Safety & enforcement

Gate & enforcement

Every call is evaluated by the gate before reaching the provider. The gate runs up to ten checks in order and returns one of three decisions: Approve, Block, or Review required.

Gate decisions

What the gate checks

Checks run in order — the first one that triggers wins. Later checks are skipped.

1. Session limit — if the session has exceeded its spend cap, block (402).
2. High-risk payment pattern — tool names suggest a payment transaction with high confidence. Block (403).
3. Per-model budget — model's daily cap is exhausted or model is paused. Block (402).
4. Model blocked — this model is in the key's block list. Block (403).
5. Model not allowed — key has an allow list and this model is not on it. Block (403).
6. Max tokens exceeded — max_tokens on the request exceeds the key's per-call limit. Block (403).
7. Daily budget exceeded — company, team, or project budget is exhausted. Block (402) or proceed per overage policy.
8. Premium model + high cost — claude-opus or gpt-4 with an estimated cost above $0.005. Review required.
9. Medium-risk payment pattern — tool names suggest a payment transaction with medium confidence. Review required.
10. Vendor payment intent — high-confidence intent detection of payment-related content. Review required.

The review queue

When the gate returns Review required, the call still completes and your app gets the AI response. However:

• The response includes the header X-Zelyx-Review-Required: true.
• The call appears in the Gate dashboard under Pending reviews.
• Budget is reserved — the call did consume tokens and incur cost.

A workspace manager or team lead can review flagged calls on the Gate page and mark them as approved or rejected. Rejection is informational — it does not reverse the cost or undo the call.

Reading the header in your code

If your app needs to handle review-required differently (e.g. hold the response until a human approves):

response_raw = client.chat.completions.with_raw_response.create(
    model="claude-opus-4-7",
    messages=messages,
)

if response_raw.headers.get("X-Zelyx-Review-Required") == "true":
    # Flag for human review in your own system before acting on the output
    flag_for_review(response_raw.parse())
else:
    act_on(response_raw.parse())

Shadow mode

Shadow mode runs all gate checks but never blocks. Every call that would have been blocked or flagged is recorded as a shadow decision, but the call proceeds normally.

Use shadow mode to:

• Understand what the gate would do before enabling enforcement
• Tune budgets and policies without disrupting your team
• Identify false positives before going live

Enable shadow mode in Settings. Results are visible on the Gate page under Shadow mode results, broken down by reason.