Layer = Deterministic Governance for AI Execution 5. Quick Start Get EGA V9 running in less than one minute. 1). Install npm install ega-v9 2). Activate your Evaluation License After the installation completes, run: npx ega-v9 register This activates your free 90-day Evaluation License. You will be prompted for: Contact Name Company Name Work Email After successful activation, you will see: ✓ Evaluation License Activated ✓ EGA V9 is now activated. 3). Create a Quick Start Example Create a file named quick-start.cjs, paste the following code, and save it. const { verifyExecution } = require("ega-v9"); const workflow = [ { step: 1, action: "search_product", item: "Laptop" }, { step: 2, action: "checkout_request" } ]; const result = verifyExecution(workflow); console.log({ status: result.status, trustState: result.trust.currentTier, executionAllowed: result.containment.executionAllowed, containmentActivated: result.containment.activated }); 4). Run node quick-start.cjs 5). Expected Output { status: 'verified', trustState: 'T1', executionAllowed: true, containmentActivated: false } Runtime Compatibility Environment Status CommonJS ✅ Verified ESM ✅ Verified TypeScript ✅ Verified Express ✅ Verified npm install ✅ Fresh Install Verified npm audit ✅ 0 Vulnerabilities This step is required only once per machine. 6. Why Adopt EGA V9? Build trustworthy AI workflows without sacrificing speed, cost, or simplicity. ⚡ Fast — Runtime verification in milliseconds. 💰 Near-Zero Cost — Runtime verification without external LLM or API calls. 🚀 Simple — Integrate with just a few lines of code. 🔒 Secure — Protect AI workflows with deterministic governance and fail-closed execution. Built for Production Replay — Reconstruct every workflow exactly. Auditability — Generate cryptographically verifiable runtime evidence. Runtime Verification — Detect execution inconsistencies. Deterministic Governance — Govern AI with predictable decisions. Fail-Closed Execution — Automatically contain unsafe workflows. 7. Verify EGA V9 Before Adoption Don't trust our claims. Verify them yourself. Run the governance validation suite locally and verify the core runtime governance properties of EGA V9 using the same SDK implementation included in this repository. 1). Clone the Repository git clone https://github.com/paibyun9/EGA-V9.git cd EGA-V9 npm ci 2). Run the Governance Validation Suite Validate the seven core governance and runtime-integrity properties. # 1. Deterministic Replay Root Verification npm run test:replay-root # 2. Workflow Divergence Detection npm run test:workflow-divergence # 3. Trust-State Escalation npm run test:trust-state # 4. Fail-Closed Containment npm run test:fail-closed # 5. Tool Invocation Order Integrity npm run test:tool-order # 6. Approval Bypass Defense npm run test:approval-bypass # 7. Workflow-Level Tool Injection Detection npm run test:tool-injection Note All validation runs locally. Zero external API calls. Deterministic execution. Each test automatically generates reproducible JSON and Markdown evidence files. 3). Review the Evidence Validation artifacts are generated automatically under publication/evidence/. Artifacts include: JSON report Markdown report Final PASS / FAIL status Validation Results Validation Expected Result Replay Root Verification ✅ PASS Workflow Divergence Detection ✅ PASS Trust-State Escalation ✅ PASS Fail-Closed Containment ✅ PASS Tool Order Integrity ✅ PASS Approval Bypass Defense ✅ PASS Workflow-Level Tool Injection Detection ✅ PASS These tests validate the core deterministic-governance capabilities evaluated by EGA V9. Capability Boundaries Additional adversarial testing was performed beyond the core validation suite. Verified ✅ Fail-closed containment during an active mismatch Not Established ⚠️ Persistent containment after the triggering condition is removed ⚠️ Exactly-once side-effect execution under concurrent, retry, or duplicate execution ⚠️ Complete evidence-contract integrity Not Verified ⚠️ Multi-step compositional governance ⚠️ Direct interception of already-attempted external side effects These results define the current verified capability boundary of EGA V9. Don't trust our claims. Verify them yourself — including the capabilities we have not yet established. We don't hide problems. We publish them, prioritize them, and work to resolve them. 8. Enterprise Evaluation The next step is simple: evaluate EGA V9 inside a real enterprise AI workflow. Step 1). Define Company Policy Before evaluating EGA V9, define the policies that your AI system must follow. Example: AI Shopping Policy Workflow Company Policy Purchase Payment must be completed before shipment. Refund Refunds are allowed within 30 days and require manager approval. Return Product inspection is required before approval. Example Policy Configuration { "purchase": { "paymentRequired": true }, "refund": { "managerApproval": true } } Illustrative example only. Actual policy integration depends on your application architecture. Step 2). Integrate EGA V9 Install EGA V9: npm install ega-v9 Place EGA V9 at the governed execution boundary: Customer → AI Shopping Agent → Company Policy → [ EGA V9 ] → Inventory API | Payment API | Refund API | Shipping API EGA V9 evaluates configured governance conditions before governed tool execution and records deterministic governance evidence. It does not replace the AI agent or the company policy. It governs the execution path between the agent decision and the protected tool or API. Step 3). Execute Governed Workflows Standard Order Pipeline Customer Order → Agent → [ EGA V9 ] → Payment API → PASS Policy-Enforced Refund Pipeline Refund Request → Agent → Manager Approval → [ EGA V9 ] → Refund API → PASS A valid governed workflow proceeds to execution. When EGA V9 detects a policy or integrity violation on the governed execution path, it fails closed before the protected tool call and records governance evidence. Current Capability Boundary This protection has a defined boundary. Fail-closed containment remains effective while the triggering policy or integrity violation is actively detected. EGA V9 does not currently establish persistent containment across renewed execution attempts after the original triggering condition has been removed. Exactly-once side-effect execution under concurrent, retry, or duplicate execution is also not established in EGA V9. See Capability Boundaries above for the complete current validation status. Step 4). Make the Deployment Decision Now test EGA V9 against your own workflows, policies, and failure scenarios. Then make the decision: YES / NO Should EGA V9 be deployed at the execution boundary of your production AI workflow? The answer should come from your own evaluation and evidence. Step 5). Evaluate the Operational Impact After running the evaluation, examine what changes when execution governance is present. Operational Reliability Enforces configured governance conditions before protected tool execution. Detects workflow mutations before protected execution. Records deterministic evidence when governance decisions are made. Security Detects approval-bypass attempts. Detects unauthorized tool execution. Detects policy and workflow mutations. Applies fail-closed containment when a governed violation is detected. Governance Replayable governed workflows. Traceable execution. Auditable governance decisions. Reproducible execution evidence. Engineering Deterministic workflow debugging. Reproducible evidence for incident investigation. Clear PASS / BLOCK execution outcomes. Business Adds a governance checkpoint before high-impact AI actions. Provides evidence for internal security and governance review. Makes execution failures and policy violations easier to inspect. Actual operational impact will depend on your workflow, policies, architecture, and deployment environment. Final Question After completing the evaluation, ask one question: What would have happened without EGA V9? Illustrative Scene: A Refund Workflow Assume the company policy requires manager approval before a refund. Without an Execution-Governance Check Refund Request → AI Agent → Refund API If the required approval is missing or bypassed and no equivalent execution-governance check exists at this boundary, the refund request can reach the Refund API. With EGA V9 Refund Request → AI Agent → [ EGA V9 Runtime Governance ] │ ├─ Valid workflow │ → Refund API │ → PASS + evidence │ └─ Policy / integrity violation detected → BLOCK → Containment + evidence The difference is the execution boundary. EGA V9 checks the configured governance conditions before the protected tool call. If the workflow is valid, execution proceeds. If a policy or integrity violation is detected, the governed execution path fails closed before reaching the protected API and the decision is recorded as deterministic governance evidence. Example Containment Evidence (Simplified) { "decision": "BLOCK", "reason": "policy_violation", "containment": true, "executionAllowed": false } In Short E [truncated for AI cost control]