> ## Documentation Index
> Fetch the complete documentation index at: https://docs.interchange.io/llms.txt
> Use this file to discover all available pages before exploring further.
# Acceptance policy
> The content rules your Merchandising Agent judges buyer transactions against — what it reads, the three verdicts it returns, and the deterministic guarantees behind creative review.
Your **acceptance policy** is the set of content rules your storefront judges
buyer transactions against — which advertisers, categories, and creatives you
will and won't carry. It is the *content* half of [Govern](/v2/concepts/merchandising-agent#govern):
the policy says **what** is acceptable; your [approval settings](/v2/storefront/approvals/overview)
say **whether a human reviews** at all. They're independent dials, and your
media-buy policy does its work in `manual` mode — where it decides which buys
clear on their own and which are held for you. In `auto` mode the gate is skipped
entirely: every buy forwards and nothing is evaluated.
You can author it either way:
* Open the portable **Acceptance Policy** page from Murph, Claude, ChatGPT, or
another MCP Apps host to create or edit the full policy inline.
* Tell Murph the exact rules you want. Murph shows you the complete markdown for
confirmation, then writes it through the same typed configuration contract.
Creating or activating policy versions requires an account administrator. Other
storefront members can review the active policy and immutable history without
receiving a hidden mutation path through the Page or MCP API.
You don't pick from a fixed menu of categories. You describe your rules, and the
agent reads what **you** wrote. Every save creates an immutable version and can
activate it atomically, so version history never contains a half-applied update.
## Edit and restore versions
The Acceptance Policy page always reads the current active policy and its version
history. **Save and activate** creates a new version without overwriting the old
one. If versions exist but none is active, the history remains visible so an
administrator can preview, reactivate, or reuse a version as a new draft. The
page performs those operations directly; it does not turn a button click into a
new chat prompt or depend on a particular host model.
## Policy vs. approval mode
These are two separate dials. Keeping them straight is the most common point of
confusion:
| | What it is | Where it lives |
| --------------------- | ---------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------ |
| **Acceptance policy** | The content rules you judge against ("no gambling", "fast-food needs review", "we protect our exclusive auto sponsor") | The policy text you author with Murph |
| **Approval mode** | Whether a human reviews before buyers go live (`auto` vs `manual`, per `creativeApproval` / `mediaBuyApproval`) | [Approval settings](/v2/storefront/approvals/overview) |
In `auto` mode the policy isn't consulted — everything forwards. In `manual` mode
with no policy written, nothing pre-clears, so every media buy is held for you.
With both in place, the agent clears clearly-on-policy buys and holds the rest.
## How the agent reads your policy
The pre-screen reads your **written** policy against the buyer's submission. It is
**not** a fixed category list — a rule like "we don't accept fast-food advertisers"
is honored even though "fast food" isn't a built-in category, because the agent
derives the subjects you wrote and matches them against the submission. It also
recognizes common evasion (look-alike characters, light obfuscation) so a rule
isn't sidestepped by spelling. When it cannot judge a submission with confidence,
it **errs toward holding it for you**.
When no policy document is available, the Acceptance Policy Page starts with a
conservative draft: no category is pre-cleared, every advertiser, category, and
creative needs review, and illegal, malicious, deceptive, discriminatory,
exploitative, phishing, impersonation, or destination-concealing submissions
are flagged for operator rejection. This draft has no effect until an account
administrator reviews, saves, and activates it.
While that starter remains active, its review-all contract also makes the
deterministic pre-screen hold every submission for an operator; it is not just
instructional copy.
The three verdicts
When `mediaBuyApproval` is `manual`, the pre-screen sorts each media buy into one
of three buckets (in `auto` mode none of this runs — every buy forwards):
* **Clearly on policy** → auto-forwarded without waiting for you. The buy books
quoted products and hits no rule you wrote.
* **Needs a look** → queued for you. Evidence is thin, the policy is ambiguous, or
a rule says this category needs review.
* **Clearly against an explicit rule you wrote** → surfaced as a strong
**decision-support** signal on the queued item.
That last bucket is a **recommendation, not an action**: the buy still queues for
you, and you make the call.
The platform does **not** auto-reject on its own yet. A "clearly against policy"
verdict is recorded and surfaced on the queued item to speed your decision — it
never rejects the buy automatically. You always decide.
## The second opinion
On top of the deterministic pre-screen, an AI layer gives a **second opinion** —
but only as a **one-way veto on auto-approval**. It runs only when the pre-screen
would otherwise auto-clear a buy, and it can only **downgrade** that to "needs a
look" if it spots a policy conflict the rules missed (for example, a competitive-
exclusivity clause you wrote in prose). It can never approve a buy on its own,
never reject one, and never overrides a hold.
It also **fails closed**: if the AI layer is unavailable, errors, or is uncertain,
the buy escalates to you rather than auto-forwarding. The agent never auto-clears
a buy without a corroborating second opinion.
## How creatives are evaluated against your policy
Creative review applies the same content rules to a buyer's submitted creative,
plus a **rendered-image** check that looks at the actual pixels for brand cues,
prohibited categories, and third-party marks. As with media buys, the AI layer can
route a creative to human review but cannot approve one over a clear policy
violation the deterministic checks already found.
For **locked assets** — a product shot, logo, or book cover you've marked as
exact-fidelity — there's a stronger, **deterministic** guarantee that the AI layer
**cannot override**:
* A creative whose final render goes through the platform's locked-asset
compositor preserves your supplied asset **exactly**, and clears that check.
* A final-stage creative that carries your locked asset but did **not** go through
that deterministic path — for example, one that re-generated or edited your
product from a reference — is **blocked from auto-finalizing and held back for a
human to inspect**. A buyer's generated creative cannot quietly substitute a
look-alike stand-in for your locked product at final approval.
These fidelity checks are deterministic and run independently of the AI layer, so
a crafted creative or brief can't talk the reviewer into clearing them — they
**err toward human review** whenever exact fidelity isn't deterministically
guaranteed.
## What the agent never does on its own
* It **never auto-rejects**. The strongest automated action against a buy is a
decision-support flag on your queue.
* It **never changes an approval gate** for you. As it watches your decisions it
may **suggest** opting a consistently-clean buyer into
[per-buyer auto-approve](/v2/storefront/buyer-auto-approvals/overview) — but you
author the carve-out. See the learned suggestion in
[seller analytics](/v2/storefront/analytics).
* It **never auto-clears without a second opinion**. If the AI layer can't
corroborate, the buy goes to you.
The two settings and the lowest-risk posture.
The queue and the auto-clear vs. escalate decision.
How submitted creatives move through review.
Opting one trusted buyer past the queue.