# Safety & Guardrails (/docs/concepts/safety-guardrails)



Every automated decision in Layers is gated by safety guardrails. These are
**non-overrideable** unless explicitly stated.

## Optimizer guardrails [#optimizer-guardrails]

Applied to every pause / budget / bid decision the optimizer proposes:

1. **Min age 3 days** — no action on an entity younger than 3 days.
2. **Min 100 impressions** — budget/bid changes require at least 100
   impressions of data.
3. **Min $5 spend** (content-refresh disable step) — ads with under $5 spend
   are not disabled.
4. **Never disable `TOP_PERFORMER` or `HEALTHY`** ads.
5. **Never empty an adset/adgroup** — if pausing would leave zero active
   ads, the optimizer holds.
6. **Max 50% change and $100 absolute** per budget move.
7. **Action caps per cycle**: no more than 40 total actions or 20 pauses in
   a single optimizer cycle.
8. **Cooldowns** between repeat actions on the same entity: 48h structural,
   48h budget, 24h bid.
9. **Min confidence 0.5** — proposals below this confidence score are
   rejected.

## Creative generation guardrails [#creative-generation-guardrails]

Applied to generation workflows:

1. **Ad eligibility / safety review** — content whose upstream safety check
   fails (`safety.pass = false` or `metadata.adEligibility.eligible = false`)
   is marked `safetyFailed = true` and forced to `organic_score = 0`,
   permanently excluding it from ad selection.
2. **Approval policy** — projects with
   `metadata.approval_policy.requiresApproval = true` block scheduling of any
   content whose `approval_status` isn't `approved` or `not_required`.
3. **Credit gate** — generation commands check available org credits
   before queuing. If the org is out of credits the request fails with
   HTTP 402.

## Distribution guardrails [#distribution-guardrails]

Applied to social posts:

1. **Posts-per-day cap** — Social Distribution limits posts per platform
   per day based on the selected provider (e.g. 2 for `ds`, 4 otherwise).
2. **Account health** — if a social account's OAuth token is invalid or the
   account has been flagged, distribution activities error out for that
   account.
3. **Approval gate** — if the project's approval policy requires human
   approval, scheduled posts wait until a human approves before publishing.

## Spending guardrails [#spending-guardrails]

1. **Generation credit cap** — enforced per request against the org's
   credit balance. Overages are billed automatically at the end of the
   billing period via Stripe.
2. **Agency wallet balance** — if insufficient, agency-mode ads are paused
   and you're notified.

## How to override [#how-to-override]

* **Manual pauses / resumes** on ads, adsets, campaigns — done in the app or
  via API; the optimizer's cooldowns apply equally to manual changes within
  the project window.
* **Creative overrides** — use `project_ads_content.override = 'include'` /
  `'exclude'` to force a creative in or out of ad selection.

## What to do when a guardrail fires [#what-to-do-when-a-guardrail-fires]

Check the workflow run in the layer's **Runs** tab. Every proposal the
optimizer rejects records a reason (e.g., "Entity too young (2d \< 3d
minimum)", "Change exceeds 50% cap"). Fix the underlying condition (age up,
spend up, manual approval) or wait for the next cycle.