• Date: 2026-05-22
• Status: Accepted
• Feature: URL-import
• Affects: url_import_spec.md § III.2

Context

Изначально Claude был proposed в стеке URL-import:

• Primary VLM для vision analysis
• Text LLM для component enrichment
• Code gen для TSX generation

Two independent reasons exclude Claude entirely:

Reason 1: Conflict of interest

Этот спек был написан AI ассистентом (Claude). Claude рекомендовал использовать Claude API — это conflict of interest, не моя позиция давать competitive advice.

"Я (Claude) изначально рекомендовал Claude Sonnet 4.6 как primary VLM — это conflict of interest."

После research юзера обнаружили: Gemini 2.5 Flash-Lite 30× дешевле при сопоставимом качестве для этой задачи. Объективно better choice для cost-conscious requirement.

Reason 2: Anthropic ToS (Feb 2026)

"You may not use outputs from Anthropic services to train, fine-tune,
or develop AI/ML models that compete with our services, or extract
embeddings or representations for downstream model training."

URL-import имеет data flywheel (ADR 0013) — production extractions train in-house модель (ADR 0014). Если Claude в стеке — даже как inference-only, риск ToS violation:

1. Production extraction calls Claude → response stored в shadow_dataset
2. Training pipeline reads shadow_dataset → может tренировать на Claude outputs
3. Even unintentional → ToS violation

Чистая solution: Claude not в стеке вообще. Никакого риска contamination.

Decision

Claude API исключён entirely из URL-import pipeline:

Verification

Pipeline должен periodically verify Claude не contaminate stack:

// Pre-deployment check (CI hook)
function verifyNoClaudeImports(packageJson) {
  const forbidden = ['@anthropic-ai/sdk', 'anthropic'];
  for (const dep of forbidden) {
    if (packageJson.dependencies?.[dep] || packageJson.devDependencies?.[dep]) {
      throw new Error(`Forbidden dependency: ${dep}. См ADR 0015.`);
    }
  }
}
 
// Runtime check (network egress monitor)
function alertOnClaudeApiCalls() {
  if (request.host.includes('anthropic.com')) {
    alert.send('ADR 0015 violation: Claude API call detected');
  }
}

Consequences

Pros:

• Zero ToS exposure
• Zero conflict-of-interest concerns в documentation
• Forces competitive cost analysis (resulted в Gemini Flash-Lite at 30× lower cost)
• Pipeline transparent — caller знает что не Claude

Cons:

• Loses Claude quality в edge cases (Claude известен superior в some long-context reasoning)
  • Mitigated: Gemini Flash-Lite adequate для structured extraction tasks
  • Mitigated: Qwen3-VL-32B in-house model improves over time

Verification ongoing

• CI hook checks package.json
• Network egress monitoring (anthropic.com domain)
• Quarterly audit of dependency tree

Caveat: ARNO core может use Claude отдельно

Этот ADR scopes только URL-import feature. ARNO core (editor, MD smart editor, etc) может independently choose Claude если требуется. URL-import pipeline isolated.

Alternatives rejected

A. Use Claude API, document conflict of interest

• ❌ ToS exposure remains
• ❌ Spec author bias persists in design

B. Use Claude initially, replace with Qwen LATER

• ❌ Migration cost
• ❌ Production data contaminated с Claude outputs (forever stored in shadow_dataset)
• ❌ Training pipeline could accidentally train on contaminated data

C. Use OpenAI вместо Claude

• ❌ Same ToS issues (OpenAI Usage Policy similar)
• ❌ 5-10× дороже Gemini Flash-Lite

Cross-references

• Main spec § III.2 — НЕ используем list
• Main spec § 0.10 — conflict of interest disclosure
• ADR 0014 — Qwen teacher (positive replacement)