# Banned Providers

This document is a hard policy. It is not advisory. It is not aspirational.
Any agent, wizard, or automated process that violates this policy is broken
and must be fixed immediately.

## Permanently Banned

### Anthropic (Claude)

**Status:** BANNED — April 2026
**Scope:** All configs, fallback chains, playbooks, wizard bootstraps, and fleet scripts.
**Enforcement:** Pre-commit hook, architecture linter, sovereignty enforcement tests.

No Anthropic model (Claude Opus, Sonnet, Haiku, or any variant) may appear as:
- A primary provider
- A fallback provider
- An OpenRouter model slug (e.g. `anthropic/claude-*`)
- An API endpoint (api.anthropic.com)
- A required dependency (`anthropic` pip package)
- An environment variable (`ANTHROPIC_API_KEY`, `ANTHROPIC_TOKEN`)

### What to use instead

| Was | Now |
|-----|-----|
| claude-opus-4-6 | kimi-k2.5 |
| claude-sonnet-4-20250514 | kimi-k2.5 |
| claude-haiku | google/gemini-2.5-pro |
| anthropic (provider) | kimi-coding |
| anthropic/claude-* (OpenRouter) | google/gemini-2.5-pro |
| ANTHROPIC_API_KEY | KIMI_API_KEY |

### Exceptions

The following files may reference Anthropic for **historical or defensive** purposes:

- `training/` — Training data must not be altered
- `evaluations/` — Historical benchmark results
- `RELEASE_*.md` — Changelogs
- `metrics_helpers.py` — Historical cost calculation
- `pre-commit.py` — Detects leaked Anthropic keys (defensive)
- `secret-scan.yml` — Detects leaked Anthropic keys (defensive)
- `architecture_linter.py` — Warns/blocks Anthropic usage (enforcement)
- `test_sovereignty_enforcement.py` — Tests that Anthropic is blocked (enforcement)

### Golden State

```yaml
fallback_providers:
  - provider: kimi-coding
    model: kimi-k2.5
    reason: Primary
  - provider: openrouter
    model: google/gemini-2.5-pro
    reason: Cloud fallback
  - provider: ollama
    model: gemma4:latest
    base_url: http://localhost:11434/v1
    reason: Terminal fallback — never phones home
```

*Sovereignty and service always.*