@jlongo78/agent-spaces 0.7.6 → 0.8.0

package/.next/standalone/docs/superpowers/plans/2026-03-15-cortex-v2-boundary-engine.md

@@ -1,853 +1,853 @@
-# Cortex v2 — Pillar 4: Boundary Engine
-
-> **For agentic workers:** REQUIRED: Use superpowers:subagent-driven-development (if subagents available) or superpowers:executing-plans to implement this plan. Steps use checkbox (`- [ ]`) syntax for tracking.
-
-**Goal:** Add three-layer access control to Cortex: auto-classification of sensitivity, organizational policies, and creator overrides — ensuring confidential knowledge never leaks across scope boundaries.
-
-**Architecture:** A new `src/lib/cortex/boundary/` module with three components: a `Classifier` (regex-based sensitivity detection), a `PolicyEngine` (rule evaluation), and an `AccessFilter` (query-time enforcement). The AccessFilter integrates with the ContextEngine to pre-filter scopes before search. An audit log records all access decisions in the entity graph's SQLite database.
-
-**Tech Stack:** TypeScript, better-sqlite3 (existing graph DB), vitest
-
-**Spec:** `docs/superpowers/specs/2026-03-14-cortex-v2-design.md` — Pillar 4
-
-**Depends on:** Pillar 1 (Entity Graph) + Pillar 2 (Knowledge Unit Evolution) — both completed
-
----
-
-## File Structure
-
-```
-New files:
-├── src/lib/cortex/boundary/classifier.ts    — Auto-classification (Layer 1)
-├── src/lib/cortex/boundary/policy.ts        — Policy engine (Layer 2)
-├── src/lib/cortex/boundary/access.ts        — Query-time access filter (Layer 3 + enforcement)
-├── src/lib/cortex/boundary/audit.ts         — Audit trail logging
-├── src/lib/cortex/boundary/index.ts         — Barrel export
-
-Modified files:
-├── src/lib/cortex/ingestion/pipeline.ts     — Auto-classify on ingestion
-├── src/lib/cortex/retrieval/context-engine.ts — Pre-filter with AccessFilter
-├── src/lib/cortex/config.ts                 — Add policies to config
-
-Test files:
-├── tests/lib/cortex/boundary/classifier.test.ts
-├── tests/lib/cortex/boundary/policy.test.ts
-├── tests/lib/cortex/boundary/access.test.ts
-├── tests/lib/cortex/boundary/audit.test.ts
-```
-
----
-
-## Chunk 1: Auto-Classification and Policy Engine
-
-### Task 1: Sensitivity auto-classifier
-
-**Files:**
-- Create: `src/lib/cortex/boundary/classifier.ts`
-- Create: `tests/lib/cortex/boundary/classifier.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/boundary/classifier.test.ts
-import { describe, it, expect } from 'vitest';
-import { classifySensitivity } from '@/lib/cortex/boundary/classifier';
-
-describe('classifySensitivity', () => {
-  it('classifies secrets as confidential', () => {
-    expect(classifySensitivity('Set API_KEY=sk-ant-abc123 in .env')).toBe('confidential');
-    expect(classifySensitivity('password: hunter2')).toBe('confidential');
-    expect(classifySensitivity('DATABASE_URL=postgres://user:pass@host')).toBe('confidential');
-  });
-
-  it('classifies personnel content as confidential', () => {
-    expect(classifySensitivity('Alice performance review: exceeds expectations')).toBe('confidential');
-    expect(classifySensitivity('salary adjustment from 120k to 140k')).toBe('confidential');
-  });
-
-  it('classifies security content as restricted', () => {
-    expect(classifySensitivity('Found SQL injection vulnerability in login endpoint')).toBe('restricted');
-    expect(classifySensitivity('CVE-2024-1234 affects our auth library')).toBe('restricted');
-  });
-
-  it('classifies business content as restricted', () => {
-    expect(classifySensitivity('Q3 revenue was $2.5M, below target')).toBe('restricted');
-    expect(classifySensitivity('Unreleased product launch planned for April')).toBe('restricted');
-  });
-
-  it('classifies technical content as internal', () => {
-    expect(classifySensitivity('We decided to use PostgreSQL for the new service')).toBe('internal');
-    expect(classifySensitivity('The auth middleware handles JWT validation')).toBe('internal');
-  });
-
-  it('classifies general content as public', () => {
-    expect(classifySensitivity('How to use git rebase')).toBe('public');
-    expect(classifySensitivity('JavaScript array methods')).toBe('public');
-  });
-
-  it('returns most restrictive when multiple detectors match', () => {
-    // Has both secret (confidential) and security (restricted) signals
-    expect(classifySensitivity('API_KEY leaked in CVE-2024-1234')).toBe('confidential');
-  });
-});
-```
-
-- [ ] **Step 2: Implement classifier**
-
-```typescript
-// src/lib/cortex/boundary/classifier.ts
-import type { SensitivityClass } from '../knowledge/types';
-
-interface Detector {
-  sensitivity: SensitivityClass;
-  patterns: RegExp[];
-  priority: number;  // higher = more restrictive
-}
-
-const DETECTORS: Detector[] = [
-  {
-    sensitivity: 'confidential',
-    priority: 4,
-    patterns: [
-      // Secrets
-      /\b(api[_-]?key|secret[_-]?key|access[_-]?token|auth[_-]?token)\s*[=:]/i,
-      /\b(password|passwd|pwd)\s*[=:]/i,
-      /\bsk-[a-z]{2,4}-[a-zA-Z0-9]{10,}/,  // Anthropic/OpenAI key format
-      /\b(DATABASE_URL|REDIS_URL|MONGO_URI)\s*=/i,
-      /\bpostgres:\/\/\w+:\w+@/i,
-      /\b(private[_-]?key|ssh[_-]?key)\b/i,
-      // Personnel
-      /\b(performance\s+review|annual\s+review)\b/i,
-      /\b(salary|compensation|pay\s+raise|pay\s+cut)\b/i,
-      /\b(hiring|termination|fired|let\s+go)\b/i,
-      /\b(1:1\s+notes?|one[\s-]on[\s-]one)\b/i,
-    ],
-  },
-  {
-    sensitivity: 'restricted',
-    priority: 3,
-    patterns: [
-      // Security
-      /\b(vulnerab|exploit|attack\s+vector|injection|xss|csrf|ssrf)\b/i,
-      /\bCVE-\d{4}-\d+/i,
-      /\b(incident\s+report|security\s+breach|data\s+leak)\b/i,
-      // Business
-      /\b(revenue|profit|loss|earnings|ARR|MRR)\b/i,
-      /\b(unreleased|pre-launch|confidential\s+plan|roadmap)\b/i,
-      /\bcustomer\s+(data|records|PII)\b/i,
-    ],
-  },
-  {
-    sensitivity: 'internal',
-    priority: 2,
-    patterns: [
-      /\b(we\s+decided|architecture|design\s+pattern|refactor)\b/i,
-      /\b(middleware|service|endpoint|database|schema)\b/i,
-      /\b(deployment|CI\/CD|pipeline|infrastructure)\b/i,
-      /\b(bug\s+fix|pull\s+request|code\s+review)\b/i,
-    ],
-  },
-];
-
-export function classifySensitivity(text: string): SensitivityClass {
-  let highestPriority = 0;
-  let result: SensitivityClass = 'public';
-
-  for (const detector of DETECTORS) {
-    if (detector.priority <= highestPriority) continue;
-    for (const pattern of detector.patterns) {
-      if (pattern.test(text)) {
-        highestPriority = detector.priority;
-        result = detector.sensitivity;
-        break;
-      }
-    }
-  }
-
-  return result;
-}
-```
-
-- [ ] **Step 3: Run tests, commit**
-
-```bash
-git commit -m "feat(cortex): add sensitivity auto-classifier for boundary engine"
-```
-
----
-
-### Task 2: Policy engine
-
-**Files:**
-- Create: `src/lib/cortex/boundary/policy.ts`
-- Create: `tests/lib/cortex/boundary/policy.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/boundary/policy.test.ts
-import { describe, it, expect } from 'vitest';
-import { PolicyEngine } from '@/lib/cortex/boundary/policy';
-import type { Policy } from '@/lib/cortex/boundary/policy';
-
-describe('PolicyEngine', () => {
-  it('returns empty actions when no policies match', () => {
-    const engine = new PolicyEngine([]);
-    const actions = engine.evaluate({ type: 'decision', sensitivity: 'internal' });
-    expect(actions).toEqual([]);
-  });
-
-  it('matches by knowledge type', () => {
-    const policies: Policy[] = [{
-      name: 'arch-decisions-propagate',
-      match: { type: 'decision' },
-      action: { propagate_to: [{ level: 'department' }] },
-    }];
-    const engine = new PolicyEngine(policies);
-    const actions = engine.evaluate({ type: 'decision', sensitivity: 'internal' });
-    expect(actions).toHaveLength(1);
-    expect(actions[0].propagate_to).toEqual([{ level: 'department' }]);
-  });
-
-  it('matches by sensitivity', () => {
-    const policies: Policy[] = [{
-      name: 'lock-confidential',
-      match: { sensitivity: 'confidential' },
-      action: { cannot_propagate: true },
-    }];
-    const engine = new PolicyEngine(policies);
-    const actions = engine.evaluate({ type: 'pattern', sensitivity: 'confidential' });
-    expect(actions).toHaveLength(1);
-    expect(actions[0].cannot_propagate).toBe(true);
-  });
-
-  it('matches by topics', () => {
-    const policies: Policy[] = [{
-      name: 'security-routing',
-      match: { topics: ['security', 'vulnerability'] },
-      action: { max_scope: 'department', propagate_to: [{ level: 'team', entity_id: 'team-security' }] },
-    }];
-    const engine = new PolicyEngine(policies);
-    const actions = engine.evaluate({ type: 'error_fix', sensitivity: 'restricted', topics: ['security'] });
-    expect(actions).toHaveLength(1);
-    expect(actions[0].max_scope).toBe('department');
-  });
-
-  it('returns multiple matching policies', () => {
-    const policies: Policy[] = [
-      { name: 'p1', match: { type: 'decision' }, action: { trickle_down: true } },
-      { name: 'p2', match: { sensitivity: 'internal' }, action: { max_scope: 'organization' } },
-    ];
-    const engine = new PolicyEngine(policies);
-    const actions = engine.evaluate({ type: 'decision', sensitivity: 'internal' });
-    expect(actions).toHaveLength(2);
-  });
-
-  it('does not match when criteria do not overlap', () => {
-    const policies: Policy[] = [{
-      name: 'security-only',
-      match: { topics: ['security'] },
-      action: { cannot_propagate: true },
-    }];
-    const engine = new PolicyEngine(policies);
-    const actions = engine.evaluate({ type: 'decision', sensitivity: 'internal', topics: ['architecture'] });
-    expect(actions).toHaveLength(0);
-  });
-});
-```
-
-- [ ] **Step 2: Implement policy engine**
-
-```typescript
-// src/lib/cortex/boundary/policy.ts
-import type { KnowledgeType, SensitivityClass, ScopeLevel } from '../knowledge/types';
-
-export interface PropagationTarget {
-  level: ScopeLevel;
-  entity_id?: string;
-}
-
-export interface PolicyAction {
-  max_scope?: ScopeLevel;
-  propagate_to?: PropagationTarget[];
-  trickle_down?: boolean;
-  cannot_propagate?: boolean;
-}
-
-export interface Policy {
-  name: string;
-  match: {
-    type?: KnowledgeType;
-    topics?: string[];
-    sensitivity?: SensitivityClass;
-    scope_level?: ScopeLevel;
-  };
-  action: PolicyAction;
-}
-
-export interface PolicyMatchInput {
-  type: KnowledgeType | string;
-  sensitivity: SensitivityClass | string;
-  topics?: string[];
-  scope_level?: ScopeLevel | string;
-}
-
-export class PolicyEngine {
-  constructor(private policies: Policy[]) {}
-
-  evaluate(input: PolicyMatchInput): PolicyAction[] {
-    const matched: PolicyAction[] = [];
-
-    for (const policy of this.policies) {
-      if (this.matches(policy, input)) {
-        matched.push(policy.action);
-      }
-    }
-
-    return matched;
-  }
-
-  private matches(policy: Policy, input: PolicyMatchInput): boolean {
-    const { match } = policy;
-
-    if (match.type && match.type !== input.type) return false;
-    if (match.sensitivity && match.sensitivity !== input.sensitivity) return false;
-    if (match.scope_level && match.scope_level !== input.scope_level) return false;
-
-    if (match.topics && match.topics.length > 0) {
-      const inputTopics = input.topics ?? [];
-      const hasOverlap = match.topics.some(t => inputTopics.includes(t));
-      if (!hasOverlap) return false;
-    }
-
-    return true;
-  }
-}
-```
-
-- [ ] **Step 3: Run tests, commit**
-
-```bash
-git commit -m "feat(cortex): add policy engine for boundary enforcement"
-```
-
----
-
-## Chunk 2: Access Filter and Audit Trail
-
-### Task 3: Query-time access filter
-
-**Files:**
-- Create: `src/lib/cortex/boundary/access.ts`
-- Create: `tests/lib/cortex/boundary/access.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/boundary/access.test.ts
-import { describe, it, expect } from 'vitest';
-import { AccessFilter } from '@/lib/cortex/boundary/access';
-import type { ScoredKnowledge } from '@/lib/cortex/knowledge/types';
-
-function makeUnit(overrides: Partial<ScoredKnowledge> = {}): ScoredKnowledge {
-  return {
-    id: 'k1', vector: [], text: 'test', type: 'decision', layer: 'personal',
-    workspace_id: null, session_id: null, agent_type: 'claude',
-    project_path: null, file_refs: [], confidence: 0.8,
-    created: new Date().toISOString(), source_timestamp: new Date().toISOString(),
-    stale_score: 0, access_count: 0, last_accessed: null, metadata: {},
-    relevance_score: 0.9, similarity: 0.9,
-    sensitivity: 'internal',
-    scope: { level: 'personal', entity_id: 'person-alice' },
-    origin: { source_type: 'conversation', source_ref: '', creator_entity_id: 'person-alice' },
-    ...overrides,
-  };
-}
-
-describe('AccessFilter', () => {
-  const filter = new AccessFilter({
-    requesterId: 'person-alice',
-    requesterScope: { level: 'team', entity_id: 'team-platform' },
-    requesterOrg: 'organization-acme',
-  });
-
-  it('allows public knowledge from anywhere in org', () => {
-    const unit = makeUnit({ sensitivity: 'public', scope: { level: 'organization', entity_id: 'org-acme' } });
-    expect(filter.canAccess(unit)).toBe(true);
-  });
-
-  it('allows internal knowledge within org', () => {
-    const unit = makeUnit({ sensitivity: 'internal', scope: { level: 'team', entity_id: 'team-other' } });
-    expect(filter.canAccess(unit)).toBe(true);
-  });
-
-  it('allows restricted knowledge within same scope', () => {
-    const unit = makeUnit({
-      sensitivity: 'restricted',
-      scope: { level: 'team', entity_id: 'team-platform' },
-    });
-    expect(filter.canAccess(unit)).toBe(true);
-  });
-
-  it('denies restricted knowledge from different department', () => {
-    const unit = makeUnit({
-      sensitivity: 'restricted',
-      scope: { level: 'department', entity_id: 'department-sales' },
-    });
-    // Alice is in team-platform, not department-sales
-    expect(filter.canAccess(unit)).toBe(false);
-  });
-
-  it('allows confidential knowledge from self', () => {
-    const unit = makeUnit({
-      sensitivity: 'confidential',
-      scope: { level: 'personal', entity_id: 'person-alice' },
-      origin: { source_type: 'conversation', source_ref: '', creator_entity_id: 'person-alice' },
-    });
-    expect(filter.canAccess(unit)).toBe(true);
-  });
-
-  it('denies confidential knowledge from others', () => {
-    const unit = makeUnit({
-      sensitivity: 'confidential',
-      scope: { level: 'personal', entity_id: 'person-bob' },
-      origin: { source_type: 'conversation', source_ref: '', creator_entity_id: 'person-bob' },
-    });
-    expect(filter.canAccess(unit)).toBe(false);
-  });
-
-  it('respects creator_scope override (further restriction)', () => {
-    const unit = makeUnit({
-      sensitivity: 'internal',  // normally accessible org-wide
-      creator_scope: { max_level: 'personal' },  // but creator restricted to personal only
-      scope: { level: 'personal', entity_id: 'person-bob' },
-    });
-    // Alice can't access Bob's personal-restricted knowledge even though it's internal
-    expect(filter.canAccess(unit)).toBe(false);
-  });
-
-  it('filters a list of results', () => {
-    const results = [
-      makeUnit({ id: 'a', sensitivity: 'public' }),
-      makeUnit({ id: 'b', sensitivity: 'confidential', origin: { source_type: 'conversation', source_ref: '', creator_entity_id: 'person-bob' } }),
-      makeUnit({ id: 'c', sensitivity: 'internal' }),
-    ];
-    const filtered = filter.filterResults(results);
-    expect(filtered).toHaveLength(2);  // a and c pass, b denied
-    expect(filtered.map(r => r.id)).toEqual(['a', 'c']);
-  });
-});
-```
-
-- [ ] **Step 2: Implement access filter**
-
-```typescript
-// src/lib/cortex/boundary/access.ts
-import type { ScoredKnowledge, Scope, ScopeLevel, SensitivityClass } from '../knowledge/types';
-
-const SCOPE_HIERARCHY: Record<ScopeLevel, number> = {
-  personal: 0,
-  team: 1,
-  department: 2,
-  organization: 3,
-};
-
-export interface AccessFilterConfig {
-  requesterId: string;
-  requesterScope: Scope;        // requester's team scope
-  requesterOrg: string;         // requester's org entity_id
-  grants?: Set<string>;         // knowledge IDs the requester has been granted access to
-}
-
-export class AccessFilter {
-  private config: AccessFilterConfig;
-
-  constructor(config: AccessFilterConfig) {
-    this.config = config;
-  }
-
-  canAccess(unit: ScoredKnowledge): boolean {
-    const sensitivity = (unit.sensitivity ?? 'internal') as SensitivityClass;
-    const unitScope = unit.scope ?? { level: 'personal' as ScopeLevel, entity_id: '' };
-    const creatorId = unit.origin?.creator_entity_id ?? '';
-
-    // Check creator_scope override first — most restrictive wins
-    if (unit.creator_scope) {
-      const maxLevel = SCOPE_HIERARCHY[unit.creator_scope.max_level] ?? 0;
-      const requesterLevel = this.getRequesterProximityLevel(unitScope);
-      if (requesterLevel > maxLevel) return false;
-    }
-
-    switch (sensitivity) {
-      case 'public':
-        return true;
-
-      case 'internal':
-        // Accessible to anyone in the same org
-        return true;
-
-      case 'restricted':
-        // Accessible within same scope or if policy grants cross-scope access
-        return this.isWithinScope(unitScope);
-
-      case 'confidential':
-        // Only creator or explicit grant
-        if (this.config.requesterId === creatorId) return true;
-        if (this.config.grants?.has(unit.id)) return true;
-        return false;
-
-      default:
-        return false;
-    }
-  }
-
-  filterResults(results: ScoredKnowledge[]): ScoredKnowledge[] {
-    return results.filter(r => this.canAccess(r));
-  }
-
-  private isWithinScope(unitScope: Scope): boolean {
-    // Same entity_id = same scope
-    if (unitScope.entity_id === this.config.requesterScope.entity_id) return true;
-    if (unitScope.entity_id === this.config.requesterId) return true;
-    // Same org for org-level scopes
-    if (unitScope.level === 'organization' && unitScope.entity_id === this.config.requesterOrg) return true;
-    return false;
-  }
-
-  private getRequesterProximityLevel(unitScope: Scope): number {
-    // How "far" is the requester from the unit's scope?
-    // 0 = same entity, 1 = same scope level, 2+ = further
-    if (unitScope.entity_id === this.config.requesterId) return 0;
-    if (unitScope.entity_id === this.config.requesterScope.entity_id) return 0;
-    return SCOPE_HIERARCHY[unitScope.level] ?? 3;
-  }
-}
-```
-
-- [ ] **Step 3: Run tests, commit**
-
-```bash
-git commit -m "feat(cortex): add query-time access filter for boundary enforcement"
-```
-
----
-
-### Task 4: Audit trail
-
-**Files:**
-- Create: `src/lib/cortex/boundary/audit.ts`
-- Create: `tests/lib/cortex/boundary/audit.test.ts`
-
-- [ ] **Step 1: Write failing tests**
-
-```typescript
-// tests/lib/cortex/boundary/audit.test.ts
-import { describe, it, expect, beforeEach, afterEach } from 'vitest';
-import fs from 'fs';
-import path from 'path';
-import os from 'os';
-import Database from 'better-sqlite3';
-import { AuditLog } from '@/lib/cortex/boundary/audit';
-
-describe('AuditLog', () => {
-  let tmpDir: string;
-  let audit: AuditLog;
-
-  beforeEach(() => {
-    tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'cortex-audit-'));
-    const db = new Database(path.join(tmpDir, 'graph.db'));
-    audit = new AuditLog(db);
-  });
-
-  afterEach(() => {
-    fs.rmSync(tmpDir, { recursive: true, force: true });
-  });
-
-  it('logs an access decision', () => {
-    audit.log({
-      requesterId: 'person-alice',
-      knowledgeId: 'k1',
-      action: 'allowed',
-      reason: 'public sensitivity',
-    });
-    const entries = audit.query({ requesterId: 'person-alice' });
-    expect(entries).toHaveLength(1);
-    expect(entries[0].action).toBe('allowed');
-  });
-
-  it('logs denied access', () => {
-    audit.log({
-      requesterId: 'person-alice',
-      knowledgeId: 'k2',
-      action: 'denied',
-      reason: 'confidential, not creator',
-    });
-    const entries = audit.query({ requesterId: 'person-alice' });
-    expect(entries).toHaveLength(1);
-    expect(entries[0].action).toBe('denied');
-  });
-
-  it('queries by time range', () => {
-    audit.log({ requesterId: 'person-alice', knowledgeId: 'k1', action: 'allowed', reason: 'public' });
-    const recent = audit.query({ since: new Date(Date.now() - 60000).toISOString() });
-    expect(recent).toHaveLength(1);
-
-    const future = audit.query({ since: new Date(Date.now() + 60000).toISOString() });
-    expect(future).toHaveLength(0);
-  });
-
-  it('supports retention cleanup', () => {
-    audit.log({ requesterId: 'person-alice', knowledgeId: 'k1', action: 'allowed', reason: 'test' });
-    // Cleanup entries older than 0 days (everything)
-    audit.cleanup(0);
-    const entries = audit.query({});
-    expect(entries).toHaveLength(0);
-  });
-});
-```
-
-- [ ] **Step 2: Implement audit log**
-
-```typescript
-// src/lib/cortex/boundary/audit.ts
-import type Database from 'better-sqlite3';
-
-export interface AuditEntry {
-  requesterId: string;
-  knowledgeId: string;
-  action: 'allowed' | 'denied';
-  reason: string;
-}
-
-export interface AuditQueryFilter {
-  requesterId?: string;
-  since?: string;  // ISO timestamp
-  limit?: number;
-}
-
-export interface AuditRecord extends AuditEntry {
-  timestamp: string;
-}
-
-export class AuditLog {
-  private db: InstanceType<typeof Database>;
-
-  constructor(db: InstanceType<typeof Database>) {
-    this.db = db;
-    this.db.exec(`
-      CREATE TABLE IF NOT EXISTS audit_log (
-        id INTEGER PRIMARY KEY AUTOINCREMENT,
-        requester_id TEXT NOT NULL,
-        knowledge_id TEXT NOT NULL,
-        action TEXT NOT NULL,
-        reason TEXT NOT NULL,
-        timestamp TEXT NOT NULL
-      );
-      CREATE INDEX IF NOT EXISTS idx_audit_requester ON audit_log(requester_id);
-      CREATE INDEX IF NOT EXISTS idx_audit_timestamp ON audit_log(timestamp);
-    `);
-  }
-
-  log(entry: AuditEntry): void {
-    this.db.prepare(`
-      INSERT INTO audit_log (requester_id, knowledge_id, action, reason, timestamp)
-      VALUES (?, ?, ?, ?, ?)
-    `).run(entry.requesterId, entry.knowledgeId, entry.action, entry.reason, new Date().toISOString());
-  }
-
-  query(filter: AuditQueryFilter): AuditRecord[] {
-    let sql = 'SELECT * FROM audit_log WHERE 1=1';
-    const params: any[] = [];
-
-    if (filter.requesterId) {
-      sql += ' AND requester_id = ?';
-      params.push(filter.requesterId);
-    }
-    if (filter.since) {
-      sql += ' AND timestamp >= ?';
-      params.push(filter.since);
-    }
-
-    sql += ' ORDER BY timestamp DESC';
-
-    if (filter.limit) {
-      sql += ' LIMIT ?';
-      params.push(filter.limit);
-    }
-
-    return (this.db.prepare(sql).all(...params) as any[]).map(row => ({
-      requesterId: row.requester_id,
-      knowledgeId: row.knowledge_id,
-      action: row.action,
-      reason: row.reason,
-      timestamp: row.timestamp,
-    }));
-  }
-
-  cleanup(retentionDays: number): void {
-    const cutoff = new Date(Date.now() - retentionDays * 86400000).toISOString();
-    this.db.prepare('DELETE FROM audit_log WHERE timestamp < ?').run(cutoff);
-  }
-}
-```
-
-- [ ] **Step 3: Run tests, commit**
-
-```bash
-git commit -m "feat(cortex): add audit trail for access decisions"
-```
-
----
-
-## Chunk 3: Integration
-
-### Task 5: Integrate with ingestion pipeline
-
-**Files:**
-- Modify: `src/lib/cortex/ingestion/pipeline.ts`
-
-- [ ] **Step 1: Read pipeline.ts**
-
-- [ ] **Step 2: Add auto-classification during ingestion**
-
-Import the classifier:
-```typescript
-import { classifySensitivity } from '../boundary/classifier';
-```
-
-In the KnowledgeUnit construction (where v2 fields are set), replace the hardcoded `sensitivity: 'internal'` with:
-```typescript
-sensitivity: classifySensitivity(chunk.text),
-```
-
-This ensures every ingested knowledge unit gets an auto-classified sensitivity level.
-
-- [ ] **Step 3: Run pipeline tests to verify no regressions**
-
-```bash
-npx vitest run tests/lib/cortex/ingestion/
-```
-
-- [ ] **Step 4: Commit**
-
-```bash
-git commit -m "feat(cortex): auto-classify sensitivity during ingestion"
-```
-
----
-
-### Task 6: Integrate AccessFilter with ContextEngine
-
-**Files:**
-- Modify: `src/lib/cortex/retrieval/context-engine.ts`
-
-- [ ] **Step 1: Read context-engine.ts**
-
-- [ ] **Step 2: Add AccessFilter to the pipeline**
-
-Import AccessFilter:
-```typescript
-import { AccessFilter } from '../boundary/access';
-```
-
-Add to `ContextEngineDeps`:
-```typescript
-accessFilter?: AccessFilter;
-```
-
-In the `assemble()` method, after fusion/ranking (Stage 5) but before conflict detection (Stage 6), apply the access filter:
-
-```typescript
-// Stage 5.5: Access control filtering
-let accessible = fused;
-if (this.deps.accessFilter) {
-  accessible = this.deps.accessFilter.filterResults(fused);
-}
-```
-
-Then pass `accessible` (not `fused`) to conflict detection and formatting.
-
-- [ ] **Step 3: Run context-engine tests to verify no regressions**
-
-```bash
-npx vitest run tests/lib/cortex/retrieval/context-engine.test.ts
-```
-
-- [ ] **Step 4: Commit**
-
-```bash
-git commit -m "feat(cortex): integrate access filter into context assembly pipeline"
-```
-
----
-
-### Task 7: Barrel export and config integration
-
-**Files:**
-- Create: `src/lib/cortex/boundary/index.ts`
-- Modify: `src/lib/cortex/config.ts` (add policies array to config)
-
-- [ ] **Step 1: Create barrel export**
-
-```typescript
-// src/lib/cortex/boundary/index.ts
-export { classifySensitivity } from './classifier';
-export { PolicyEngine } from './policy';
-export type { Policy, PolicyAction, PropagationTarget } from './policy';
-export { AccessFilter } from './access';
-export type { AccessFilterConfig } from './access';
-export { AuditLog } from './audit';
-export type { AuditEntry, AuditRecord, AuditQueryFilter } from './audit';
-```
-
-- [ ] **Step 2: Add policies to CortexConfig**
-
-In `src/lib/cortex/config.ts`, add to the `CortexConfig` interface:
-```typescript
-policies?: Policy[];  // Organizational boundary policies
-```
-
-And in `DEFAULT_CORTEX_CONFIG`:
-```typescript
-policies: [],
-```
-
-Import the Policy type:
-```typescript
-import type { Policy } from './boundary/policy';
-```
-
-- [ ] **Step 3: Run full test suite**
-
-```bash
-npx vitest run tests/lib/cortex/
-```
-
-- [ ] **Step 4: Commit**
-
-```bash
-git commit -m "feat(cortex): add boundary module barrel export and config integration"
-```
-
----
-
-## Summary
-
-| Task | Component | Tests | Status |
-|------|-----------|-------|--------|
-| 1 | Sensitivity classifier | 7 | |
-| 2 | Policy engine | 6 | |
-| 3 | Access filter | 8 | |
-| 4 | Audit trail | 4 | |
-| 5 | Pipeline integration | regression | |
-| 6 | ContextEngine integration | regression | |
-| 7 | Barrel export + config | regression | |
-
-**Total: 7 tasks, ~25 new tests, 3 chunks**
-
-**Key design decisions:**
-- Auto-classification is regex-based (no LLM call) — fast, deterministic, auditable
-- Policies are stored in CortexConfig (admin-editable JSON) — not code
-- AccessFilter is a separate class that can be composed with ContextEngine or used standalone
-- Audit log uses the entity graph's SQLite DB (same file, new table) — no new database
-- Most restrictive classification always wins (confidential > restricted > internal > public)
+# Cortex v2 — Pillar 4: Boundary Engine
+
+> **For agentic workers:** REQUIRED: Use superpowers:subagent-driven-development (if subagents available) or superpowers:executing-plans to implement this plan. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Add three-layer access control to Cortex: auto-classification of sensitivity, organizational policies, and creator overrides — ensuring confidential knowledge never leaks across scope boundaries.
+
+**Architecture:** A new `src/lib/cortex/boundary/` module with three components: a `Classifier` (regex-based sensitivity detection), a `PolicyEngine` (rule evaluation), and an `AccessFilter` (query-time enforcement). The AccessFilter integrates with the ContextEngine to pre-filter scopes before search. An audit log records all access decisions in the entity graph's SQLite database.
+
+**Tech Stack:** TypeScript, better-sqlite3 (existing graph DB), vitest
+
+**Spec:** `docs/superpowers/specs/2026-03-14-cortex-v2-design.md` — Pillar 4
+
+**Depends on:** Pillar 1 (Entity Graph) + Pillar 2 (Knowledge Unit Evolution) — both completed
+
+---
+
+## File Structure
+
+```
+New files:
+├── src/lib/cortex/boundary/classifier.ts    — Auto-classification (Layer 1)
+├── src/lib/cortex/boundary/policy.ts        — Policy engine (Layer 2)
+├── src/lib/cortex/boundary/access.ts        — Query-time access filter (Layer 3 + enforcement)
+├── src/lib/cortex/boundary/audit.ts         — Audit trail logging
+├── src/lib/cortex/boundary/index.ts         — Barrel export
+
+Modified files:
+├── src/lib/cortex/ingestion/pipeline.ts     — Auto-classify on ingestion
+├── src/lib/cortex/retrieval/context-engine.ts — Pre-filter with AccessFilter
+├── src/lib/cortex/config.ts                 — Add policies to config
+
+Test files:
+├── tests/lib/cortex/boundary/classifier.test.ts
+├── tests/lib/cortex/boundary/policy.test.ts
+├── tests/lib/cortex/boundary/access.test.ts
+├── tests/lib/cortex/boundary/audit.test.ts
+```
+
+---
+
+## Chunk 1: Auto-Classification and Policy Engine
+
+### Task 1: Sensitivity auto-classifier
+
+**Files:**
+- Create: `src/lib/cortex/boundary/classifier.ts`
+- Create: `tests/lib/cortex/boundary/classifier.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/boundary/classifier.test.ts
+import { describe, it, expect } from 'vitest';
+import { classifySensitivity } from '@/lib/cortex/boundary/classifier';
+
+describe('classifySensitivity', () => {
+  it('classifies secrets as confidential', () => {
+    expect(classifySensitivity('Set API_KEY=sk-ant-abc123 in .env')).toBe('confidential');
+    expect(classifySensitivity('password: hunter2')).toBe('confidential');
+    expect(classifySensitivity('DATABASE_URL=postgres://user:pass@host')).toBe('confidential');
+  });
+
+  it('classifies personnel content as confidential', () => {
+    expect(classifySensitivity('Alice performance review: exceeds expectations')).toBe('confidential');
+    expect(classifySensitivity('salary adjustment from 120k to 140k')).toBe('confidential');
+  });
+
+  it('classifies security content as restricted', () => {
+    expect(classifySensitivity('Found SQL injection vulnerability in login endpoint')).toBe('restricted');
+    expect(classifySensitivity('CVE-2024-1234 affects our auth library')).toBe('restricted');
+  });
+
+  it('classifies business content as restricted', () => {
+    expect(classifySensitivity('Q3 revenue was $2.5M, below target')).toBe('restricted');
+    expect(classifySensitivity('Unreleased product launch planned for April')).toBe('restricted');
+  });
+
+  it('classifies technical content as internal', () => {
+    expect(classifySensitivity('We decided to use PostgreSQL for the new service')).toBe('internal');
+    expect(classifySensitivity('The auth middleware handles JWT validation')).toBe('internal');
+  });
+
+  it('classifies general content as public', () => {
+    expect(classifySensitivity('How to use git rebase')).toBe('public');
+    expect(classifySensitivity('JavaScript array methods')).toBe('public');
+  });
+
+  it('returns most restrictive when multiple detectors match', () => {
+    // Has both secret (confidential) and security (restricted) signals
+    expect(classifySensitivity('API_KEY leaked in CVE-2024-1234')).toBe('confidential');
+  });
+});
+```
+
+- [ ] **Step 2: Implement classifier**
+
+```typescript
+// src/lib/cortex/boundary/classifier.ts
+import type { SensitivityClass } from '../knowledge/types';
+
+interface Detector {
+  sensitivity: SensitivityClass;
+  patterns: RegExp[];
+  priority: number;  // higher = more restrictive
+}
+
+const DETECTORS: Detector[] = [
+  {
+    sensitivity: 'confidential',
+    priority: 4,
+    patterns: [
+      // Secrets
+      /\b(api[_-]?key|secret[_-]?key|access[_-]?token|auth[_-]?token)\s*[=:]/i,
+      /\b(password|passwd|pwd)\s*[=:]/i,
+      /\bsk-[a-z]{2,4}-[a-zA-Z0-9]{10,}/,  // Anthropic/OpenAI key format
+      /\b(DATABASE_URL|REDIS_URL|MONGO_URI)\s*=/i,
+      /\bpostgres:\/\/\w+:\w+@/i,
+      /\b(private[_-]?key|ssh[_-]?key)\b/i,
+      // Personnel
+      /\b(performance\s+review|annual\s+review)\b/i,
+      /\b(salary|compensation|pay\s+raise|pay\s+cut)\b/i,
+      /\b(hiring|termination|fired|let\s+go)\b/i,
+      /\b(1:1\s+notes?|one[\s-]on[\s-]one)\b/i,
+    ],
+  },
+  {
+    sensitivity: 'restricted',
+    priority: 3,
+    patterns: [
+      // Security
+      /\b(vulnerab|exploit|attack\s+vector|injection|xss|csrf|ssrf)\b/i,
+      /\bCVE-\d{4}-\d+/i,
+      /\b(incident\s+report|security\s+breach|data\s+leak)\b/i,
+      // Business
+      /\b(revenue|profit|loss|earnings|ARR|MRR)\b/i,
+      /\b(unreleased|pre-launch|confidential\s+plan|roadmap)\b/i,
+      /\bcustomer\s+(data|records|PII)\b/i,
+    ],
+  },
+  {
+    sensitivity: 'internal',
+    priority: 2,
+    patterns: [
+      /\b(we\s+decided|architecture|design\s+pattern|refactor)\b/i,
+      /\b(middleware|service|endpoint|database|schema)\b/i,
+      /\b(deployment|CI\/CD|pipeline|infrastructure)\b/i,
+      /\b(bug\s+fix|pull\s+request|code\s+review)\b/i,
+    ],
+  },
+];
+
+export function classifySensitivity(text: string): SensitivityClass {
+  let highestPriority = 0;
+  let result: SensitivityClass = 'public';
+
+  for (const detector of DETECTORS) {
+    if (detector.priority <= highestPriority) continue;
+    for (const pattern of detector.patterns) {
+      if (pattern.test(text)) {
+        highestPriority = detector.priority;
+        result = detector.sensitivity;
+        break;
+      }
+    }
+  }
+
+  return result;
+}
+```
+
+- [ ] **Step 3: Run tests, commit**
+
+```bash
+git commit -m "feat(cortex): add sensitivity auto-classifier for boundary engine"
+```
+
+---
+
+### Task 2: Policy engine
+
+**Files:**
+- Create: `src/lib/cortex/boundary/policy.ts`
+- Create: `tests/lib/cortex/boundary/policy.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/boundary/policy.test.ts
+import { describe, it, expect } from 'vitest';
+import { PolicyEngine } from '@/lib/cortex/boundary/policy';
+import type { Policy } from '@/lib/cortex/boundary/policy';
+
+describe('PolicyEngine', () => {
+  it('returns empty actions when no policies match', () => {
+    const engine = new PolicyEngine([]);
+    const actions = engine.evaluate({ type: 'decision', sensitivity: 'internal' });
+    expect(actions).toEqual([]);
+  });
+
+  it('matches by knowledge type', () => {
+    const policies: Policy[] = [{
+      name: 'arch-decisions-propagate',
+      match: { type: 'decision' },
+      action: { propagate_to: [{ level: 'department' }] },
+    }];
+    const engine = new PolicyEngine(policies);
+    const actions = engine.evaluate({ type: 'decision', sensitivity: 'internal' });
+    expect(actions).toHaveLength(1);
+    expect(actions[0].propagate_to).toEqual([{ level: 'department' }]);
+  });
+
+  it('matches by sensitivity', () => {
+    const policies: Policy[] = [{
+      name: 'lock-confidential',
+      match: { sensitivity: 'confidential' },
+      action: { cannot_propagate: true },
+    }];
+    const engine = new PolicyEngine(policies);
+    const actions = engine.evaluate({ type: 'pattern', sensitivity: 'confidential' });
+    expect(actions).toHaveLength(1);
+    expect(actions[0].cannot_propagate).toBe(true);
+  });
+
+  it('matches by topics', () => {
+    const policies: Policy[] = [{
+      name: 'security-routing',
+      match: { topics: ['security', 'vulnerability'] },
+      action: { max_scope: 'department', propagate_to: [{ level: 'team', entity_id: 'team-security' }] },
+    }];
+    const engine = new PolicyEngine(policies);
+    const actions = engine.evaluate({ type: 'error_fix', sensitivity: 'restricted', topics: ['security'] });
+    expect(actions).toHaveLength(1);
+    expect(actions[0].max_scope).toBe('department');
+  });
+
+  it('returns multiple matching policies', () => {
+    const policies: Policy[] = [
+      { name: 'p1', match: { type: 'decision' }, action: { trickle_down: true } },
+      { name: 'p2', match: { sensitivity: 'internal' }, action: { max_scope: 'organization' } },
+    ];
+    const engine = new PolicyEngine(policies);
+    const actions = engine.evaluate({ type: 'decision', sensitivity: 'internal' });
+    expect(actions).toHaveLength(2);
+  });
+
+  it('does not match when criteria do not overlap', () => {
+    const policies: Policy[] = [{
+      name: 'security-only',
+      match: { topics: ['security'] },
+      action: { cannot_propagate: true },
+    }];
+    const engine = new PolicyEngine(policies);
+    const actions = engine.evaluate({ type: 'decision', sensitivity: 'internal', topics: ['architecture'] });
+    expect(actions).toHaveLength(0);
+  });
+});
+```
+
+- [ ] **Step 2: Implement policy engine**
+
+```typescript
+// src/lib/cortex/boundary/policy.ts
+import type { KnowledgeType, SensitivityClass, ScopeLevel } from '../knowledge/types';
+
+export interface PropagationTarget {
+  level: ScopeLevel;
+  entity_id?: string;
+}
+
+export interface PolicyAction {
+  max_scope?: ScopeLevel;
+  propagate_to?: PropagationTarget[];
+  trickle_down?: boolean;
+  cannot_propagate?: boolean;
+}
+
+export interface Policy {
+  name: string;
+  match: {
+    type?: KnowledgeType;
+    topics?: string[];
+    sensitivity?: SensitivityClass;
+    scope_level?: ScopeLevel;
+  };
+  action: PolicyAction;
+}
+
+export interface PolicyMatchInput {
+  type: KnowledgeType | string;
+  sensitivity: SensitivityClass | string;
+  topics?: string[];
+  scope_level?: ScopeLevel | string;
+}
+
+export class PolicyEngine {
+  constructor(private policies: Policy[]) {}
+
+  evaluate(input: PolicyMatchInput): PolicyAction[] {
+    const matched: PolicyAction[] = [];
+
+    for (const policy of this.policies) {
+      if (this.matches(policy, input)) {
+        matched.push(policy.action);
+      }
+    }
+
+    return matched;
+  }
+
+  private matches(policy: Policy, input: PolicyMatchInput): boolean {
+    const { match } = policy;
+
+    if (match.type && match.type !== input.type) return false;
+    if (match.sensitivity && match.sensitivity !== input.sensitivity) return false;
+    if (match.scope_level && match.scope_level !== input.scope_level) return false;
+
+    if (match.topics && match.topics.length > 0) {
+      const inputTopics = input.topics ?? [];
+      const hasOverlap = match.topics.some(t => inputTopics.includes(t));
+      if (!hasOverlap) return false;
+    }
+
+    return true;
+  }
+}
+```
+
+- [ ] **Step 3: Run tests, commit**
+
+```bash
+git commit -m "feat(cortex): add policy engine for boundary enforcement"
+```
+
+---
+
+## Chunk 2: Access Filter and Audit Trail
+
+### Task 3: Query-time access filter
+
+**Files:**
+- Create: `src/lib/cortex/boundary/access.ts`
+- Create: `tests/lib/cortex/boundary/access.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/boundary/access.test.ts
+import { describe, it, expect } from 'vitest';
+import { AccessFilter } from '@/lib/cortex/boundary/access';
+import type { ScoredKnowledge } from '@/lib/cortex/knowledge/types';
+
+function makeUnit(overrides: Partial<ScoredKnowledge> = {}): ScoredKnowledge {
+  return {
+    id: 'k1', vector: [], text: 'test', type: 'decision', layer: 'personal',
+    workspace_id: null, session_id: null, agent_type: 'claude',
+    project_path: null, file_refs: [], confidence: 0.8,
+    created: new Date().toISOString(), source_timestamp: new Date().toISOString(),
+    stale_score: 0, access_count: 0, last_accessed: null, metadata: {},
+    relevance_score: 0.9, similarity: 0.9,
+    sensitivity: 'internal',
+    scope: { level: 'personal', entity_id: 'person-alice' },
+    origin: { source_type: 'conversation', source_ref: '', creator_entity_id: 'person-alice' },
+    ...overrides,
+  };
+}
+
+describe('AccessFilter', () => {
+  const filter = new AccessFilter({
+    requesterId: 'person-alice',
+    requesterScope: { level: 'team', entity_id: 'team-platform' },
+    requesterOrg: 'organization-acme',
+  });
+
+  it('allows public knowledge from anywhere in org', () => {
+    const unit = makeUnit({ sensitivity: 'public', scope: { level: 'organization', entity_id: 'org-acme' } });
+    expect(filter.canAccess(unit)).toBe(true);
+  });
+
+  it('allows internal knowledge within org', () => {
+    const unit = makeUnit({ sensitivity: 'internal', scope: { level: 'team', entity_id: 'team-other' } });
+    expect(filter.canAccess(unit)).toBe(true);
+  });
+
+  it('allows restricted knowledge within same scope', () => {
+    const unit = makeUnit({
+      sensitivity: 'restricted',
+      scope: { level: 'team', entity_id: 'team-platform' },
+    });
+    expect(filter.canAccess(unit)).toBe(true);
+  });
+
+  it('denies restricted knowledge from different department', () => {
+    const unit = makeUnit({
+      sensitivity: 'restricted',
+      scope: { level: 'department', entity_id: 'department-sales' },
+    });
+    // Alice is in team-platform, not department-sales
+    expect(filter.canAccess(unit)).toBe(false);
+  });
+
+  it('allows confidential knowledge from self', () => {
+    const unit = makeUnit({
+      sensitivity: 'confidential',
+      scope: { level: 'personal', entity_id: 'person-alice' },
+      origin: { source_type: 'conversation', source_ref: '', creator_entity_id: 'person-alice' },
+    });
+    expect(filter.canAccess(unit)).toBe(true);
+  });
+
+  it('denies confidential knowledge from others', () => {
+    const unit = makeUnit({
+      sensitivity: 'confidential',
+      scope: { level: 'personal', entity_id: 'person-bob' },
+      origin: { source_type: 'conversation', source_ref: '', creator_entity_id: 'person-bob' },
+    });
+    expect(filter.canAccess(unit)).toBe(false);
+  });
+
+  it('respects creator_scope override (further restriction)', () => {
+    const unit = makeUnit({
+      sensitivity: 'internal',  // normally accessible org-wide
+      creator_scope: { max_level: 'personal' },  // but creator restricted to personal only
+      scope: { level: 'personal', entity_id: 'person-bob' },
+    });
+    // Alice can't access Bob's personal-restricted knowledge even though it's internal
+    expect(filter.canAccess(unit)).toBe(false);
+  });
+
+  it('filters a list of results', () => {
+    const results = [
+      makeUnit({ id: 'a', sensitivity: 'public' }),
+      makeUnit({ id: 'b', sensitivity: 'confidential', origin: { source_type: 'conversation', source_ref: '', creator_entity_id: 'person-bob' } }),
+      makeUnit({ id: 'c', sensitivity: 'internal' }),
+    ];
+    const filtered = filter.filterResults(results);
+    expect(filtered).toHaveLength(2);  // a and c pass, b denied
+    expect(filtered.map(r => r.id)).toEqual(['a', 'c']);
+  });
+});
+```
+
+- [ ] **Step 2: Implement access filter**
+
+```typescript
+// src/lib/cortex/boundary/access.ts
+import type { ScoredKnowledge, Scope, ScopeLevel, SensitivityClass } from '../knowledge/types';
+
+const SCOPE_HIERARCHY: Record<ScopeLevel, number> = {
+  personal: 0,
+  team: 1,
+  department: 2,
+  organization: 3,
+};
+
+export interface AccessFilterConfig {
+  requesterId: string;
+  requesterScope: Scope;        // requester's team scope
+  requesterOrg: string;         // requester's org entity_id
+  grants?: Set<string>;         // knowledge IDs the requester has been granted access to
+}
+
+export class AccessFilter {
+  private config: AccessFilterConfig;
+
+  constructor(config: AccessFilterConfig) {
+    this.config = config;
+  }
+
+  canAccess(unit: ScoredKnowledge): boolean {
+    const sensitivity = (unit.sensitivity ?? 'internal') as SensitivityClass;
+    const unitScope = unit.scope ?? { level: 'personal' as ScopeLevel, entity_id: '' };
+    const creatorId = unit.origin?.creator_entity_id ?? '';
+
+    // Check creator_scope override first — most restrictive wins
+    if (unit.creator_scope) {
+      const maxLevel = SCOPE_HIERARCHY[unit.creator_scope.max_level] ?? 0;
+      const requesterLevel = this.getRequesterProximityLevel(unitScope);
+      if (requesterLevel > maxLevel) return false;
+    }
+
+    switch (sensitivity) {
+      case 'public':
+        return true;
+
+      case 'internal':
+        // Accessible to anyone in the same org
+        return true;
+
+      case 'restricted':
+        // Accessible within same scope or if policy grants cross-scope access
+        return this.isWithinScope(unitScope);
+
+      case 'confidential':
+        // Only creator or explicit grant
+        if (this.config.requesterId === creatorId) return true;
+        if (this.config.grants?.has(unit.id)) return true;
+        return false;
+
+      default:
+        return false;
+    }
+  }
+
+  filterResults(results: ScoredKnowledge[]): ScoredKnowledge[] {
+    return results.filter(r => this.canAccess(r));
+  }
+
+  private isWithinScope(unitScope: Scope): boolean {
+    // Same entity_id = same scope
+    if (unitScope.entity_id === this.config.requesterScope.entity_id) return true;
+    if (unitScope.entity_id === this.config.requesterId) return true;
+    // Same org for org-level scopes
+    if (unitScope.level === 'organization' && unitScope.entity_id === this.config.requesterOrg) return true;
+    return false;
+  }
+
+  private getRequesterProximityLevel(unitScope: Scope): number {
+    // How "far" is the requester from the unit's scope?
+    // 0 = same entity, 1 = same scope level, 2+ = further
+    if (unitScope.entity_id === this.config.requesterId) return 0;
+    if (unitScope.entity_id === this.config.requesterScope.entity_id) return 0;
+    return SCOPE_HIERARCHY[unitScope.level] ?? 3;
+  }
+}
+```
+
+- [ ] **Step 3: Run tests, commit**
+
+```bash
+git commit -m "feat(cortex): add query-time access filter for boundary enforcement"
+```
+
+---
+
+### Task 4: Audit trail
+
+**Files:**
+- Create: `src/lib/cortex/boundary/audit.ts`
+- Create: `tests/lib/cortex/boundary/audit.test.ts`
+
+- [ ] **Step 1: Write failing tests**
+
+```typescript
+// tests/lib/cortex/boundary/audit.test.ts
+import { describe, it, expect, beforeEach, afterEach } from 'vitest';
+import fs from 'fs';
+import path from 'path';
+import os from 'os';
+import Database from 'better-sqlite3';
+import { AuditLog } from '@/lib/cortex/boundary/audit';
+
+describe('AuditLog', () => {
+  let tmpDir: string;
+  let audit: AuditLog;
+
+  beforeEach(() => {
+    tmpDir = fs.mkdtempSync(path.join(os.tmpdir(), 'cortex-audit-'));
+    const db = new Database(path.join(tmpDir, 'graph.db'));
+    audit = new AuditLog(db);
+  });
+
+  afterEach(() => {
+    fs.rmSync(tmpDir, { recursive: true, force: true });
+  });
+
+  it('logs an access decision', () => {
+    audit.log({
+      requesterId: 'person-alice',
+      knowledgeId: 'k1',
+      action: 'allowed',
+      reason: 'public sensitivity',
+    });
+    const entries = audit.query({ requesterId: 'person-alice' });
+    expect(entries).toHaveLength(1);
+    expect(entries[0].action).toBe('allowed');
+  });
+
+  it('logs denied access', () => {
+    audit.log({
+      requesterId: 'person-alice',
+      knowledgeId: 'k2',
+      action: 'denied',
+      reason: 'confidential, not creator',
+    });
+    const entries = audit.query({ requesterId: 'person-alice' });
+    expect(entries).toHaveLength(1);
+    expect(entries[0].action).toBe('denied');
+  });
+
+  it('queries by time range', () => {
+    audit.log({ requesterId: 'person-alice', knowledgeId: 'k1', action: 'allowed', reason: 'public' });
+    const recent = audit.query({ since: new Date(Date.now() - 60000).toISOString() });
+    expect(recent).toHaveLength(1);
+
+    const future = audit.query({ since: new Date(Date.now() + 60000).toISOString() });
+    expect(future).toHaveLength(0);
+  });
+
+  it('supports retention cleanup', () => {
+    audit.log({ requesterId: 'person-alice', knowledgeId: 'k1', action: 'allowed', reason: 'test' });
+    // Cleanup entries older than 0 days (everything)
+    audit.cleanup(0);
+    const entries = audit.query({});
+    expect(entries).toHaveLength(0);
+  });
+});
+```
+
+- [ ] **Step 2: Implement audit log**
+
+```typescript
+// src/lib/cortex/boundary/audit.ts
+import type Database from 'better-sqlite3';
+
+export interface AuditEntry {
+  requesterId: string;
+  knowledgeId: string;
+  action: 'allowed' | 'denied';
+  reason: string;
+}
+
+export interface AuditQueryFilter {
+  requesterId?: string;
+  since?: string;  // ISO timestamp
+  limit?: number;
+}
+
+export interface AuditRecord extends AuditEntry {
+  timestamp: string;
+}
+
+export class AuditLog {
+  private db: InstanceType<typeof Database>;
+
+  constructor(db: InstanceType<typeof Database>) {
+    this.db = db;
+    this.db.exec(`
+      CREATE TABLE IF NOT EXISTS audit_log (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        requester_id TEXT NOT NULL,
+        knowledge_id TEXT NOT NULL,
+        action TEXT NOT NULL,
+        reason TEXT NOT NULL,
+        timestamp TEXT NOT NULL
+      );
+      CREATE INDEX IF NOT EXISTS idx_audit_requester ON audit_log(requester_id);
+      CREATE INDEX IF NOT EXISTS idx_audit_timestamp ON audit_log(timestamp);
+    `);
+  }
+
+  log(entry: AuditEntry): void {
+    this.db.prepare(`
+      INSERT INTO audit_log (requester_id, knowledge_id, action, reason, timestamp)
+      VALUES (?, ?, ?, ?, ?)
+    `).run(entry.requesterId, entry.knowledgeId, entry.action, entry.reason, new Date().toISOString());
+  }
+
+  query(filter: AuditQueryFilter): AuditRecord[] {
+    let sql = 'SELECT * FROM audit_log WHERE 1=1';
+    const params: any[] = [];
+
+    if (filter.requesterId) {
+      sql += ' AND requester_id = ?';
+      params.push(filter.requesterId);
+    }
+    if (filter.since) {
+      sql += ' AND timestamp >= ?';
+      params.push(filter.since);
+    }
+
+    sql += ' ORDER BY timestamp DESC';
+
+    if (filter.limit) {
+      sql += ' LIMIT ?';
+      params.push(filter.limit);
+    }
+
+    return (this.db.prepare(sql).all(...params) as any[]).map(row => ({
+      requesterId: row.requester_id,
+      knowledgeId: row.knowledge_id,
+      action: row.action,
+      reason: row.reason,
+      timestamp: row.timestamp,
+    }));
+  }
+
+  cleanup(retentionDays: number): void {
+    const cutoff = new Date(Date.now() - retentionDays * 86400000).toISOString();
+    this.db.prepare('DELETE FROM audit_log WHERE timestamp < ?').run(cutoff);
+  }
+}
+```
+
+- [ ] **Step 3: Run tests, commit**
+
+```bash
+git commit -m "feat(cortex): add audit trail for access decisions"
+```
+
+---
+
+## Chunk 3: Integration
+
+### Task 5: Integrate with ingestion pipeline
+
+**Files:**
+- Modify: `src/lib/cortex/ingestion/pipeline.ts`
+
+- [ ] **Step 1: Read pipeline.ts**
+
+- [ ] **Step 2: Add auto-classification during ingestion**
+
+Import the classifier:
+```typescript
+import { classifySensitivity } from '../boundary/classifier';
+```
+
+In the KnowledgeUnit construction (where v2 fields are set), replace the hardcoded `sensitivity: 'internal'` with:
+```typescript
+sensitivity: classifySensitivity(chunk.text),
+```
+
+This ensures every ingested knowledge unit gets an auto-classified sensitivity level.
+
+- [ ] **Step 3: Run pipeline tests to verify no regressions**
+
+```bash
+npx vitest run tests/lib/cortex/ingestion/
+```
+
+- [ ] **Step 4: Commit**
+
+```bash
+git commit -m "feat(cortex): auto-classify sensitivity during ingestion"
+```
+
+---
+
+### Task 6: Integrate AccessFilter with ContextEngine
+
+**Files:**
+- Modify: `src/lib/cortex/retrieval/context-engine.ts`
+
+- [ ] **Step 1: Read context-engine.ts**
+
+- [ ] **Step 2: Add AccessFilter to the pipeline**
+
+Import AccessFilter:
+```typescript
+import { AccessFilter } from '../boundary/access';
+```
+
+Add to `ContextEngineDeps`:
+```typescript
+accessFilter?: AccessFilter;
+```
+
+In the `assemble()` method, after fusion/ranking (Stage 5) but before conflict detection (Stage 6), apply the access filter:
+
+```typescript
+// Stage 5.5: Access control filtering
+let accessible = fused;
+if (this.deps.accessFilter) {
+  accessible = this.deps.accessFilter.filterResults(fused);
+}
+```
+
+Then pass `accessible` (not `fused`) to conflict detection and formatting.
+
+- [ ] **Step 3: Run context-engine tests to verify no regressions**
+
+```bash
+npx vitest run tests/lib/cortex/retrieval/context-engine.test.ts
+```
+
+- [ ] **Step 4: Commit**
+
+```bash
+git commit -m "feat(cortex): integrate access filter into context assembly pipeline"
+```
+
+---
+
+### Task 7: Barrel export and config integration
+
+**Files:**
+- Create: `src/lib/cortex/boundary/index.ts`
+- Modify: `src/lib/cortex/config.ts` (add policies array to config)
+
+- [ ] **Step 1: Create barrel export**
+
+```typescript
+// src/lib/cortex/boundary/index.ts
+export { classifySensitivity } from './classifier';
+export { PolicyEngine } from './policy';
+export type { Policy, PolicyAction, PropagationTarget } from './policy';
+export { AccessFilter } from './access';
+export type { AccessFilterConfig } from './access';
+export { AuditLog } from './audit';
+export type { AuditEntry, AuditRecord, AuditQueryFilter } from './audit';
+```
+
+- [ ] **Step 2: Add policies to CortexConfig**
+
+In `src/lib/cortex/config.ts`, add to the `CortexConfig` interface:
+```typescript
+policies?: Policy[];  // Organizational boundary policies
+```
+
+And in `DEFAULT_CORTEX_CONFIG`:
+```typescript
+policies: [],
+```
+
+Import the Policy type:
+```typescript
+import type { Policy } from './boundary/policy';
+```
+
+- [ ] **Step 3: Run full test suite**
+
+```bash
+npx vitest run tests/lib/cortex/
+```
+
+- [ ] **Step 4: Commit**
+
+```bash
+git commit -m "feat(cortex): add boundary module barrel export and config integration"
+```
+
+---
+
+## Summary
+
+| Task | Component | Tests | Status |
+|------|-----------|-------|--------|
+| 1 | Sensitivity classifier | 7 | |
+| 2 | Policy engine | 6 | |
+| 3 | Access filter | 8 | |
+| 4 | Audit trail | 4 | |
+| 5 | Pipeline integration | regression | |
+| 6 | ContextEngine integration | regression | |
+| 7 | Barrel export + config | regression | |
+
+**Total: 7 tasks, ~25 new tests, 3 chunks**
+
+**Key design decisions:**
+- Auto-classification is regex-based (no LLM call) — fast, deterministic, auditable
+- Policies are stored in CortexConfig (admin-editable JSON) — not code
+- AccessFilter is a separate class that can be composed with ContextEngine or used standalone
+- Audit log uses the entity graph's SQLite DB (same file, new table) — no new database
+- Most restrictive classification always wins (confidential > restricted > internal > public)