User Task
  │
  ▼
┌──────────────────────────────────────────────────────────────┐
│ STEP 0: PONYTAIL LADDER                                       │
│ stdlib → native → dep → shortest                              │
│ 0 tokens · <100ms · $0.0000                                   │
│ 57% of tasks RESOLVED HERE                                    │
└──────────────────────────────────────────────────────────────┘
  │ (only if Ponytail returns empty)
  ▼
┌──────────────────────────────────────────────────────────────┐
│ STEP 1: GLM 5.2 PRE-DESIGN                   Touchpoint 1    │
│ Architecture sketch before a single line of code              │
│ Identifies: files, risks, alternatives, complexity           │
└──────────────────────────────────────────────────────────────┘
  │
  ▼
┌──────────────────────────────────────────────────────────────┐
│ STEP 2: KIMI K2.7 CODE GENERATION                            │
│ Primary implementation based on validated architecture        │
└──────────────────────────────────────────────────────────────┘
  │
  ▼
┌──────────────────────────────────────────────────────────────┐
│ STEP 3: DUAL REVIEW                           Touchpoint 2   │
│ Kimi K2.7 + GLM 5.2 BOTH review the code                     │
│ Different architectures → different blind spots              │
│ Scores averaged                                              │
└──────────────────────────────────────────────────────────────┘
  │
  ▼
┌──────────────────────────────────────────────────────────────┐
│ STEP 4: GLM 5.2 ARBITRATION                  Touchpoint 3   │
│ Final synthesis incorporating all critiques                  │
└──────────────────────────────────────────────────────────────┘
  │
  ▼
┌──────────────────────────────────────────────────────────────┐
│ STEP 5: QUALITY GATE 10/10 + VERIFY          $0 · Offline    │
│ Error handling · Type safety · Edge cases                    │
│ Failure modes (3x) · DRY · Dead code · Clarity               │
│ Type-check + test suite execution                            │
└──────────────────────────────────────────────────────────────┘

You: "Build an HTTP client with retry logic"

Model: Generates 80 lines of code with:
  - Custom HTTP wrapper class
  - Manual retry loop with exponential backoff
  - Custom error handling
  - 3 new dependencies suggested
  - No edge case coverage for rate limiting
  - No type validation on responses

Cost: ~$0.008 in API calls
Time: ~3 seconds
Review: Self-review only — the same blind spots apply twice

You: "Build an HTTP client with retry logic"

Step 0 - Ponytail Ladder:
  Keywords detected: "HTTP", "client", "retry"
  Match found: Python requests.Session + urllib3.Retry
  Solution: requests.Session() with built-in retry adapter
  Cost: $0.0000 | Time: 98ms

Step 1-4 skipped — task resolved at Step 0

Step 5 - Quality Gate:
  PASS (10/10) — Standard library, well-tested, no new code needed

Cost: $0.0000
Time: <100ms
Review: Deterministic pattern match — no model hallucinations possible

Task: "Implement a distributed rate limiter with Redis backend"

Single Model:
  - Generates code based on ONE architectural perspective
  - Self-reviews with the same blind spots
  - No systematic quality check beyond model's own judgment
  - Ships whatever the sampling process produces
  - Cost: ~$0.012

Onklaud 5:
  Step 1 - GLM Pre-Design: Validates architecture — sliding window vs token bucket,
           Redis Lua scripting for atomicity, race condition analysis
  Step 2 - Kimi Code: Implements based on validated architecture
  Step 3 - Dual Review: Kimi AND GLM independently review — Kimi catches missing
           edge case on Redis connection timeout, GLM catches race condition
           in the Lua script
  Step 4 - GLM Arbitration: Synthesizes fixes for both issues
  Step 5 - Quality Gate: 10/10 check — error handling, type safety, edge cases

  Result: Production-ready code with 2 independently verified fixes
  Cost: ~$0.015 (3 touchpoints)
  Confidence: Cross-model agreement + quality gate enforcement