Article VI: The Shadow Stack Prohibition

The Law of Uniform Enforcement

A constitution that only applies in production is not a constitution. It is a suggestion that ends at the staging environment boundary.

Section 1: No Ungoverned Inference

Every inference call — production, staging, development, and CI — must route through the same metering and routing infrastructure.

The most common constitutional failure in developer stacks: production has metering, but engineers bypass it with direct API keys in local environments, notebooks, and eval scripts. Those ungoverned calls train bad habits, leak credentials, and produce cost data that does not reflect real usage patterns.

Requirements:

• No raw provider API keys in application code or environment variables accessible to developers
• All environments route through a proxy layer that enforces Articles I–V
• Local development uses the same routing and metering stack (with separate budgets)

Section 2: The Eval Tax

Offline evaluation suites that call live APIs must carry explicit budgets and produce ledger entries.

Benchmarking is not free. Running 10,000 eval prompts against GPT-4 is a real expense that must be:

• Budgeted before execution
• Tagged with feature: "eval-suite" and the specific benchmark name
• Included in monthly FinOps reports alongside production spend

Ungoverned eval pipelines are how "we only spent $50 in production" becomes "$12,000 on the invoice."

Section 3: Framework Neutrality

Third-party agent frameworks, orchestration libraries, and no-code AI tools must not bypass constitutional enforcement.

If your agent framework makes a direct OpenAI call without passing through your metering proxy, the framework is not a productivity tool — it is a constitutional bypass. Wrap it, fork it, or replace it.

Previous: Article V: Prompt Schema Standards · Back to: The Constitution