Skip to main content
Certification is how Apostra decides that a sales agent can take buys. Apostra places a small test buy with your agent for each use case the agent declares, lets it serve through your real ad server, and judges the result from Apostra’s own tracker and from the reports your agent sends. Certification is per use case. An agent that has passed display can take display buys while video is still waiting for its result. Apostra does not wait for everything to be perfect before it runs anything.

The short answers

  • What is checked. Whether the buy goes live, whether your agent hands back the context Apostra sends, whether Apostra’s macros are replaced when the ad serves, whether clicks and video quartiles arrive, whether your reports arrive on time, and whether your numbers agree with Apostra’s tracker and your ad server. See What each check judges.
  • How an agent earns it. You allow certification buys, and Apostra starts a run on its own. An agent is certified for a use case when its test buy passes. See How a developer starts.
  • How it keeps it. It stays certified while live traffic for that use case keeps passing. Real buys count as evidence, and Apostra runs a synthetic buy when none is live. See How certification is kept.
  • What it unlocks. A passed use case lets publishers in other organizations go live on your agent, which is what the agent marketplace offers: your agent, used by other organizations. See What a pass unlocks.
  • What it costs. Apostra charges no fee to certify your own agent. Test buys are capped at USD 10 each and USD 50 per run (converted to your currency when you price in another), and they run on your sandbox account, which serves through your own ad server. Your sandbox decides whether any money moves, and Apostra does not pay for test impressions. See Test buy size.
  • Which agents. Sales agents can be certified today. Creative agents and buyer agents cannot be certified yet. Creative agents are tested but stop there, and a buyer agent has no certification path yet.
  • The deadline. Apostra sends new buys to a sales agent only for the use cases it has passed, unless Apostra has exempted the agent from certification. Apostra has exempted existing sales agents until 1 February 2027, and the agent’s page shows the date. Buys already running are not affected. It applies to sales agents because only sales agents can be certified. To be ready, turn on Allow certification buys on the agent’s page now. See The certification deadline.
Not building with Apostra yet? See Building on Apostra or on AdCP directly for what certification adds on top of the protocol.
Turn on Allow certification buys on the agent’s page, or send it with save_agent. Once it is on, Apostra starts certification on its own. If no run starts for your agent within a day and the Agent page shows no setup step to fix, ask Apostra support. The rest of this page describes how certification works once a run starts.
Certification is for sales agents. It is not available for buyer agents or creative agents yet, so a buyer agent’s status stops at the last stage Apostra can prove and says that certification isn’t available.

Which Source certification buys through

Apostra’s certification buyer buys through a Source that uses your agent. Two kinds of Source qualify:
  • A Source that is active.
  • A Source that is still pending, in your own organization, whose health is healthy or degraded. Failing health, or no health reading in the last seven days, holds it out until the health recovers. For this, a creative your agent refused with VALIDATION_ERROR, INVALID_REQUEST or POLICY_VIOLATION does not count against the Source: certification tests your agent’s creative handling itself, family by family, and a pending Source takes no buyer buys. A creative sync that failed any other way, such as a timeout, an authentication failure or a server error, still counts.
A certification buy spends nothing. It is capped, and it runs on your sandbox account (see Test buy size). So a new agent does not have to be live before Apostra can test it. The seller whose Storefront it runs on sees it among their media buys, with a next action that starts “Certification buy from Apostra, not a buyer’s order.” It never counts as that Storefront having gone live. A Source you connect starts pending. Turn on Allow certification buys on the agent’s page, and Apostra certifies through your pending Source. When the agent passes, the Source goes live under the rules in What a pass unlocks. Only Apostra’s certification buys use a pending Source. A buyer’s buy through a pending Source is refused, and setting the Source to ACTIVE yourself is refused until the agent has passed or Apostra has exempted it. A pending Source of another organization is not certified through: it keeps the partner connection path described under What a pass unlocks.

What a pass unlocks

Once its exemption ends, Apostra sends new buys to a sales agent only for the use cases it has passed. Buys already running are not affected. See The certification deadline. One passed use case completes the Successful transaction go-live step for every Source that uses your agent. While Apostra has exempted your agent from certification, the exemption completes the step the same way, and the step says the agent is exempt, not certified. That is the step a seller’s Storefront needs before it can go live with a sales agent. A Source whose agent is not registered with Apostra cannot complete it: the step asks you to register the agent in your organization’s Agents account and connect it to the Source, because Apostra certifies a registered agent. See Connect an agent from its URL and Seller readiness. A pass is also what lets a Source go live on your agent. Apostra certifies through a pending Source in your own organization, so a new agent needs no exemption to get its Source live.
  • A Source in your own organization, or one that was already live before its connection to your agent was recorded and has not since been activated through the partner connection, goes live and stays live while your agent has passed or Apostra has exempted it from certification, and the Source’s own health is not failing. Degraded health is fine; failing health, or no health reading in the last seven days, holds the Source in pending. A Source held this way cannot be set live, and Apostra rechecks live Sources about once a minute and sets one back to pending when its health starts failing or its readings stop. A seller’s agent refusing a buyer’s creative, for example with CREATIVE_REJECTED, FORMAT_NOT_SUPPORTED or CREATIVE_SIZE_MISMATCH, does not make the Source’s health fail: the agent answered, so the refusal shows as degraded, which keeps it visible without taking the Source out of service. A creative sync that fails for any other reason still does, including a timeout, an authentication failure, a server error, or a VALIDATION_ERROR, which can mean the whole request was invalid rather than one creative refused. The exception is a certification run: when your agent refuses one of certification’s own test creatives with VALIDATION_ERROR, INVALID_REQUEST or POLICY_VIOLATION, the refusal shows as degraded, because it is about Apostra’s test creative, not your Source. The run reports the refusal on that format’s result instead. Any other failure, including rate limiting, still counts.
  • Any other Source of a publisher in another organization goes through the partner connection instead. It activates only once your agent has passed or Apostra has exempted it from certification, and then goes through authorization, validation, health and commercial checks, which keep applying while it is live.
  • A pass counts for the agent a Source was using when the run happened. It does not follow a Source that is later moved to a different agent.
  • A pass stands until a later result for the same use case and product family fails. Apostra judges a run’s use case again every hour while its evidence is open, and the latest judgement of that run decides: if it ends inconclusive after a pass, for example because a check could not be judged on that buy, the pass no longer counts until a run passes again. A newer run that is still inconclusive never removes an earlier run’s pass.
  • A compact lifecycle result never counts toward a pass or a failure of the use cases it is shown beside. See Use cases. From 1 February 2027 an agent that has not passed it does not count as certified, as the item below says.
  • From a capability’s date, a pass does not count while the agent lacks it (1 February 2027 for all four dated capabilities): its AdCP 4.0 readiness lists it under datedGaps. That includes an agent that passed earlier, and a capability we have not yet proved, such as the compact lifecycle. It counts again once the agent has the capability and the Agent page no longer lists it. Nothing changes before the date. A connection that is already live stays live.
  • A pass does not count while the agent’s behaviour baseline shows a current Blocker on any active Source that uses the agent, on one of the checks that can stop new buys. A Blocker on any other check shows on the Agent page and leaves the pass alone. The pass counts again once the baseline no longer shows a Blocker on one of those checks; the baseline is re-evaluated over the last seven days of traffic. A connection that is already live stays live.
  • Certification does not publish your agent, grant Partner Program status or replace a seller’s credentials, account mapping and health checks. Those stay separate.
  • Partner Program readiness uses this certification. An organization can operate agents for other organizations’ clients only when it owns, through a claimed provider, at least one active sales agent that has passed here, as counted by the rules above, or that Apostra has exempted from certification and that has no current Blocker on one of the checks above. Passing does not grant the right: registration, the commercial agreement and payment standing are separate. If readiness says certification is missing, turn on Allow certification buys on the agent’s page and Apostra starts certification automatically. Publishers see your agent in their list of partner agents, and can connect to it without an invitation, only while your organization holds this right and the agent has a published connection setup.

The certification deadline

Without a certification exemption, Apostra sends new buys to a sales agent only for the base use cases it has passed. Apostra has exempted the sales agents that existed on 22 October 2026 until 1 February 2027, so for them the rule starts when that exemption ends. The date was 1 December 2026 until 15 November 2026, when Apostra moved it to give sellers more time to complete certification. A sales agent created after 22 October 2026 has no exemption, so the rule applies to it from the start. It does not need an exemption to get started: turn on Allow certification buys and Apostra certifies through your pending Source (see Which Source certification buys through). To ask for an exemption, contact Apostra support. The agent’s page shows its exemption date, and so does certificationExemption on the V3 get read of the agent. See See whether Apostra exempted an agent from certification. The rule applies to every sales agent, including one that only your own organization uses, and to sales agents only. Creative agents and buyer agents cannot be certified. Once it applies to your agent:
  • It is per base use case. A pass on one product family of a base use case lets your agent take new buys for every product in that use case. A product that falls in two use cases, such as one sold as both display and linear TV, needs a pass in each. A base use case with no pass takes no new buys. See Use cases.
  • A refused buy never reaches your agent. A new buy for a product whose base use case your agent has not passed is refused with PRODUCT_UNAVAILABLE, and the buyer is asked to choose a different product. A product Apostra cannot place in a certification use case, or one no longer in your catalog, is refused the same way.
  • Running buys continue. Buys already running, and changes to them, are not affected.
  • Some buys are not held. Apostra’s own certification buys, which are how your agent passes, and sandbox buys, which spend nothing, go through as before.
  • A pass can stop counting. The conditions in What a pass unlocks apply here too, such as a current Blocker on a check that can stop new buys.
Do these now, so a pass is in place well before your exemption ends:
  1. Make sure the agent has a Source and a sandbox account that serves. See How a developer starts and A sandbox account that serves.
  2. Turn on Allow certification buys on the agent’s page.
  3. Read the result under Live certification and fix what a failed or inconclusive use case names. A certification buy serves for about three days, plus the reports due after it, so start early. See How long it takes.

How a developer starts

Every agent lives in your organization’s Agents account. See Connect an agent in an Agents account. Certification buys go through a seller Source that uses your agent, so an agent needs a Source before it can be certified.
  • From your Agents account, connect the agent to your organization’s test storefront: a free storefront Apostra creates for you, never listed, that holds only your organization’s own agents. See Test and certify your agent.
  • If your organization runs its own storefront, you can connect the agent to it as a Source instead. See Connect your own sales agent.
  • If you build an agent for other companies to use, ask Apostra support to start. Apostra arranges the connection today, and your agent is certified through the seller Source that uses it.
Then allow Apostra to send certification buys. You do not need to prove control of the agent’s URL first to certify an agent through your own organization’s storefront: the agent is reached only through Apostra there, so nothing is exposed. Proof is needed for a marketplace listing, Partner standing and use by another organization.
  1. Turn on Allow certification buys. Use the agent’s page, or send certificationRequested: true with V3 save_agent. See Allow certification buys. An Operator or Admin of the Agents account that owns the agent can set it, signed in as themselves, and only a sales agent can be requested. An Operator is enough because a certification buy spends nothing. See Who can do what with an agent. Turn it off, or send certificationRequested: false, and Apostra sends no new certification buys; buys already running continue. With it off, only your real traffic can show your agent working. Read The deadline for what that means once buys depend on a pass.
  2. Apostra starts a run on its own, with no one at Apostra acting. It runs on each active Source that uses your agent, usually within the hour. It checks the agent-wide baseline, then runs a certification buy for each use case the agent declares, on each matching product family.
  3. Results appear on the Agent page under Live certification, and in the V3 get read of the agent. Each use case shows Passed, Failed or Inconclusive, when it was last judged and a reason when it did not pass.

Channels not yet covered

If your storefront sells only channels that certification’s standard briefs do not cover, the agent page names the channels and no certification run starts. You do not need to change your storefront’s channels. Apostra retries automatically; a run can start when certification covers a channel your storefront serves and the other certification requirements are met.

What makes an agent ready

An agent is ready for certification when two things are true. The agent-wide baseline holds, and the agent has a sandbox account that serves through your ad server.

The agent-wide baseline

Apostra reads these from your get_adcp_capabilities response and the products your agent returns. They apply to the whole agent, not to one use case.
  • AdCP 3. adcp.major_versions includes 3, and responses validate against the version Apostra negotiates.
  • A billing model Apostra supports. supported_billing includes agent or operator, or both.
  • The buying modes it serves. media_buy.buying_modes lists brief, wholesale, or both, with exactly the paths the agent serves. Certification tests every path in the agent’s product mode. Apostra takes the product mode from this declaration. Every Source the agent serves runs each declared path. A Source whose agent has no usable declaration has no product mode. See Declare your buying modes.
  • A way to receive audiences, if it declares audience targeting. An agent that declares audience_include or audience_exclude in media_buy.execution.targeting also lists sync_audiences or declares media_buy.audience_targeting, the AdCP declaration for it. This is reported only: it stops neither certification nor buying. See Declare how audiences reach you.
  • A creative kind Apostra supplies. At least one product accepts display_tag, video_vast or audio_daast, or declares legacy format_ids.
Agent requirements covers the baseline and the seven-day behaviour thresholds in full. What we need from reporting lists the thirteen things Apostra asks a seller’s reporting to carry. They are judged as report-only findings: none stops a buy or a certification pass.

Behaviour on live traffic

Apostra also measures how the agent behaves on its own traffic over the last seven days, such as whether it answers in time, completes async tasks, and returns the same account for a repeat sync_accounts. Every agent owner sees this behaviour baseline on the Agent page, under Agent-wide behaviour. Each check shows Pass, Flag or Blocker, with what Apostra saw and how to fix it. When a check is a Blocker, fixing it is the agent’s next step, and the page links the check’s section of the requirements page. A missing declaration from the list above still comes first. A check can also read Not read this run. That means Apostra could not read the check’s evidence in the latest refresh because the read took too long. It tries again on the next refresh, and you have nothing to do. A check that was not read is never counted as a pass: an Agent-wide check does not pass while any Source’s read of it is missing, although a flag or Blocker that another Source showed still stands. A Blocker found earlier stays a Blocker until a refresh reads the check again, so a slow read never lifts a block on new buys. The baseline also covers a pending Source of your own organization whose health is not failing, the Source certification runs through, so you see these checks while the agent is being certified. A blocker shows on the Agent page. A blocker on one of these checks, on an active Source, stops the agent’s certification pass counting, and an agent without a current certification exemption takes no new buys while its pass does not count. A blocker on a pending Source stops nothing until that Source is active. Running buys continue. Only these checks can stop new buys: Lifecycle agreement checks that your agent tells one story about each buy: get_media_buys and get_media_buy_delivery report the same status, and a buy never reads active while every package that can still serve is paused. It applies from the day it ships, with no report-only period, and to the embedded sales agent Apostra runs. A disagreement that lasts more than 15 minutes is a flag, and a second one within seven days is a blocker. Changes are announced checks that an agent that declares the change feed puts every lifecycle change on it, within five minutes, under the right account. It applies from the day it ships, with no report-only period, and to the embedded sales agent Apostra runs. One miss is a flag, and a second one within seven days is a blocker. Ended buys are not reported active checks that your agent stops reporting a buy active once the buy’s end has passed. It is new from 26 October 2026 and applies to every agent, including the embedded sales agent Apostra runs. One buy reported active more than 24 hours after its end is a flag, and a second within seven days is a blocker. From that day a flag or a blocker shows on the Agent page, and nothing stops new buys because of it before 1 February 2027: an agent that existed on 22 October 2026 holds a certification exemption until then, so a blocker cannot take its new buys, and stopping new buys on blockers is switched on for no account. Rejected buys are not reported active is a finding today. It shows as a flag on the Agent page when your agent reports a buy active for more than 15 minutes while every package that can still serve has had all of its creatives rejected, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. A buy keeps one status is a finding today. It shows as a flag on the Agent page when a buy your agent reports goes back and forth between the same two statuses, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. A live buy has a serving ad is a finding today. It shows as a flag on the Agent page when your agent reports a buy active inside its flight for more than 15 minutes while no package that can still serve holds a creative, or while a package is paused, beside one that still serves, that Apostra never asked it to pause, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. A synced creative shows in creative approvals is a finding today. It shows as a flag on the Agent page when your agent answers a creative sync for a buy with success and then never lists a creative the sync carried in creative_approvals, on any read made more than an hour later, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Creative changes declared is a finding today. It shows as a flag on the Agent page when your agent does not say which creative changes it takes on a live buy, and it stops nothing: no new buy is held back and no certification pass stops counting because of it. Whether it becomes required is a later decision, and this page will say so before it changes. Delivery reports use the AdCP envelope shows as a flag on the Agent page when a source sends the AdCP 3.0 bare delivery report, or sends delivery reports that fail the schema. From 1 December 2026 it can also be a blocker: two delivery reports from one source refused with 400 within seven days, both received on or after that date. A 3.0 report Apostra accepts stays a flag. Before that date it stops nothing. An agent that existed on 22 October 2026 holds a certification exemption until 1 February 2027, so until then a blocker cannot take its new buys. Package assignments are answered as AdCP allows is a finding today. It shows as a flag on the Agent page when your agent refuses package assignments on a creative sync in any way other than a correctable UNSUPPORTED_FEATURE that names assignments, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Creative assignments on buys are answered as AdCP allows is a finding today. It shows as a flag on the Agent page when your agent refuses package creative_assignments on create_media_buy or update_media_buy in a way AdCP does not define, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Creative approvals use buyer ids is a finding today. It shows as a flag on the Agent page when your agent lists, in a package’s creative_approvals on get_media_buys, a creative_id that is not a creative Apostra sent it, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. An identical retry replays its first answer is a finding today. It shows as a flag on the Agent page when your agent answers IDEMPOTENCY_CONFLICT to a request Apostra sends again, unchanged and under the same idempotency_key, after it answered the first send, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Creative errors carry a recovery is a finding today. It shows as a flag on the Agent page when an error in your agent’s answer to a creative sync carries a recovery that is not transient, correctable or terminal, or, in an answer served as AdCP 3.2 or later, carries none, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Creative review verdicts can be read is a finding today. It shows as a flag on the Agent page when your agent takes a creative into review without a task and its verdict never reaches Apostra before the delivery’s deadline, through a callback or list_creatives, and it stops nothing. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Delivery names only booked packages is a finding today. It shows as a flag on the Agent page when a delivery report from your agent, for a buy that has delivered, names a package the buy never booked, and it stops nothing. A buy that has delivered nothing yet is not judged. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Discovery brief accepted is a finding today. It shows as a flag on the Agent page when your agent refuses a get_products brief that passes the AdCP request schema, again and again, with INVALID_REQUEST or VALIDATION_ERROR, and answers none, and it stops nothing. AdCP gives a seller that needs more from a brief the input-required task status, and one that declines it the rejected outcome. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Delivery by creative is its own and adds up is a finding today. It shows as a flag on the Agent page when a delivery report from your agent carries by_creative rows that are equal shares of the package total, or that do not add up to the package’s impressions or spend, and it stops nothing. A report with no creative rows is not judged. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. It is requirement WNR-2, WNR-3 and WNR-4 on What we need from reporting. Stays reachable checks that your agent keeps answering the calls Apostra makes to it. It is new from 29 October 2026 and applies to every agent, including the embedded sales agent Apostra runs. A stretch of more than 15 minutes in which Apostra called your agent and got no response is a flag, and a second stretch within seven days, or one that lasts six hours, is a blocker. From that day a flag or a blocker shows on the Agent page, and nothing stops new buys because of it before 1 February 2027: an agent that existed on 22 October 2026 holds a certification exemption until then, so a blocker cannot take its new buys, and stopping new buys on blockers is switched on for no account. An agent whose deploys or restarts leave it unanswering for a few minutes is not affected. Payment terms, account sync is idempotent and declarations match behaviour show a flag or a blocker but do not stop buys. This page will say so before that changes. Valid responses shows a flag when Apostra refuses a reply your agent sent as not valid AdCP, and it never stops a buy. Machine-readable errors is a best practice and never stops a buy. See What a pass unlocks for how a Blocker affects the pass.

The embedded sales agent Apostra runs

The embedded sales agent Apostra runs for sellers is judged by the same checks as your agent, over the same seven days of live traffic and with the same thresholds. Each ad server it runs (Google Ad Manager, FreeWheel, SpringServe, AdsWizz and Broadsign) is its own Agent, so a finding on one ad server never marks another down. A source the embedded sales agent serves counts towards the Agent for its ad server, and that Agent’s page shows its findings to Apostra’s team. Apostra never places a certification buy on a seller’s own ad server through the embedded sales agent. Its certification evidence comes from live traffic only. A blocker on the embedded sales agent does not stop new buys or a certification pass yet. This page will say so, with the date, before that changes.

Lifecycle agreement finding

Threshold. For each Source, over seven days, Apostra checks two things each time it reads a buy:
  • The two status answers agree. get_media_buys and get_media_buy_delivery report the same status. Apostra sets each delivery answer against the get_media_buys answer it read within five minutes of it.
  • The buy agrees with its packages. A buy that reads active while every package that can still serve is paused is a disagreement. A package can still serve when it is not paused and not finished. It is finished when canceled is true or its end_time has passed. When every package is paused or finished and at least one is paused, AdCP reads the buy as paused.
A disagreement counts once it has lasted more than 15 minutes: Apostra read it twice, more than 15 minutes apart, with no agreeing read between. One that clears within 15 minutes counts for nothing, so your agent has that long to reflect a change. One disagreement is a flag. Two within the seven days are a blocker, and two buys that disagree count as two. When Apostra has no pair of reads to compare and no get_media_buys answer that carries its packages, the check is not evaluated. Each of the two parts is judged on its own: a get_media_buys answer without packages still has its status set against the delivery answer, and only the package part is skipped. The check applies from the day it ships, with no report-only period, and to the embedded sales agent Apostra runs. A blocker stops the agent’s certification pass counting, as for the other checks that can stop new buys, and running buys continue. Usual causes.
  • The agent updates its own record of a buy only when it makes a change itself, so a change made directly in the ad server never reaches get_media_buys.
  • A pause reaches the packages that can still serve, but the agent reports the whole pause as failed because a finished package refused it. The ad server is changed, the agent’s record is not, and the buy keeps reading active.
  • get_media_buy_delivery derives its status from flight dates, while get_media_buys reads the agent’s record.
Next step. Answer both reads from one record, and bring that record in line with your ad server within 15 minutes of any change, including one made directly there. Treat a package whose flight has ended as finished: it is never a reason a pause fails. Report the buy paused once every package that can still serve is paused. The finding clears as the seven-day window moves on.

Changes announced finding

Applies when your agent declares account.change_feed in get_adcp_capabilities. An agent that does not declare it is read by polling and is not judged here. Threshold. For each Source, over seven days, Apostra counts two kinds of miss:
  • A change the feed did not announce. Apostra’s scheduled read of a buy found its lifecycle status changed, and your list_account_changes feed carried no record, on the feed for the account the buy is in, naming that buy (a media buy, one of its packages, one of its creative assignments, or a creative record that names the buy as its parent) recorded between the earlier read and five minutes after the read that found the change, and that Apostra received within 15 minutes after that. Apostra reads every feed at least every 15 minutes, so a record your feed exposes later than that does not count, whatever time it carries. A record whose account does not resolve never counts as the announcement; it is counted once, as the record below, and the change it names is not counted a second time. A record announces one change, so two changes need two records. A change Apostra read because the feed announced it is announced by definition. Apostra judges a change only once that account’s feed has been read successfully after that window, so a feed it could not read is never counted as a missing record. Apostra judges only changes whose earlier read came after it began keeping your feed’s records, so a change from before the check shipped is never counted against you.
  • A record that does not resolve to an account. A record names an account other than the one the feed was read for. Each record counts once, by its change_id, however often it is read again, and it stops counting once your agent returns it under the right account_id.
One miss is a flag. Two within the seven days are a blocker. When the feed has shown Apostra nothing to check, the check is not evaluated. It applies from the day it ships, with no report-only period, and to the embedded sales agent Apostra runs. A blocker stops the agent’s certification pass counting, as for the other checks that can stop new buys, and running buys continue. What Apostra keeps. For each record your feed announces, Apostra keeps the change_id, the time your agent gave, the buy the record names, the action, the resource type and whether the record’s account resolves. It keeps no other seller text. A record is kept for 14 days after Apostra received it. A cleanup that runs every 15 minutes then deletes it, removing up to 100,000 records a run, so a very large backlog can take longer. A record is deleted sooner when the Source or its feed is removed. One exception: a record that does not resolve to an account is kept past 14 days until your feed has been read again without it, or your agent returns it under the right account_id. A feed that is still registered but never answers again keeps those records until the feed is removed. Usual causes.
  • The agent records a change in its own store without writing it to the feed, so a change made directly in the ad server never reaches the feed.
  • The feed returns records under a placeholder account id such as default instead of the account_id the feed is read for. Apostra then cannot use the page at all.
  • Package changes are written to the feed with no parent_ids.media_buy_id, so they do not name the buy they belong to.
Next step. Write every change to a buy or a package, whoever made it, to the account.change_recorded feed within five minutes of recording it, with resource.account_id set to the account the feed is read for and, for a package, parent_ids.media_buy_id. If your agent cannot, stop declaring account.change_feed; Apostra then reads your buys by polling. The finding clears as the seven-day window moves on.

Ended buys not active finding

Applies when Apostra read a buy from your agent more than 24 hours after the buy’s end. A Source with no such read is not judged. Threshold. For each Source, over seven days, Apostra counts each buy it read as active through get_media_buys more than 24 hours after the buy’s end. The end is the end_time your agent sent for the buy on that read, read as the moment it names, so an offset such as +02:00 is fine. When your agent sent none, it is the end Apostra booked for the buy. When it sent one Apostra cannot read as an RFC 3339 timestamp, Apostra does not judge that read. A buy that your agent reports paused, completed, canceled or in any other status is not counted, so only a buy reported running after its end has passed is a miss. A buy counts once, however often Apostra reads it, and it stays in the window until it is seven days old. One buy is a flag. Two within the seven days are a blocker. Apostra does not complete a buy because its end date has passed: your agent owns the buy’s state, so a buy it still reports active stays active in Apostra’s records, and this check is how the disagreement reaches you. Usual causes.
  • The agent keeps reporting the status it held in its own store, and does not follow its ad server once the flight has ended.
  • The agent reports the end Apostra booked, but the buy was extended or moved in the ad server, so the real end is later. Report the end it actually runs to.
  • The agent reports active for a buy its ad server finished or removed.
Next step. When a buy’s end passes, report it completed (or canceled) in get_media_buys. If a buy runs past the end Apostra booked, report the later end_time it runs to; Apostra judges a buy against the end your agent reports. The finding clears as the seven-day window moves on.

Rejected buys not active finding

Applies when Apostra read a buy from your agent through get_media_buys, and a package in the answer listed the review state of its creatives in creative_approvals. A Source whose packages list no creative approvals is not judged. Threshold. For each Source, over seven days, Apostra counts each buy that reads active while every package that can still serve has all of its creatives rejected. A package can still serve when it is not finished: it is finished when canceled is true or its end_time has passed. A package has all of its creatives rejected when creative_approvals lists at least one creative and every entry has approval_status of rejected. A package that holds any creative that is approved, partly approved or still in review is not wholly rejected, so a buy with one such package is not counted. No package of such a buy can deliver, and AdCP does not let it read active: it reads pending_creatives while the buyer can replace the rejected creatives, or rejected when your agent declines the buy. A contradiction counts once it has lasted more than 15 minutes: Apostra read it twice, more than 15 minutes apart, with no agreeing read between. One that clears within 15 minutes counts for nothing, so your agent has that long to reflect a review outcome. A buy counts once however often Apostra reads it. One buy is a flag. The check is report-only today: it never becomes a blocker, it does not stop certification, and it does not stop new buys. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Usual causes.
  • The agent reports the buy status it held when the buyer created the buy, and does not move it when the ad platform later rejects the ad set, ad group or creatives.
  • The agent reads a package’s own review outcome from the ad platform, but never carries it into the buy’s status.
  • The agent reports the buy’s status from its flight dates, not from whether any package can deliver.
Next step. When your ad platform rejects what a package needs to deliver, report the rejection on the package in creative_approvals, with approval_status rejected and a rejection_reason, in AdCP’s shape. Report a media buy status that reflects it within 15 minutes: pending_creatives while the buyer can replace the creatives, or rejected when your agent declines the buy. Validate the answer with the pinned AdCP SDK schema. The finding clears as the seven-day window moves on.

Reconciliation not stuck finding

Applies when Apostra flagged one of your buys for reconciliation because it could not confirm whether a targeting, budget/bid price, name, or flight-date change you were sent actually applied — a transient sync failure, for example. A Source with no such flagged buy is not judged. Threshold. For each Source, over seven days, Apostra counts each buy that carries the flag. A buy counts as a miss once the flag has stood for more than 24 hours without being confirmed or refused. One buy is a flag. The check is report-only today: it never becomes a blocker, it does not stop certification, and it does not stop new buys. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Usual causes.
  • The agent’s next status read repeats the value it held before the change, instead of the value the change actually produced.
  • The agent’s read answers with an ambiguous or missing value for the changed field, so Apostra cannot tell whether the change applied.
  • A transient failure on the agent’s side during the change leaves it unable to say, on its next read, whether the change took effect.
Next step. When you cannot confirm whether a change applied, make sure your next get_media_buys/get_media_buy_delivery read answers with a clear, consistent value for the changed field, so Apostra is not left guessing. Nothing automatically re-checks a flagged buy today, so the finding itself stays until the buy is canceled or Apostra ships the automated re-verification this finding is meant to motivate; a clear answer is exactly what that re-verification will act on once it exists.

Status not flapping finding

Applies when your agent reported a change to the status of a buy Apostra placed with it, through a status read (get_media_buys) or a webhook. A Source whose buys had no such change in the last seven days is not judged. Threshold. For each Source, over seven days, Apostra counts each buy whose status your agent changed back and forth between the same two statuses three times in a row, each change within two hours of the last. A change reverses the one before it when it puts the buy back to the status it had just left, so the status goes active, pending_start, active, pending_start, active. A status the buyer asked for, such as a pause, is never counted, and a buy your agent pauses and resumes once is not counted. A buy counts once however often it flaps. One buy is a flag. The check is report-only today: it never becomes a blocker, it does not stop certification, and it does not stop new buys. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Usual causes.
  • get_media_buys and get_media_buy_delivery name different statuses for the same buy. The most common case is a buy booked and enabled before its flight starts: the status read says pending_start and the delivery read says active, so each read undoes the other. This is the case that first showed the problem, on an OpenAI Ads buy that changed state more than 20 times in eight hours.
  • A status derived from one field, such as the platform’s enabled flag, in one read and from the flight dates in another.
  • A webhook and a status read that describe the buy at different moments.
Next step. Report one status for a buy at any moment, in every read. Before a buy’s start_time, report pending_start in get_media_buys, in get_media_buy_delivery and in every webhook, even when the buy is enabled at your platform, and move to active when the flight starts. See Report pending start until the flight starts. The finding clears as the seven-day window moves on.

Live buys serving finding

Applies when your agent reports a buy Apostra placed with it active in get_media_buys, with the buy’s packages listed, once the buy’s flight has run for an hour. The flight is the start_time and end_time your agent reports for the buy, or the ones Apostra booked when it reports none. A Source with no such read in the last seven days is not judged. Threshold. For each Source, over seven days, Apostra counts each buy that reads active inside its flight for more than 15 minutes, in consecutive reads with no agreeing read between them, while either of these holds:
  • No ad. At least one package can still serve (it is not paused, canceled or ended), every such package lists the creatives it holds in creative_approvals, and none of them holds a creative, whether approved, pending or rejected.
  • Paused with no pause request. A package that is not finished is paused while another package of the buy can still serve, and Apostra never asked your agent to pause the buy or that package. A pause the buyer asked for is never counted, and a request to pause one package does not excuse the pause of another.
A buy with every package paused is judged by the lifecycle agreement check, and a buy with every creative rejected by rejected buys not active, so neither counts here. A seller that leaves creative_approvals off a package that can still serve is not judged for the first condition. A buy counts once however often it contradicts. One buy is a flag. The check is report-only today: it never becomes a blocker, it does not stop certification, and it does not stop new buys. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. Usual causes.
  • An ad platform creates a campaign, an ad group or an ad paused, and nothing switches it on once the creative is in. This is the case that first showed the problem, on an OpenAI Ads buy whose only ad group stayed paused for ten hours while the buy read active. With every package paused the lifecycle agreement finding catches it, and with one of several paused this one does.
  • A buy reported active from the platform’s campaign status alone, without looking at whether the buy holds an ad.
  • A creative that never reached the platform, or was removed, while the package still reads as live.
Next step. Report a buy active only when it can serve. List the creatives each package holds in creative_approvals on get_media_buys. When none of the packages that can still serve holds one, report pending_creatives, as AdCP has a buy without creatives read. Report a package as paused only when Apostra asked you to pause it, and switch on a package your platform created paused once its creative is live. The finding clears as the seven-day window moves on.

Creative approvals after sync finding

creative_approvals on get_media_buys is how your agent tells Apostra what review each creative on a buy has reached. When your agent takes a creative for a buy and never lists it there, the buyer cannot see whether the creative was approved or rejected: it stays unreviewed, and a rejection never arrives. Apostra matches each creative by the creative_id it sent on sync_creatives, so a buy that lists one creative does not excuse another that your agent accepted and left out. Applies when Apostra sent your agent a sync_creatives request for a buy it placed with it, that carried at least one creative, your agent answered with success, and an hour has passed since that answer. It also needs your agent to list creative_approvals on at least one package of the buy that can still serve (not paused, canceled or ended), on a read that keeps the buy’s packages. A Source with no such buy in the last seven days is not judged, and neither is a seller that leaves creative_approvals off those packages: an omitted list says nothing about what the package holds. Threshold. For each Source, over seven days, Apostra counts each such buy where a creative the sync carried is named by creative_id in no package’s creative_approvals, whether approved, pending or rejected, on any read made more than an hour after your agent answered the sync, over more than 15 minutes between the first read and the last. The creative on any later read stops it counting. A buy counts once however many of its creatives or reads miss. One buy is a flag. The check is report-only today: it never becomes a blocker, it does not stop certification, and it does not stop new buys. Whether it becomes required is a later decision, and this page will say so, with the date, before it changes. This is different from the package assignments check, which asks whether each creative a sync assigned to a package appears there. This check asks it of the whole buy and does not need the sync to carry assignments. Usual causes.
  • The review status is read from one place when the buy is created and from another when it is polled, so the poll finds no record of the creative. This is the case that first showed the problem, on an OpenAI Ads buy whose ad was live and whose creative_approvals stayed empty for days.
  • A creative that your platform holds under its own id, and that your agent cannot map back to the creative_id Apostra sent.
  • creative_approvals filled in only for a package your platform created from the creative, and left empty when the creative is on the buy another way.
Next step. List each creative your agent holds for the buy in creative_approvals on its package, under the creative_id Apostra sent on sync_creatives, with the review status it holds: pending_review while it is being reviewed, then approved or rejected. The finding clears as the seven-day window moves on.

Creative changes declared finding

Apostra checks what your agent declares in available_actions (or the older valid_actions) before it changes a live buy, and refuses a change your agent does not take. Adding a creative to a package, taking one off and swapping one in place on a live buy (update_creative_assignments, remove_creative and replace_creative) are the exception today: when your agent declares no action list at all, Apostra lets the change through to it, because many sales agents do not declare yet. This check reports that gap so you can close it. Applies when Apostra read a live buy (active or paused) from your agent through get_media_buys in the last seven days. A Source with no such read is not judged. Threshold. For each Source, over seven days, Apostra takes the latest read of each live buy and checks whether your agent sent an action list on it, either available_actions or a non-empty valid_actions. A buy read with no action list is a miss. A buy counts once, however often Apostra reads it. An agent that sends a list is not judged on what is in it: leaving a creative action out of a list that is present is a valid way to say the agent does not take that change. While this is a finding, one or more misses show as a flag on the Agent page. It is never a blocker today, so it stops nothing: no buy is held back, no certification pass stops counting, and no action is needed to keep selling. Whether a missing declaration becomes required is a later decision, which the changelog and this page will announce before it takes effect. The check also shows, for the buys that did declare a list, how many take each of the three creative changes, in any mode (self_serve, conditional_self_serve, requires_approval or seller_managed). Usual causes.
  • The agent declares available_actions on some responses and not on get_media_buys, which is where Apostra reads it.
  • The agent sends no action list at all, because it predates AdCP 3.
  • The agent sends a list, but the creative actions are missing from it. That is not a miss for this check: it tells Apostra the agent does not take the change.
Next step. Declare the live creative changes you take in available_actions on every media-buy response (get_media_buys, create_media_buy and update_media_buy): update_creative_assignments to add a creative to a package, remove_creative to take one off and replace_creative to swap one in place, each with its mode. A product that should not take a live creative change leaves that action out of a list that is present: an absent action means not supported, an absent list means you have not said. Declaring support is what lets a buyer change creatives on a live buy from Apostra’s tools, so it is the outcome to aim for. The finding clears as the seven-day window moves on. Certification also tries the changes you declare, and the ones you leave out, on a live test buy: see Live creative changes.

Delivery report format finding

A source that pushes delivery reports sends each one to Apostra’s delivery webhook as a media_buy_delivery callback. Apostra records every callback it receives from an authenticated sender, and this check reads those records. Each refused report is also in the agent’s webhook results, with the first failing field. Applies when a source sent a delivery report to Apostra in the last seven days. A source that sent none is not judged, for example because Apostra reads its delivery by polling. Threshold. For each Source, over seven days, Apostra counts the delivery reports it answered 200 or 400:
  • A report refused with 400 is a miss. The message and the webhook results name the first failing field, whichever field failed.
  • A report in the AdCP 3.0 shape, with media_buy_deliveries at the top level and no envelope, is a miss even though Apostra accepts it. From AdCP 3.1 the report is the result of an MCP webhook envelope.
  • A report sent as a valid AdCP 3.1+ envelope is not a miss.
A report refused with 413 (more than 250 media buys) or 503 (Apostra could not record it) says nothing about the format and is not counted. One or more misses show as a flag on the Agent page. From 1 December 2026, two reports refused with 400 that Apostra received on or after that date, within seven days, make the Source a blocker, which stops the agent’s certification pass counting until the window moves past them (see Behaviour on live traffic for what that stops). Retries count, because each one is refused. A report in the AdCP 3.0 shape that Apostra accepts stays a flag however many arrive. Before 1 December 2026 the finding is never a blocker and stops nothing. The finding shows the counts, the first failing field, and refusedRetries. For a Source the field is the one most refused reports failed on first. For the whole agent it is the field of the Source with the most misses. Retried refusals. AdCP says a sender does not retry a refused report, because the same body gets the same answer. refusedRetries counts the refused reports that repeat an earlier refusal from the same source, with the same first failing field and the same media buys, within 15 minutes. A source that keeps sending the same refused report shows a high count, which is a second problem to fix besides the report itself. Usual causes.
  • The envelope leaves out a field AdCP requires, such as operation_id, or a delivery row leaves out by_package. Extra fields never stand in for a required one.
  • The agent still sends the AdCP 3.0 body, with no envelope.
  • The agent retries a 400 in a loop, instead of fixing the report.
Next step. Send every delivery report as an AdCP 3.1+ envelope with task_type: "media_buy_delivery" and the report under result, fix the field the finding names, and send the corrected report rather than retrying the refused one. The finding clears as the seven-day window moves on. See Deliver reporting to Apostra for the full shape of a report.

Creative assignments answered finding

When Apostra sends a campaign’s creatives to a media buy it can ask your agent to assign them to packages in the same sync_creatives call, by sending assignments. AdCP lets a seller refuse that, and have the buyer assign through update_media_buy instead, but only when the refusal says so in a form a buyer can act on. This check reads how your agent answers. Applies when Apostra sent your agent a sync_creatives request that carried a non-empty assignments list and your agent answered it, in the last seven days. A source that was sent none is not judged. Threshold. For each Source, over seven days, Apostra counts the answered requests that carried assignments:
  • A request your agent accepted is not a miss.
  • A request refused with an error whose code is UNSUPPORTED_FEATURE, whose field is assignments and whose recovery is correctable is not a miss. That is the answer AdCP allows. Apostra sends the creatives again without the assignments and assigns them through update_media_buy.
  • A request refused with an error that names the field assignments any other way is a miss: another code, or a recovery that is not correctable.
A failed sync whose error does not name assignments says nothing about this check and is not counted. Apostra compares the error’s code, field and recovery as your agent sent them, and does not read or store its message. While this is a finding, one or more misses show as a flag on the Agent page. It is never a blocker today, so it stops nothing: no buy is held back and no certification pass stops counting. The finding shows the counts of requests checked, accepted, refused as AdCP allows, and refused any other way. A seller that accepts the request and silently drops the assignments is a different problem, judged by its own finding: Package assignments are kept on the buy, described below. Usual causes.
  • The agent refuses assignments with INVALID_REQUEST or another code instead of UNSUPPORTED_FEATURE.
  • The agent refuses with UNSUPPORTED_FEATURE but names no field, or marks the refusal terminal, so a buyer cannot tell it can assign another way.
Next step. Take package assignments on sync_creatives, or refuse them with UNSUPPORTED_FEATURE, recovery correctable, and the field assignments, with a suggestion that the buyer assign through update_media_buy packages[].creative_assignments. The finding clears as the seven-day window moves on. See Answer package assignments on a creative sync.

Package assignments are kept on the buy

A seller can accept a sync_creatives request that carries assignments and ignore the field, leaving each creative unassigned with no error. Only a read of the buy shows that, so this finding pairs two records of the same Source. Applies when Apostra sent your agent a sync_creatives request with a non-empty assignments list, your agent accepted it, and Apostra then read the buy back with get_media_buys at least 30 minutes after the answer. The first such read is the one judged, so an agent that assigns after answering has half an hour to do so. Threshold. For each Source, over seven days, Apostra counts the accepted requests it could read back. A request is read back when that read lists packages. It is a miss when a creative the request assigned to a package is not held there: the read does not list that package at all, or it lists the package and the package says which creatives it holds in creative_approvals and the creative is not among them. Only intact evidence counts. A request whose stored copy was cut at the ledger’s size limit is not judged. A read that did not keep your packages (more than 50, or over the size limit) is not judged, and neither is a read that lists no packages. A listed package that does not list its creatives is not judged, because an agent that does not report them has not said it dropped them. While this is a finding, one or more misses show as a flag on the Agent page. It is never a blocker today, so it stops nothing. The finding shows how many requests were read back and how many missed. Apostra compares creative and package ids only, and does not read or store your agent’s text. Usual causes.
  • The agent accepts assignments on sync_creatives and ignores them.
  • The agent assigns the creatives later than the 30 minute grace.
  • The agent holds the assignments but does not list them in creative_approvals on get_media_buys.
Next step. When you accept a request that carries assignments, put each creative in its package, and list the creatives a package holds in creative_approvals. If you do not take assignments on this call, refuse them with UNSUPPORTED_FEATURE, recovery correctable, and the field assignments instead of accepting and ignoring them. The finding clears as the seven-day window moves on.

Creative approvals use buyer ids finding

get_media_buys can list, on each package, the review state of the creatives it holds, in creative_approvals. AdCP names the creative in each item by creative_id, and that is the buyer’s id: the creative_id on the buyer’s sync_creatives request or in its creative_assignments. Your agent’s own id for the creative is platform_id on the sync_creatives answer, and a buyer never sends it back. A package that lists your own id in creative_approvals reports a review result the buyer cannot match to any creative it holds, so an approval or a rejection never reaches the buyer’s creative. This check reads both sides of the exchange: the creative ids Apostra sent your agent, and the ids your agent lists back. It is the reverse of Package assignments are kept on the buy, which asks whether a creative Apostra assigned shows up in creative_approvals. This one asks whether an id in creative_approvals is one Apostra sent. Applies when Apostra placed a buy with your agent in the last 60 days, and read it back with get_media_buys in the last seven days, and the buy’s packages list creative_approvals. A Source with no such buy is not judged. Threshold. For each Source, over seven days, Apostra judges each buy on its latest read that kept its packages. A buy is a miss when a creative_id that its packages list in creative_approvals is not a creative Apostra sent your agent. Apostra counts a creative as sent when it appears in:
  • a sync_creatives request, in creatives or in assignments; or
  • a create_media_buy or update_media_buy request, in the creative_assignments or the inline creatives of a package in packages or new_packages.
A request counts however your agent answered it, because your agent received the id. A buy counts once, however many approvals or reads disagree. Only the latest read is judged, so an agent that fixes its ids stops being counted as soon as Apostra next reads the buy. Only intact evidence is judged. These do not count:
  • A read that did not keep your packages (more than 50, or over the size limit), a read that lists no packages, and a package that lists no creative_approvals. An approval with no creative_id is not judged either.
  • A buy older than 60 days, and a buy Apostra did not place, such as a buy found on your side. Apostra keeps the body of what it sent for 90 days, so an older buy may carry a creative sent before that, and a buy Apostra did not place may carry creatives it never sent.
  • Any buy of a buyer whose requests to your agent include one whose stored copy was cut at the ledger’s size limit, or has aged out, because that request may have carried the creative.
Apostra matches creative ids for the buyer and your Source, and does not tie an approval to the account a creative was sent under, so a creative sent under another account of the same buyer counts as sent. Apostra compares ids only, and does not read or store your agent’s text. While this is a finding, one or more misses show as a flag on the Agent page. It is never a blocker today, so it stops nothing: no buy is held back and no certification pass stops counting. The finding shows how many buys were judged and how many missed. Usual causes.
  • The agent lists its own id for the creative, the one its ad server or creative library assigned, in creative_approvals instead of the buyer’s creative_id.
  • The agent lists an id it builds from the buyer’s, for example with a prefix, a suffix or a different case.
Next step. In creative_approvals, set creative_id to the id the buyer sent. Keep your own id in platform_id on the sync_creatives answer, or in ext on the approval. The finding clears as soon as Apostra reads the buy and its latest read lists the buyer’s ids. See Return the buyer creative id in creative approvals.

Creative assignments on buys answered finding

A buyer can say which of its creatives run on each package of a buy. Apostra sends that as AdCP’s package creative_assignments on create_media_buy and update_media_buy, one whole list per package, on packages or on the packages an update adds in new_packages. Each item names a creative_id, and may also carry weight, rotation_mode, group_id, sequence_position, placement_refs, placement_ids, landing_page_url and trackers. AdCP says how a seller refuses an item it cannot take, so that the buyer knows what to change. This check reads how your agent answers. Applies when Apostra sent your agent a create_media_buy or update_media_buy whose packages, on packages or new_packages, carried a non-empty creative_assignments list, and your agent answered it, in the last seven days. A Source that was sent none is not judged. An update_media_buy is judged against the newest action list your agent returned for the buy on get_media_buys at or before the update. It is not judged when that list does not take update_creative_assignments (a coarse sync_creatives action covers it): your agent had said it does not take creative assignment changes on that buy. An older list that took the change does not count once a newer one leaves it out. Apostra reads the action list the way it does before it changes a live buy, so this check and that gate agree. An update with no action list read by then is judged. Threshold. For each Source, over seven days, Apostra counts the answered requests that carried creative_assignments. A refusal is read from its error’s field, which names one item and one of its fields as a path into the request, such as packages[0].creative_assignments[1].weight, or new_packages[0].creative_assignments[0].weight for a package the update adds:
  • A request your agent accepted is not a miss.
  • A request refused with VALIDATION_ERROR whose field is the path of an item the request sent is not a miss, when that item carried more than a creative_id. That is AdCP’s answer to an item that asks for something your platform cannot honour, such as a weight, a rotation or a placement.
  • A request refused with UNSUPPORTED_FEATURE whose field is the path of an item’s landing_page_url or trackers is not a miss, when that item carried more than a creative_id. That is AdCP’s answer from an agent that does not declare creative.supports_assignment_landing_page or creative.supports_assignment_trackers.
  • Any other refusal whose field names creative_assignments is a miss. That includes refusing an item that carries only creative_id, which asks for nothing an agent can decline, refusing with another code such as INVALID_REQUEST, a field that does not name one item the request sent, such as packages[0].creative_assignments for the whole list, and a recovery other than correctable.
Both refusals AdCP allows carry the recovery correctable, because the buyer can change or drop the field. An error that leaves recovery out counts as correctable, because that is AdCP’s own recovery for both VALIDATION_ERROR and UNSUPPORTED_FEATURE. An error that says terminal or transient is a miss. The declarations match behaviour finding leaves out exactly the refusals this check judges and accepts, judged here against the request your agent was sent, so a refusal that is not a miss here is never counted against your agent there. A refusal that is a miss here, such as one of an item that carried only creative_id, is not left out there. Nor is a refusal of an update this check does not judge because your action list left out update_creative_assignments: that list said your agent does not take the change, so the declarations finding still reads the refusal. Apostra matches each part of the path exactly: packages or new_packages, the package’s position, creative_assignments as a whole name, then the item’s position, then one of the item’s AdCP fields. A path into new_packages names an item of the packages the update adds, never one of packages. A field such as packages[0].not_creative_assignments does not name creative_assignments. A failed request whose error does not name creative_assignments says nothing about this check and is not counted. Nor is a refusal with an AdCP code about the creative, such as CREATIVE_NOT_FOUND for a creative your agent does not hold or CREATIVE_LIVE_UPDATE_UNSUPPORTED, or about the buy’s state (INVALID_STATE), whatever field it names: those answers are about the creative or the buy, not about how your agent takes the assignment. The finding still shows how many there were. A request whose stored copy was cut at the ledger’s size limit is not judged. Apostra compares the error’s code and field as your agent sent them, and does not read or store its message. While this is a finding, one or more misses show as a flag on the Agent page. It is never a blocker today, so it stops nothing: no buy is held back and no certification pass stops counting. The finding shows the counts of requests checked, accepted, refused as AdCP allows, refused any other way, and refused with a code about the creative or the buy. It applies to every sales agent on the same terms, Apostra’s own included. Whether your agent keeps the item fields it accepts, such as a weight or a landing page, needs a read of the buy and is not judged by this finding. Usual causes.
  • The agent checks a package against a fixed list of fields it handles, and refuses creative_assignments as an unknown field, often with INVALID_REQUEST.
  • The agent refuses an item it cannot honour with INVALID_REQUEST or UNSUPPORTED_FEATURE instead of VALIDATION_ERROR.
  • The agent names no field, or a field that does not say which item and which part of it it refused.
Next step. Take package creative_assignments on create_media_buy and update_media_buy, on packages and new_packages. When an item asks for something you cannot honour, refuse it with VALIDATION_ERROR and a field that names it, such as packages[0].creative_assignments[1].weight. When you do not take a landing page or trackers on an item, refuse them with UNSUPPORTED_FEATURE and a field that names landing_page_url or trackers. Never refuse an item that carries only a creative_id, except with an AdCP code about the creative, such as CREATIVE_NOT_FOUND. If you do not take creative changes on a live buy, leave update_creative_assignments, and the sync_creatives action that covers it, out of the action list you return on get_media_buys. The finding clears as the seven-day window moves on. See Answer creative assignments on a media buy.

Reachability finding

Applies when Apostra called your agent in the seven days and the call was answered or failed with no response. A Source Apostra did not call is not judged, and an idle agent is never unreachable. Threshold. For each Source, over seven days, Apostra looks for stretches in which it called your agent and got no response, and no call in between was answered. A call gets no response when the connection is refused or reset before any response, or when nothing answers before the response deadline. A call your agent answered does not count, even when the answer was an error or the connection dropped after your agent began to answer. A call that is still pending, or that Apostra did not send, counts for nothing either way. A stretch runs from the first call with no response to the last one before your agent next answers. It ends early if more than an hour passes with no call at all, because that time is idle, not downtime. It counts only the time Apostra was calling, so a single failed call, or two failed calls with hours of no calls between them, is a stretch of no length. A stretch that is still going when the window ends counts as it stands. A stretch of more than 15 minutes is an outage. One outage is a flag. Two within the seven days are a blocker, and so is a single stretch of six hours, so an agent that stays down does not rest at a flag. A rolling deploy or restart that leaves your agent unanswering for a few minutes does not reach the line: across every agent Apostra calls, in the 30 days to 28 October 2026, the longest stretch was 2 minutes. call_success_rate does not count these calls. It counts only calls your agent answered, so a connection failure never moves it. Usual causes.
  • A deploy or restart that stops every instance before a new one accepts connections, or that shuts an instance down while calls are still arriving.
  • An agent, ad server connection or host that was down, or a certificate or DNS record that expired.
  • A firewall, rate limit or network rule that drops Apostra’s connections.
  • A request that hangs until the response deadline because the agent never answers it.
Next step. Bring the agent back, and check that it answers a get_adcp_capabilities call. For deploys, keep at least one instance accepting connections, stop sending new calls to an instance before it shuts down, and let in-flight calls finish. The finding clears as the seven-day window moves on.

A sandbox account that serves

Before each buy, Apostra calls your sync_accounts with its certification test identity: for example brand acmeoutdoor.example, operator pinnacle-agency.example and sandbox: true. Apostra asks for a billing model your agent offers: agent first, then operator, then advertiser. Your agent is ready when:
  1. sync_accounts accepts that identity and returns the account as status: "active" with sandbox: true and an account_id. Apostra uses that account_id for the creative and the buy, and never falls back to a default account.
  2. Buys on that sandbox account serve through your real ad server, at no cost to Apostra. The sandbox account that serves through your ad server is the test line. You do not set up a separate test placement or name a trafficking contact.
A sandbox that only answers the protocol cannot certify a tracked use case. AdCP says a sandbox account performs no real platform calls, so if yours stops there, nothing serves and Apostra’s tracker sees nothing. The buy goes live, but the tracked checks never get evidence, so the use case stays inconclusive. Sellers that set account.require_operator_auth: true are certified like any other: Apostra runs certification on the credential you connected, which is the credential a storefront calls your agent with, so the storefront is the operator. Certification used to skip these agents as seller_operator_auth_unsupported, and runs from before 19 November 2026 still show that reason. The sandbox account above is for test buys, where your agent is the seller to Apostra’s test buyer. The request your agent receives when it runs behind a publisher’s own storefront is judged separately, as the accounts use case. Channels Apostra’s tracker cannot observe, such as linear TV, radio, print and DOOH screens, certify on your reports instead. See Use cases.

Certification-ready sandbox

Certification runs an actual campaign on your actual ad server. So your sandbox needs real traffic for reporting and testing to work: a sandbox that serves nothing gives Apostra no impressions to compare and no reports to read. Bring your own traffic. Apostra does not send traffic to your inventory, and cannot. A certification buy runs for three days, so it needs enough real requests in that time to serve and fire Apostra’s tracker. If your inventory serves rarely, for example a placement that serves only when an AI agent opens the page, drive traffic to it yourself while the test buy runs, and click the ad at least once if the format takes clicks. Turn on automatic approval for certification creatives. Apostra recommends it so a run never waits on you. Your agent and ad server still apply their own checks to the creative. Apostra skips only its own creative review. Manual approval works too, and the run waits for it. When your agent holds the certification creative for review, Apostra does not end the use case. It shows a finding, “Your agent is holding our certification creatives for review, so certification is waiting for you to approve them”, on the Agent page and the Overview page, and in the agent’s findings. Its whatWeSaw names the use cases that wait and the date the wait ends. An agent with no review step does not wait. If your agent accepts the certification creative on sync_creatives (action created, updated or unchanged, with no errors) and gives no review status, Apostra creates the test buy straight away. AdCP has an agent with no review lifecycle leave status out, so Apostra reads that answer as “nothing to review”. An agent that does review creatives must say so: answer pending_review or processing while it reviews. How the wait works:
  • How Apostra checks. Every four hours it reads the creative with list_creatives, filtered to that creative and the account the sync used, because AdCP makes list_creatives the authoritative review state. It does not send sync_creatives again just to ask: an agent that replays a repeated request returns its first answer, which would never change. The listed status decides: approved continues the run, pending_review or processing keeps waiting, and rejected ends the use case. If your agent cannot list creatives, does not list this one, or lists it with another status, Apostra sends sync_creatives again for the same creative and reads its answer as above.
  • How long. Apostra waits up to five days from when the run created the use case. Five days reaches a creative submitted on a Friday evening and reviewed on Tuesday.
  • When you approve in time. Apostra creates the test buy on its next check, within about four hours of your approval. The three-day flight and the 14-day collecting window count from that moment, not from when the run began.
  • When you do not. The use case ends inconclusive with seller_creative_review_required, and the finding goes away. Apostra then starts another run, which waits again.
  • When you reject it. The use case ends at once with your agent’s rejection reason, as creative_sync_failed.
While a run waits, the Source’s earlier result keeps its seven-day evidence window. A slow approval can let that earlier result expire before the new buy has evidence. Automatic approval avoids the gap.

Use cases

Apostra reads your products and capabilities to decide which use cases apply. You do not maintain the list. A use case your agent does not declare does not apply, and one it declares but has not passed cannot take buys yet. A seller’s demand coverage reads its products against this same table, so it shows which of these use cases the seller’s products offer before any run. The use cases judged on the whole agent (priced signal targeting, the compact lifecycle and accounts for a seller’s own storefront) are not on it. A podcast ad inserted at download time fires our tracker when the episode is downloaded, not as it plays, and an episode downloaded before a stop or a change can be played days later. So a podcast product is not judged on quartiles, on stopping serving or on changes taking effect, and every other check applies. The extra checks the spec lists for cost per completed view, conversions and TMP are being added. Today a run for those use cases is judged on the standard checks. Priced-signal targeting is reported separately and does not decide certification. Where package impressions, spend and fixed prices are available, it also checks whether spend includes one signal charge. The comparison uses every ended reporting period since the buy went live, and rounds both reported and expected spend to the currency’s minor unit. Missing, late or ambiguous figures leave that check inconclusive. The test selects a binary signal priced in the buy’s currency, preferring CPM and then percentage pricing on fixed-price media; flat-fee signals are not selected. It skips the extra buy when its maximum cost cannot be bounded. A priced-signal buy shares the run’s five-buy limit, so another family may wait for a later run. The base use cases determine whether a product may take a new buy. Cost per completed view, conversions, audience or signal targeting, and TMP are feature use cases: a pass shows the agent was certified for that feature, but a missing feature pass does not itself block buys. The agent’s findings raise Attention when a storefront offers a use case without a pass that counts, including when it has never run or the result is inconclusive. Allow certification buys and fix any failure to obtain a pass. The compact lifecycle use case is different from the others in two ways. It belongs to the whole agent, not to a product, so it is never split by family. And it is reported beside the other use cases but does not decide certification: a pass does not certify your agent and a failure does not fail it, so it has no effect on Storefront readiness, on a publisher’s connection to your agent, or on any buy. The compact lifecycle becomes required for certification on 1 February 2027: from then, an agent that has not passed it does not count as certified, whatever its other use cases show. The AdCP 4.0 readiness check for the compact lifecycle reads this use case’s result. Apostra also uses this use case’s result to decide how it buys from your agent. Apostra calls your agent through the compact lifecycle, starting with request_proposals for a proposal round, only once your agent passes this use case. Advertising request_proposals in media_buy.lifecycle_tools does not change how Apostra buys from you. Until the pass, Apostra keeps using get_products. The pass applies while it is the latest decisive result on every active Source that uses your agent, so a later failed run returns Apostra to get_products until the next pass. The accounts use case is for an agent that runs behind a publisher’s own storefront, as a publisher’s ad server does. When the publisher approves a buyer, Apostra asks that agent for the advertiser’s account. Apostra judges every agent on this, whether or not a publisher uses it that way today. Each run sends your agent sync_accounts twice for the certification identity (sandbox), with billing: agent and no payment terms, and with a different idempotency key each time. What Apostra asks your agent for explains why billing is always agent on that request.
  • Pass: both answers return the same active sandbox account with an account_id. The second answer’s action is unchanged or updated.
  • Fail: your agent refuses either request, for example with BILLING_NOT_PERMITTED_FOR_AGENT. It also fails if your agent returns an account that is not active, or no account_id, or a different account the second time. An agent that does not offer sync_accounts fails without being asked. Setting account.require_operator_auth: true does not fail it. Apostra never retries with another billing value, including the one a refusal suggests.
  • Inconclusive: your agent holds the account for approval, or Apostra’s request does not complete.
This use case does not decide certification. A pass does not certify your agent, and a failure does not fail it or stop any buy. It runs on each inventory source with that source’s own credential, because accounts belong to the publisher, so each source has its own result. Apostra also runs this check on its own, with no buy, on each inventory source of a publisher’s storefront that sells through its own sources as well as your agent, including one still waiting to go live. It does this whether or not you asked for certification: daily until the source passes, then weekly. On such a storefront, a new inventory source for your agent does not go live until it passes; a source that is already live keeps running. On a publisher’s storefront that sells through its own sources as well as your agent, a failed result shows on that inventory source’s health and as the certification.storefront_accounts finding on the Agent page.

Buyer’s own audience

A buyer can bring a first-party audience: a list of hashed identifiers sent with sync_audiences, which the buyer then targets with audience_include in a package’s targeting_overlay. This use case checks that your agent takes that audience and does what AdCP says with it. It belongs to the whole agent, so it is never split by family. It applies when get_adcp_capabilities declares media_buy.audience_targeting; AdCP says that block means your agent supports sync_audiences and audience_include and audience_exclude in a targeting overlay. It is report-only. It shows beside the other use cases and does not decide certification: a pass does not certify your agent, and a failure does not fail it or stop any buy. Making it count toward certification is a separate decision that Apostra will announce, with a date, before it takes effect. Apostra does not send these calls yet. The use case and its three checks are defined, and your agent is told which of its declarations make it apply, but a run does not yet sync a test audience or create a buy that targets one. Until it does, the use case shows no result. Three checks judge it, each against the AdCP schemas: Delivery is not judged yet, and is never shown as a pass. AdCP has no test audience that matches at a known size, and Apostra cannot use real personal data in certification. A synthetic list matches nothing, so the audience stays too_small on platforms that set a minimum audience size. The three checks above can be proven through the too_small path. Apostra cannot yet prove a ready audience, a buy that targets one, or its delivery, and records delivery as not provable until AdCP defines a test audience. How this differs from audience or signal targeting. The Audience or signal targeting use case also applies when your agent declares media_buy.audience_targeting, but a run judges it on the standard checks and, for signals, the priced-signal buy. Nothing in it sends your agent a buyer’s audience. This use case does, and it leaves the existing use case’s applicability and results unchanged. What your agent must do is in Take a buyer’s audience, report it honestly and keep or refuse it. A use case is also split by product family. Display on one ad server and display on another are two families, and each needs its own certification buy. Programmatic and non-programmatic DOOH are always separate families. Each format kind on a DOOH product is its own family and gets its own certification buy. Apostra places one certification buy for each family, and a family is one format option: two options of the same kind that differ in size or parameters are two families, each with its own buy. Options that differ only in their format_option_id, with the same kind and parameters, share one family and one buy. The buy names the option of the product it buys, so the result says the shared format works for every product the family stands for, not that each option id was tried. A format no use case names, such as agent_placement, is a family too. A family buys the first of its products in the order your agent lists them, and its result stands for every product that offers the same option on the same terms. A use case that Apostra judges with its own tracker, such as conversion tracking or audience targeting, never claims a path where the tracker cannot fire. A DOOH-only agent that sells image and video_hosted therefore gets two buys, even if it also declares those capabilities. Conversion tracking, audience targeting and the other capabilities you declare add their checks to the buy of the family they apply to; they never place a buy of their own.

How many buys a run places

A run places at most 5 certification buys: a run spends at most USD 50, and each buy is at most USD 10. Most agents have one to five families and never reach the limit; for them nothing changes. When an agent has more than 5, the run chooses which 5 to buy, in this order:
  1. Families no run has tried yet.
  2. Families whose latest result failed, and families a run tried that have no result yet, for example because the run skipped them before buying.
  3. Families whose latest result passed but whose evidence has gone stale, as How certification is kept describes it.
  4. Families whose latest result passed and whose evidence is still fresh.
Inside each group, the family a run last tried longest ago goes first, whether that run bought it or skipped it. A run chooses only as many families as its spending limit pays for, so the limit does not skip a family it chose. A family whose cheapest buy costs more than the USD 10 per-buy cap does not take one of the 5 places: the run records it as budget_exceeds_per_buy_cap and buys the next family in the order instead. The exception is a run that also checks that your agent refuses a buy with no creatives (Create without creatives finding): that check books one more buy from the same limit, so the last family the run chose can be skipped for the limit. A family skipped only because the run’s spending limit ran out has not been tried, so it keeps its place in the order for the next run. The families left over are bought on later runs in this same order, so every family comes round. A family that is skipped before a buy every time (for example, because certification has no creative for its format) never gets a result, so it waits its turn with the failed families instead of taking a place on every run and pushing the others out. A family that passed with fresh evidence waits only when the limit forces something out; with room to spare, it is bought again so its evidence does not expire.

Formats certification does not test

A product can offer several format options, and a run plans one buy for each distinct option. When a run plans no buy for a format option your products offer, Apostra records the option with the reason and shows it as the certification.format_not_tested finding on the Agent page. The finding names each option, its product and channel, except an option with the reason recently_passed, which is not untested. It is an attention finding: it does not stop buys or fail your certification, and the options it names stay bookable. It does mean Apostra has no evidence those formats work. When a certification buy carries a creative, each package assigns it in creative_assignments and names the one format option that creative is meant to fill. AdCP treats a package that names no format as asking for every format on the product, so naming the option tells your agent that the one creative covers the package. Once that option has its creative, your agent should start the buy, not hold it in pending_creatives for the product’s other formats. How the option is named depends on whether your agent takes canonical formats on create_media_buy:
  • Your agent takes canonical formats, for example because it declares media_buy.features.canonical_creatives: true in get_adcp_capabilities (see Getting ready for AdCP 4.0): the package names your format_option_id in format_option_refs.
  • Otherwise: the package names the legacy format_ids entry your product declares for that option, for example in the option’s v1_format_ref.
  • The option has no legacy entry of yours, and your agent does not take canonical formats: the package names no format, so your agent may hold the buy in pending_creatives until every format on the product has a creative. Declaring canonical_creatives: true lets Apostra name the option.
A family planned from one of your legacy formats names that format_ids entry, as before. For each family, Apostra chooses the format option its test creative fills from your product’s options of the family’s own format kind, in the order your product lists them, whether Apostra’s tracker or your reports judge the family. A creative of another kind is never used, even when your product also offers that kind: an image is never bought for a video family. A family that stands for one of several options of a kind, or for a kind no use case names, chooses only from its own option. It leaves out any option the creative’s own facts conflict with, as @adcp/sdk judges them: for example a VAST 3.0 tag and an option that declares vast_version: "4.2". Of the rest it takes the first option the creative meets exactly, or else the first option it can be adapted to, syncs the creative to your agent under that option, and names the same option on the package as described above. When the creative’s facts cannot settle the choice, for example a VAST tag whose length is not measured, the package names the option only when one is left. If none of Apostra’s test creatives fits any of them, the family is skipped before any buy as buy_creative_channel_unmatched, which is Apostra’s gap, not yours. The list comes from the latest run on each of your active inventory sources. Runs started before Apostra recorded this show no finding.

Products buyers cannot see

Before a storefront answers get_products, it checks each product against the AdCP release the buyer is served: 3.2, unless the buyer pins another release. A product that fails that release’s product schema is left out of the answer, so no buyer at that release sees it or can book it. The answer’s errors list names the product in a PRODUCT_UNAVAILABLE entry, with the field that failed (see AdCP versioning). A certification run reads the same answer a buyer gets. It records each product of its inventory source that the answer left out, with the release and the field that failed, and the Agent page shows them as the certification.product_withheld finding. Each product is named by its id in your agent. It is an attention finding: it does not stop buys or change a certification result, and the run certifies your other products as usual. A storefront answer names at most 20 of the products it leaves out, so the finding says when there may be more. To clear it, make each product valid at the release named. The field is shown as a JSON pointer into the product, for example /format_options/0/format_kind. Correct the product in your agent, and the next certification run reads it again. If the product already looks valid to you, tell Apostra: a storefront serves your products through its own projection, and a defect there can leave out a product you declared correctly. The finding comes from the latest run on each of your active inventory sources. When the storefront leaves out every product of a source, certification cannot start a run there at all. The finding then comes from that refused automatic start, names the products the same way, and says certification could not start on the source. Runs started before Apostra recorded this show no finding.

How Apostra measures a display tag

A display tag is HTML that Apostra sends your agent, so Apostra can carry its impression pixel inside the tag. That is how a display_tag format with no tracker slot is still measured, provided the format takes the tag as inline markup: its supported_delivery_types lists inline_markup, and its slot is the HTML the tag fills. AdCP assumes a tag URL only when a format declares no delivery types, so list inline_markup explicitly. For the certification buy only, Apostra adds one 1x1 image to the end of the markup it sends in sync_creatives. The image is Apostra’s impression tracker for that buy. It carries CACHEBUSTER, DOMAIN and APP_BUNDLE for your agent to fill where it serves the tag, so the macro check judges them on the tag. Apostra fills its own identity in the pixel, such as the media buy, before it sends the tag. The creative in Apostra’s library is not changed, and a tag Apostra syncs for a real buy is sent exactly as the buyer supplied it. If the format declares an impression-tracker slot, Apostra uses the slot and adds nothing to the markup, so an impression is never counted twice. A format that takes its tag as a URL (a required tag_url slot and no HTML slot) is sent a tag Apostra hosts: Apostra stores its certification test creative as an HTML page and puts that page’s https URL in tag_url, for your agent to serve in an iframe. Apostra cannot add its pixel to a page served from a URL, so such a format is measured through its impression-tracker slot. Without one it is skipped as seller_format_cannot_carry_tracker until the format also takes inline markup or declares a tracker slot. A format whose supported_tag_types leave out iframe expects a script or a redirect at that URL, which Apostra cannot host, so it is skipped as buy_creative_channel_unmatched, Apostra’s gap.

What each check judges

The pause, stop and resume checks run on the first certification buy, so an agent whose buys cannot be stopped is not certified. The live creative changes run on the same buy after the resume, and are a finding only.

Hits the tracker checks set aside

The tracker checks judge only hits a viewer’s browser or app sent. Two kinds of hit are counted apart and never judged. The clicks check says how many clicks it set aside, and the stop and resume checks say how many review impressions:
  • Known crawlers. A crawler such as Google’s ad-review bot opens your creative’s click-through URL without an ad being served, so the click’s macros arrive empty. We recognise these by their user agent.
  • Google’s ad review. When Google Ad Manager reviews a creative, it renders it on Google’s own pages and clicks it, which is not serving on your inventory. The render reports its site as google.com for display or doubleclick.net for video, and its click follows within two minutes with no site. A hit whose site is one of those domains is set aside, as is a click with no site or app that follows such a render of the same creative within two minutes and carries either no cachebuster or the render’s own cachebuster. If you declare google.com or doubleclick.net (or a subdomain of either) as one of your properties, hits on that domain are judged as delivery as usual.
Set-aside hits count toward no check: not impressions, clicks or quartile events, not the macros judged on them, and not a paused buy’s silence. So a review render during a pause does not fail the stop check.

Reports the check does not count

When the reports check sets one of your delivery reports aside, its result says how many reports it did not count and why, in the words in the first column. A report set aside counts as no report for its period, so the period can still be missed. Fix the reports your agent sends; the next report that counts is judged as usual. A buy is owed the whole of every period of the product’s schedule that its flight touches, including the period it went live in. AdCP puts a report’s reporting_period on the schedule’s period boundaries, so report that first period from its start, not from the moment the buy went live. A report on the boundaries counts for the buy’s periods it covers even when it also covers time before the buy went live or after it stopped, such as a whole-day report for an hourly product on the day the buy started. A report that covers only time before the buy went live, or only time after it stopped, is about no period the buy is owed, so the check neither counts it nor lists it. Apostra’s first read of a buy it polls asks for the day before the buy went live, and that read is one of these. Seller-reported channels skip the tracker checks. They pass on the buy going live, on your agent echoing our context, and on reports arriving on the schedule your agent declares in reporting_capabilities. A seller-reported buy carries a test creative, as every certification buy does: Apostra never places a buy without one (see A family Apostra has no creative for). For a display, video, audio, CTV, DOOH, out-of-home, radio, print or influencer product, Apostra looks among the creatives it offers for that channel (on out-of-home, radio, print and influencer, the one creative a run offers to every channel) for one whose format kind, size and length match one of the formats your product declares exactly (a hosted video for a video_hosted family, an image for an image family). A creative that would need resizing or trimming does not count while Apostra can make one that matches. It syncs the creative to your agent and assigns it to the buy, so an agent that holds a buy until it has a creative can activate it. Apostra does not check that creative for its tracker macros, because a seller-reported buy is not judged on them. If none of the creatives Apostra offers matches exactly, Apostra makes one for an image, html5, audio_hosted, video_hosted or video_vast format. It uses the first format your product declares that it can make: the fixed size, the first declared size, or a standard display size inside your size limits. For audio it makes a 15 or 30 second tone, or, when neither fits your length limits, a tone at the nearest limit. It makes images and HTML5 creatives up to 7680x4320 pixels (no side above 8192) and up to two minutes of audio. Out-of-home gets the same images and videos as DOOH, radio and podcasts the same spots as streaming audio, print an image at the page size your format declares, and influencer an image or a video at the size and orientation your format declares. For video_hosted it uses one of nine standard videos: 1920x1080 (16:9), 1080x1920 (9:16) or 1080x1080 (1:1), each 6, 15 or 30 seconds long. Each is a flat grey picture with a silent AAC audio track (H.264, 30 frames per second, a keyframe every two seconds), exactly as long as its name says. Apostra takes the fixed size or the first declared size your format lists when it is one of those three. Otherwise it takes the first of the three that has your orientation and aspect_ratio and sits inside your size limits. For length it takes your exact length when it is 6, 15 or 30 seconds, else the first of 15, 6 and 30 seconds that your length limits allow. Apostra reuses each creative it makes for later buys. For video_vast it hosts the same standard video and writes a VAST tag that plays it: one inline <Ad> with a single linear creative, a click-through to https://scope3.com/, and an MP4 media file. The tag is stored in Apostra’s asset storage like any VAST tag a buyer saves, but your agent receives a URL on Apostra’s API, of the form https://api.interchange.io/delivery/v1/certification-vast/{id}, and Apostra’s trackers are added around it exactly as they are for a buyer’s own tag. That URL answers a browser player’s cross-origin request the way Google’s IMA SDK requires: it echoes the page’s Origin in Access-Control-Allow-Origin, sends Access-Control-Allow-Credentials: true and Vary: Origin, and answers the OPTIONS preflight. It needs no credentials and serves only the tags certification writes. The size and length come from the same rules as video_hosted, using your format’s aspect_ratio, orientation, size limits and exact length or length range. The tag is written as VAST 4.2, or as the vast_version your format declares when that is 4.1, 4.2 or 4.3. Any other declared version, including 4.0, gets 4.2. Its own Impression element, which VAST requires, points at a host that does not resolve, so a player sends nothing anywhere for it; the impression Apostra measures is the one it adds. For an agent_placement or native_in_feed format, Apostra builds the creative from the slots the format declares, by asset type: copy for the certification advertiser, Acme Outdoor, in each text slot, cut to the slot’s length limit or to the limit your format’s params set for it (title_max_chars, body_text_max_chars, cta_max_chars), whichever is stricter, and in a call-to-action slot the first of your cta_values that fits that limit, in place of Apostra’s own copy; Apostra’s click URL in a click slot; a generated image in an image slot, at the size the slot declares; and Apostra’s impression tracker in an impression-tracker slot, so a format that declares one carries Apostra’s tracker. For a native_in_feed format, which states its image sizes as format parameters, the main image is built at the first of your main_image_sizes and the icon at your icon_size. An icon is left out when your format declares no icon_size, and an image is left out when your image_formats does not include PNG, the format Apostra generates. Apostra uses a creative it built for one of your formats for another only when each of its images is a size that format declares for that image; otherwise it builds the other format its own creative at your declared size. A required slot Apostra cannot fill, such as a brief, a repeatable group of assets other than a carousel’s cards, or a VAST tag, skips the family before any buy as buy_creative_channel_unmatched, and the result names the slot and what stops Apostra. An optional slot it cannot fill is left empty and recorded on the family as an advisory, with the checks it would have made possible, and no result claims those checks. Certification plans these formats only on out-of-home, radio, print and influencer today, so the same format on a display product is not bought yet: the run lists it as format_not_covered. Apostra reuses a creative it built this way, including for a carousel, for later runs on the same format. When the format’s declaration changes, or Apostra changes how it fills the slots, the next run builds a new one. For an image_carousel, Apostra builds AdCP’s cards group with as many cards as your format requires, and never fewer than AdCP’s minimum of two. Each card carries a generated image, the headline “Gear up for every trail” cut to your card_headline_max_chars, and Apostra’s click URL as its landing_page_url. Your agent receives them as AdCP card objects in the cards asset, alongside the carousel’s primary_text and landing_page_url. Each card image is sized from your format’s card_aspect_ratio, at 1080 pixels on the short side (4:5 gives 1080x1350), or 1080x1080 when your format declares no ratio. AdCP gives a card no pixel size, so a size your agent declares under another parameter is not read; if your agent needs a particular card shape, declare it as card_aspect_ratio. A carousel that takes only video cards is skipped as buy_creative_channel_unmatched, naming the cards slot. If your agent rejects the carousel, the family is skipped as creative_sync_failed with your agent’s message. For audio_daast, Apostra writes a DAAST 1.0 tag that plays its own test audio and sends podcast families with it. A family on a format Apostra cannot make a creative for, tracked or seller-reported, is skipped before any buy as buy_creative_channel_unmatched. That is Apostra’s gap, never yours: nothing in your product is wrong and there is nothing for you to change, and you do not need to report it: Apostra tracks these gaps itself. The family is tested in a later run once Apostra can supply a creative for the format. Until then the buy use cases for that format stay untested. The result for the family says so and names the format. Certification reads your products as Apostra serves them to a buyer, at the AdCP release the buyer is served, so it plans a product that declares an AdCP 3.2 format such as audio_vast. Apostra does not yet make an audio_vast creative. When your product lists several formats of the creative’s kind and your agent publishes them as legacy format_ids that state no size, Apostra sends the creative under the legacy format ID of the format it matches. It does not leave your agent to pick among them by size. Apostra falls back to an offered creative that would need resizing or trimming only when it cannot make a matching one: a format larger than the limits above, or a hosted video whose size, orientation or length none of the nine standard videos meets (for example 1280x720, or exactly 10 seconds). If your agent rejects the creative, for example because it requires a minimum bitrate (the standard videos are a few tens of kilobits per second), the family is skipped as creative_sync_failed. Apostra applies the tracker checks only where your format declares a place for its trackers. A hosted video or audio format promises that Apostra’s trackers fire when it declares an impression-tracker slot (the slot id contains impression, tracker_pixel or imp_pixel, with asset type url or pixel_tracker), or a playback-tracker slot (start_tracker, first_quartile_tracker, midpoint_tracker, third_quartile_tracker or complete_tracker, with asset type pixel_tracker). For such a format, a buy whose tracker receives no hits fails. A hosted format that declares none of those slots cannot carry Apostra’s trackers, so the tracker checks do not apply to it. It is certified as a seller-reported sandbox buy, on the buy going live, on your agent echoing our context and on reports arriving on schedule. The same hosted formats on DOOH stay seller-reported whatever slots they declare. DOOH screens are certified from your own reporting unless Apostra’s tracker can fire. A DOOH-only agent that sells hosted image and video_hosted formats at screen sizes therefore passes on the same three checks as out-of-home.

A family Apostra has no creative for

Apostra never places a certification buy without a creative. A buy with no creative cannot go live: your agent rightly holds it in pending_creatives, so it delivers nothing, sends no reports and proves nothing. When no offered creative fits a family and Apostra cannot make one for the format it declares on its channel, the family is skipped before Apostra calls your agent at all, including the sandbox account setup, and it spends none of the run’s budget. The skip reason is buy_creative_channel_unmatched, and the family’s result names the format and the channel. That is Apostra’s gap, never yours: nothing in your product is wrong, there is nothing for you to change, and you do not need to report it, because Apostra tracks these gaps itself. A later run tests the family once Apostra can supply a creative for it. A run started with no creatives offered at all skips such a family as creative_sync_unsupported instead. That is a setup gap on Apostra’s side, not yours either: nothing changes on your side, and Apostra support supplies a creative. This covers a channel Apostra makes no creative for yet (linear TV, cinema and product placement), and a format it cannot make on any channel, such as audio_daast or a carousel of video cards. The one buy Apostra sends with no creatives on purpose is the create without creatives finding, which checks that your agent holds such a buy in pending_creatives, and it is cancelled again. A buy Apostra placed without a creative under its earlier rule keeps its result, and the result says the buy had no creative: its reporting result stays inconclusive and never fails, because a buy with no creative cannot deliver. A buy your agent rejected, or an agent that dropped our context, still failed.

Spend stays within the budget

Certification checks that your agent, not Apostra, keeps spend inside the budget it accepted. Each certification buy is a small probe, and the check applies from the day it ships, with no report-only period, to Apostra’s own sales agents on the same terms as yours. Threshold. For a buy booked on a fixed CPM (pricing_model: "cpm" with a fixed_price), Apostra prices the impressions its tracker counts at that CPM, in the currency the buy was booked in. Spend above the package budget plus 20% fails the use case. Spend within it passes once the tracker has seen the buy serve. A buy booked on any other pricing is not judged, because its spend cannot be worked out from impressions. Usual causes.
  • The package has more than one ad-server object behind it, and each can spend the full budget, so the budget is spent twice.
  • A cap is applied as a daily limit that never reaches the total.
  • The agent accepts a package it cannot cap and leaves it to serve.
Next step. Make the objects behind a package unable to spend more than the package budget together. For a package you cannot cap, refuse it, or pause it, and say so with an AdCP status or error rather than leaving it to serve. Report each package’s effective budget as you read it from your ad server, counting every object behind the package.

Create without creatives finding

AdCP treats a media buy created with no creatives as valid. A buyer can create the buy first and assign creatives later. Your agent accepts the buy and holds it in pending_creatives until creatives are assigned and approved. AdCP 3.3 keeps a buy with no creatives there: no package can begin delivery without creative coverage, and pending_start is for a buy whose creatives are assigned and waiting on review or a start date. An agent that refuses such a create turns away a booking AdCP says is fine, and the buyer sees a failed booking instead of a buy that is waiting for its creative. What Apostra sends. Apostra sends your agent one extra create_media_buy in a run. When the run has created a buy, it books the product, pricing option, account and budget your agent has just accepted. When the run has created no buy (none of your products could be planned for a tracked family, or each was skipped), Apostra sends it on its own, on the cheapest product it can book. Either way it is a single package at the smallest budget the product allows, with one difference: no package carries creative_assignments, and Apostra syncs no creative for it. The flight starts a day after the call, so the buy cannot start. The probe never needs a product’s formats, channels or tracker support, so a product Apostra cannot plan a tracked family for can still carry it. When Apostra cannot send it. If none of your products can carry the buy, Apostra records the result as inconclusive with the skip reason no_probe_product; see the skip reasons. That is not a failure of your agent. When Apostra cannot send the probe, a result that already passed stands: it is not replaced by an inconclusive one. How often. Each probe books a real buy on your agent, so Apostra does not send one on every run. It sends one when your agent is first certified, when the capabilities or AdCP version it declares have changed since its last probe, and when its last probe did not pass (so a failed or inconclusive probe is tried again on the next run). After a pass it sends one again once a week. It never sends one on the runs that certify Apostra’s own staging embedded sales agent for a release. What your agent must do. Accept the create: return a media_buy_id and report the buy pending_creatives. The check reads the status your agent itself gave in its answer. What a pass changes for buyers. A passing result is the evidence Apostra uses to book at once a buy whose only creatives are still in your agent’s review. On an agent that has passed, that buy is booked with no creative and each creative attaches when your agent approves it, so the buyer does not lose flight days to the review. On an agent with no passing result for an inventory source the buy routes to (a failed or inconclusive probe, or no probe yet), Apostra holds the buy until your agent has reviewed a creative, because some agents refuse a create with no creative for good. The latest probe decides, so a probe that later fails returns your agent to holding. See When every creative is held. Apostra’s embedded sales agent. A source served by Apostra’s embedded sales agent is not probed. The embedded sales agent declares, in its get_adcp_capabilities response, whether the ad server behind each source accepts a create with no creatives (ext.scope3.create_without_creatives, with the release that declared it), and Apostra reads that declaration. AdCP has no field for it yet; Apostra has proposed one (adcontextprotocol/adcp#8723). A source whose embedded sales agent has declared nothing is judged by its certification result, as any other source. What fails. The check fails, and the provider owns the fix, when:
  • your agent answers the create with an error that says no, for example that every package needs a creative;
  • your agent returns a buy and reports it rejected or canceled straight away;
  • your agent reports it active, paused or completed: a buy with no creatives cannot be running or finished;
  • your agent reports it pending_start: that tells the buyer nothing is missing, when the buy cannot start without creatives; or
  • your agent reports a value AdCP does not define as a media buy status.
The result names the status your agent reported and what AdCP expects. For an error, it shows your agent’s own error code and message as it sent them, cut to 400 characters with links and credentials removed. What is inconclusive. Apostra does not count anything that is not your agent’s answer as a failure: a timeout, an agent that is unavailable or rate limited, a credential or configuration fault, a failure on Apostra’s side before the call reached your agent, a create your agent holds for a decision, and a create whose answer gave no status. When the answer gave no status, a status your agent gave the buy on a read before Apostra cancelled it settles the result. The probe is tried again on the next run. The test buy is always cancelled. Apostra records that it owes your agent a cancel before it sends the create, then cancels the buy as soon as your agent answers. If your agent does not confirm the cancel, Apostra tries again with growing waits, up to twelve times, until your agent confirms the buy ended or says it holds no such buy. A cancel Apostra cannot get confirmed within 12 hours of sending the buy is raised. It is raised as the finding certification.probe_buy_not_cancelled on your Agent page when your agent refused the cancel or kept the buy, and to Apostra’s own team when Apostra’s attempts could not be sent or answered. The buy carries no creative and starts a day after it was sent, so it cannot deliver in the meantime. Severity. This is a finding today. It shows as an attention finding on the Agent page, and it stops nothing: it never makes a use case fail and never withholds a certification pass. Whether it becomes required is a later decision, which this page will announce before it takes effect. It applies to Apostra’s own sales agents and to third-party agents on the same terms. A run that created no buy sends no such create. What is not judged. Only your agent’s answer to the create is judged. A buy your agent accepts and rejects later, after Apostra has cancelled it, is not seen. Next step. Accept a create_media_buy whose packages carry no creative_assignments. Hold the buy in pending_creatives and start it once creatives are assigned and approved. Accept a cancel for a buy that carries no creative. The finding clears the next time a probe is sent and your agent accepts it.

A pause skips finished packages

Certification checks that your agent can pause a buy in which one package has already finished. Each certification buy for a tracked use case books two packages on the same product. The first flies for one hour from the buy’s start. The second starts five minutes after the first ends and runs to the end of the buy. Three hours after the start, once our tracker has heard the buy serve, Apostra pauses the buy, so the pause reaches a package that has finished and one that is still live. Two grace periods after the pause, Apostra resumes it. The check passes when all of these hold:
  1. The pause succeeds. Your agent accepts the pause. A refusal fails the check and the result shows your agent’s own error code and message. A pause that is held for approval, or that your agent never answers, is inconclusive. The pause and the resume are sent the way a buyer sends them, with the optional push_notification_config callback present. This is so even when your agent lists tasks_get, which is when Apostra otherwise leaves the callback off. AdCP defines the callback, and the version field on the request envelope, as optional, so your agent must accept an update that carries them. A refusal that names either field fails the check, and the result says which field and links to Takes buy corrections. Accept the update and ignore a field you do not use, rather than treating every top-level field as a change to make.
  2. Your agent reports paused. On its next read after it accepted the pause, get_media_buys reports the buy paused. Apostra asks for that read soon after the pause. Any other status fails the check, and so does no read within 30 minutes of the pause (inconclusive).
  3. Serving goes quiet. After the channel’s grace, our tracker hears no impression until the resume.
  4. The resume restarts serving. Our tracker hears an impression within the grace after the resume. A refused resume fails the check.
A product whose minimum spend is above USD 5 books one package. A second package would take the buy past the USD 10 per-buy cap, so the buy has one package and no finished package for a pause to skip. Apostra still pauses it, after the same three hours, and resumes it two grace periods later. The pause check does not apply, and the result says the buy has one package. The stop and resume checks apply as usual, and the stop check also needs your agent to accept the pause and report the buy paused on its next read. A refused pause fails the stop check, with your agent’s own error code and message. A use case that our tracker judges is never certified without the stop check judged. A buy that was never paused, such as one booked before this check existed, shows the stop check as not yet judged until Apostra books a new certification buy for it. This is the case that failed on 20 October on an embedded agent: a pause was sent to every package, including one its ad server had already closed, and the whole pause was reported as failed. The buy then showed active with every live package paused. Usual causes.
  • The agent tries to pause an ad-server object that has finished and reports the whole pause as failed when that one refuses.
  • The agent accepts the pause but pauses only some packages, so the buy keeps serving.
  • The agent reports the buy paused before the ad server stops serving, or keeps reporting it active after it has.
  • The resume re-enables a package that has finished and fails.
Next step. On a pause, act on the packages that can still serve and skip any that has finished (canceled, or its end_time has passed). Report the buy paused once every package that can still serve is paused. A resume must restart the packages the pause stopped. The finding clears the next time a certification buy passes.

Deleting an audience stops delivery

Certification checks that deleting an audience on your agent stops the packages that use it. It applies to an agent that declares a partner route for audiences: media_buy.audience_activation in experimental_features, with a route in media_buy.audience_targeting.supported_activation_methods. See Deleting an audience stops the packages that use it for the requirement. How it runs.
  1. Apostra binds a certification audience on your certification account, through the partner route your agent declares.
  2. It books a package that excludes that audience in audience_exclude, and waits until our tracker hears the package serve.
  3. It deletes the audience with sync_audiences and delete: true.
The check uses an excluded audience because an audience of certification ids matches nobody: a package that targets it never serves, so there would be nothing to see stop. A package that excludes it serves normally. It is also the case that matters most, because an agent that drops a deleted exclusion and keeps serving reaches the people the buyer asked to leave out. The result.
  • Passes when, once the channel’s grace has passed after the delete, our tracker hears no impression on that package. The grace is the same as for a pause: 2 hours for display; 6 hours for video, audio and CTV; 24 hours for DOOH.
  • Fails when our tracker still hears impressions on the package after the grace.
  • Is inconclusive when the package never served before the delete.
  • Also reported: whether your agent reports the package with canceled: true on get_media_buys after the delete. This is shown, but it does not pass the check on its own.
Status. The check is report-only. It never becomes a blocker, and it does not stop certification or buying. It is not yet run for any agent: it starts once Apostra holds a certification dataset on your agent’s partner route. What a pass unlocks is described in the requirement linked above. Next step. When you receive delete: true for an audience, stop every package whose targeting names it in audience_include or audience_exclude, and report each one with canceled: true on your next get_media_buys read.

Live creative changes finding

Buyers change the creatives on a live buy: they add one, take one off and swap one for another. Certification does the same on its test buy, so Apostra learns how your agent behaves before a buyer meets it. What Apostra sends. Once the pause and the resume are done and your agent reads the buy active, Apostra sends up to three changes, one at a time, each as a package update that carries only the package’s complete creative_assignments and nothing else. The buy holds the first certification creative from the start. Apostra needs a second creative of the same format that the run offered. It syncs that creative to your agent before a change names it. The order keeps at least one creative on the buy at every point. The change goes to the package that can still serve, the second of the two. Each change your agent declares is sent. An action in your available_actions (or valid_actions) is declared, in any mode. Apostra reads the declaration from your agent’s latest get_media_buys answer with the code Apostra’s own write gate reads, so the two agree. The change passes when your agent takes it. It fails when your agent declared the action as self_serve and refused it, and the result shows your agent’s own error code and message. A change your agent declared with conditions (conditional_self_serve), or that needs approval, and then refused or held, is inconclusive: a test buy may not meet the conditions. Each change your agent leaves out is sent once, to see it refused. When your agent sent an action list and the list leaves a change out, the change is not supported, and Apostra sends it once anyway. It passes when your agent refuses it with an error code, such as ACTION_NOT_ALLOWED. Apostra reads the code its storefront relays to a buyer, which can be its own code for the kind of failure (a refusal of the change as a request reads as INVALID_REQUEST), so the result may show a different code than the one your agent sent. It fails when your agent accepts it, or answers with an error that carries no code or only INTERNAL_ERROR. The buy is not left worse off if your agent accepts: both creatives fit the product, and the steps that follow continue from whatever your agent now holds. A change the agent could not answer now (SERVICE_UNAVAILABLE, RATE_LIMITED) is inconclusive. An agent that sent no action list is not probed. Apostra lets creative changes through to an agent that declares nothing, so there is no refusal to expect. Creative changes declared reports that gap. What is not judged. Apostra judges your agent’s answer to each change. It does not read the package’s assignments back, so it does not confirm that an accepted change shows on a later read: that part is not checked yet. A change your agent holds for approval, or never answers, is inconclusive. So is a run in which no second creative of the same format was offered, and a buy booked before certification paused its buys. Severity. This is a finding today. It shows as an attention finding on the Agent page when a change your agent declared was refused, or a change it left out was accepted, and it stops nothing: it never makes a use case fail and never withholds a certification pass. Whether it becomes required is a later decision, which this page will announce before it takes effect. It applies to Apostra’s own sales agents and to third-party agents on the same terms. Next step. Take the creative changes you declare, or stop declaring the ones you do not take by leaving them out of available_actions. For a change you leave out, refuse it with an error code that says it is not supported. The finding clears the next time a certification buy is judged.

Context your agent must hand back

AdCP has your agent return, unchanged, the context object a request carries, and certification checks that it does. During a certification run:
  • Every call Apostra makes to your agent carries a context: sync_accounts for the sandbox account, get_products for a brief, sync_creatives, create_media_buy, and the get_media_buys read each evaluation makes. Each response must hand that context back unchanged, including an error response and a submitted or working answer.
  • create_media_buy carries { "storefront_media_buy_id": "<Apostra's id>" }. get_media_buys must return the buy with that same context.
  • Each package of the buy carries { "storefront_package_id": "<Apostra's id>" }. The create_media_buy response and get_media_buys must return each package with its own package context, on the package that books that product. A submitted create carries no packages yet, so they are checked on get_media_buys instead.
Values are compared as JSON, so a number you return as 1.0 instead of 1 still counts as unchanged. Keep every key, including any whose value is null. The check fails if any response leaves its context out or changes it; the result names the calls and, for each, the keys Apostra sent that your agent dropped or changed and how many keys it added, for example “The seller did not echo the context unchanged on: sync_accounts (live_certification_request_key changed; 1 key added).” The result never quotes what your agent returned. The run’s session carries those calls, so you can read the request and response Apostra judged, with the values (see Read a run and the calls Apostra made to your agent). While your agent has not accepted or reported the buy, cannot be reached, or sends a response Apostra cannot read, the check waits instead of passing. A call that did not echo during a run keeps that use case failed for the run; get_media_buys is read again on each evaluation, so a fix there passes on the next one.

Assignment operations finding

AdCP 3.2 gives sync_creatives an assignment_operations list: assign, unassign and replace. A replace swaps one creative on a package for another in one atomic step, and it is how a buyer changes a creative that is already live. The older assignments list is deprecated in its favour. The CREATIVE_LIVE_UPDATE_UNSUPPORTED error that AdCP 3.3 defines tells a buyer to sync the change as a new creative and swap it in with an assignment_operations replace. What AdCP requires. The sync_creatives request schema of AdCP 3.2 defines assignment_operations, and AdCP lists no capability a seller can decline it with. Later releases add supports_assignment_merge, which covers only the separate merge operation. AdCP does not say in so many words that a sales agent must take it. So Apostra treats a refusal as a gap, reports it, and requires nothing. What Apostra sends. Once the creative id probe has put a copy of the buy’s creative on a package of the test buy, which needs no serving, Apostra syncs your agent a second copy of that creative under a plain id and sends one sync_creatives request on the same test buy. It carries no creatives and one operation: a replace on that package, swapping the first copy for the second, with validation_mode: strict. The buy keeps the creative it was booked with. When the creative id probe did not leave a copy on the package, there is nothing to replace and the replace is not sent. It sends the request only to an agent Apostra calls at an AdCP release whose sync_creatives request defines assignment_operations. An agent called at 3.1 is not sent it. The run records what your agent answered: that it took the replace, or the error code, recovery and field of its refusal. What is flagged. A refusal shows as one of two findings, so a refusal that sends the buyer to the same operation reads differently from one that does not:
  • Takes the assignment_operations replace it advises. Your agent refused the replace, and it names assignment_operations as what to send instead, in that refusal or in an answer on recorded traffic in the last 14 days. We look for the name as one whole identifier in an error’s message, suggestion or details, so not_assignment_operations does not count. A buyer that follows the advice is refused again.
  • Takes an assignment_operations replace. Your agent refused the replace and does not name it as what to send. This is reported separately and is informational in kind, because AdCP does not require it in so many words.
What is not judged. An agent called below AdCP 3.2, a run in which no second creative reached your agent, an answer your agent did not give in time, and a transient refusal (SERVICE_UNAVAILABLE, RATE_LIMITED, TIMEOUT) are not judged. A refusal with no code, or only INTERNAL_ERROR, is judged, and the result says it must carry a code a buyer can act on. Severity. These are findings today. They show as attention findings on the Agent page and stop nothing: they never make a use case fail and never withhold a certification pass. Whether either becomes required is a later decision, which this page will announce, with the date, before it takes effect. They apply to Apostra’s own sales agents and to third-party agents on the same terms. Next step. Take an assignment_operations replace on sync_creatives. If you cannot yet, do not tell a buyer to send one: change the refusal so it names only what you accept. See A refusal suggests only what your agent accepts. The findings clear the next time a certification buy is judged.

Test buy size

Each certification buy is capped at USD 10, and each run is capped at USD 50. Apostra books the test buy in your own currency, at your own prices. It certifies any pricing option in a currency Apostra can settle in: USD, EUR, GBP, JPY, CHF, CAD, AUD, NZD, CNY, HKD, SGD, SEK, NOK, DKK, PLN, KRW, INR, MXN, BRL and ZAR. A USD option needs no conversion. For any other currency, Apostra converts the USD 10 cap at its daily locked exchange rate, the same rate it uses to quote and book buys that day, and rounds it down to the currency’s smallest unit, for example EUR 9.20 at 0.92 euros to the dollar, or JPY 1,513 at 151.37 yen. The USD 50 run cap counts every buy in USD at those same rates, so a run that mixes currencies still stops at USD 50 in total. The result records the currency, the rate and the day the rate was locked on. A buy’s budget is the smallest pricing option on the product in one currency: a flat-rate option uses its fixed price, any other option uses its minimum spend per package, and an option with no minimum uses the converted cap. Apostra never rounds a minimum down. A product whose smallest option is above its converted cap is skipped as budget_exceeds_per_buy_cap. When several options qualify, Apostra uses an option that fits its cap before one that does not, then a fixed-price option before an auction, then a USD option, then the option with the smallest value in USD at the locked rates. An option with no fixed_price is an auction, and AdCP requires each package booked on it to carry a bid. Apostra bids the median of recent winning bids the option publishes (price_guidance.p50), or its floor_price when that is higher or there is no median, in the option’s own currency. The bid is the price per unit; the budget still caps what the buy spends. Apostra never books an auction option that has neither a floor nor a median, and a product whose options are all like that is skipped as seller_auction_bid_unavailable. The buy runs for three days and asks for about 15 impressions an hour, which gives roughly 1,000 impressions to compare and three daily reports. A tracked buy books two packages and both count against the per-buy and per-run caps:
  • A product that sets no minimum books two packages at half the cap each, USD 5 each and USD 10 in total, splitting the per-buy cap between them. In another currency each half is rounded down to the currency’s smallest unit.
  • A product whose smallest minimum is at most half the cap (USD 5 or less) books each package at that minimum, so the buy books twice the minimum in total.
  • A product whose smallest minimum is above half the cap (USD 5) books one package at that minimum, and its buy is not judged on the pause.
Your sandbox decides whether money moves. Apostra does not pay for test impressions.

When Apostra starts a run

Once certificationRequested is true (Allow certification buys is on), Apostra starts a run on each active Source that uses your agent, using Apostra’s own certification buyer and creatives. Each Source gets one run at a time, and every run carries the reason it started. Apostra does not start a run beside one that is still collecting evidence. A run is collecting while its three-day buy is flying, and after that until it has a passed or failed result, for at most 14 days after the buy was planned. A Source that passes is re-certified about every two weeks while it stays requested. A Source whose run failed is tried again once that run’s buy has finished flying and the run is at least 24 hours old. When your agent refuses the certification creative with an error that AdCP says needs a changed request (its recovery is correctable or terminal, for example VALIDATION_ERROR, CREATIVE_REJECTED or CREATIVE_VALUE_NOT_ALLOWED), the use case ends at once as creative_sync_failed with your agent’s reason, instead of waiting a day to send the same creative again. A creative Apostra cannot send for this product at all ends the same way, as creative_not_sent. A test buy that Apostra could not send for a reason that can clear by itself (for example your agent was rate limiting or unavailable, or did not answer) no longer holds the Source 24 hours after that failure, and the next run starts with a fresh buy. A test buy with no recorded failure still holds it. A run waiting for you to approve the certification creative also holds the Source, for the five-day wait at most; see Certification-ready sandbox. Once a run’s buy has finished flying, nothing more is tested in it. A use case that is still inconclusive then changes only if a late report settles a check that waits on reports. A check that needs the buy itself, such as the stop check on a buy that was never paused, cannot be judged any more. The run still holds the Source until 14 days after its buy was planned. Apostra then starts the next run by itself, once the Source’s latest pass, if it has one, is three days old, as evidence_expiring above describes. Apostra judges a run’s passing use cases again every hour while the run holds the Source, so after a run with a passing use case the next run starts about 17 days after the buy was planned. While you wait, the use case’s finding and the certification.no_passing_use_case finding say that the run’s buy has finished serving and give the date the next run can start. They stop saying certification is running. A new run cancels the test buys of earlier runs that it replaces: a format it buys again, and a format it no longer plans because your agent stopped offering it. Apostra sends no cancel for a test buy your agent has already reported completed or canceled, since it has already ended; the run only records it as ended. A Source whose run cannot start gets no run and no test buy; see When a run cannot start. A run that ends with nothing to judge, for example because every family was skipped, is also retried no sooner than 24 hours later, never every hour. When a Source’s runs keep ending the same way, Apostra waits longer between them. If every family that counts toward certification was skipped for the same reasons as in the run before, and your agent’s capabilities and products are unchanged, the wait doubles with each such run: 24 hours, then 48, then 96, up to seven days. It starts over at 24 hours as soon as your agent’s capabilities change, or you turn Allow certification buys off and on again. A change to your agent’s products is seen by the next run, which then starts the count over; to have Apostra look sooner, ask Apostra support to start a run. Apostra never stops trying a Source that asked for certification; it only spaces the runs out. Apostra does not yet start a run when your agent’s products or capabilities change. After a change, the next run starts when the evidence expires, or when Apostra support starts one.

When a run cannot start

Sometimes Apostra picks a Source and cannot start a run, so there is nothing to wait for. Apostra records the attempt, and every read of your agent says so instead of saying certification is running: the agent’s stuckOn step, the certification.no_passing_use_case finding and the agent’s certificationStage all name the blocked start, the Source Apostra tried through, and whose move it is. A Source’s own read and a storefront’s publish_validation readiness step say the same. The blocked start clears the moment a run begins on the Source, from any start, and also when you turn Allow certification buys off. Once an earlier run has recorded results for your agent, the agent’s reads lead with those results instead, and the finding adds that Apostra’s latest attempt to start another run could not, so no read says that no run exists while results show. When Apostra cannot start a run because the storefront had no products from your agent, it records your agent’s answer to every call that failed during that attempt: the call, the HTTP status your agent answered with (or that it did not answer at all), and the words it used, with secrets and your endpoint’s host removed and the text cut to 400 characters. Your agent’s answer is what decides whose move it is, so an attempt that was stopped by a failed call to your agent is yours to fix even when Apostra recorded no other health reading of your agent, and an attempt where every call to your agent succeeded, or none was made, is Apostra’s. The answer stays on that attempt: it stops showing the moment a run begins on the Source, and the next attempt records its own. When a step of your sandbox account stopped the start, the attempt also records what Apostra saw at that step, so the finding names the one cause behind seller_sandbox_account_failed and the change that clears it, instead of every cause the skip reason covers: Every value Apostra records from your agent’s answer is cut short and has secrets, URLs and addresses removed first. An attempt Apostra made before it kept this detail reads as the skip reason does in Skip reasons, until the next attempt. Apostra tries the Source again by itself 24 hours after the blocked attempt, and each day after until it works, so a fix at your agent is picked up within a day with nothing for you to do. There is no call that retries sooner; ask Apostra support to start a run if a day is too long.

How long it takes

Plan for about three days of serving, plus the reports that are due after the flight. Apostra re-reads the evidence every hour, so tracker hits and reports that arrive late update the result without anyone starting a new run. Apostra compares totals against the settled report about 72 hours after each period, so the final comparison can arrive after the flight ends.

Results

Each use case and product family has one latest result. Most of a run is judged within hours of the buy going live, so you see the immediate checks as they land and then wait only for reporting. The flight stays three days. The agent read shows the results of the runs that count for your agent: runs on an active Source that uses it, or on a pending Source of your own organization, which certification buys through before it goes live, taken since the Source was connected to your agent. Each run buys a rotating handful of product families, so the read shows each Source’s newest run that recorded a result, plus any pass from an earlier run that still counts toward certification. An earlier run’s other results stay on that run’s session (see Read a run), and the certification.no_passing_use_case finding says how many, for example “2 older results from earlier runs are on their sessions.” A failure an earlier run recorded comes back when a later run buys that family again. A source’s readiness counts failed and inconclusive results by the same rule. Each use case’s result on the agent read carries displayName, the use case’s name as the table in Use cases gives it, and gating. gating is true when the result decides certification and false for a use case that is shown beside the others but never counts: a pass there certifies nothing and a failure there fails nothing. Today those are the compact lifecycle, the accounts use case and the create-without-creatives probe. The Overview page marks each use case as counting toward certification or not from this field, so a Passed on a use case that does not count is never a certification pass. Each use case’s result also carries failedCheck, the name of the first check that failed, on a failed result, and skipReason, the skip reason code, on a result from a family that has no buy: one that was skipped, or one whose buy Apostra tried to send and could not before it reached your agent. The second kind keeps waiting for its buy. When Apostra did not send its creative, skipReason is creative_not_sent and the reason is the cause Apostra recorded. Otherwise its reason first says that Apostra could not send the buy and the date it tries again by itself, no sooner than a day after the failure, then what the skip reason means, what to change and the detail Apostra recorded; when the attempt recorded no skip reason, skipReason is buy_never_sent. The use case’s finding on the agent shows its newest result, so a later run that waits on your agent replaces an earlier failure as the finding, and the earlier failure stays on its run. When no family that counts toward certification has a buy in its latest result, because each was skipped or Apostra could not send its buy, the certification.no_passing_use_case finding says so. It names the skip reasons and how many families each stopped, says whose move it is, and gives the date Apostra tries again. It does not say that certification is running, because no certification buy is serving. The owner is your agent’s developer when any of those skip reasons is yours to change (see Skip reasons), Apostra when the rest are Apostra’s, and nobody when a later run covers them all. When a use case did not pass because of one of its checks, the reason names that check and what Apostra saw. A failed use case gives the first check that failed, for example “The click_fires check failed: No click reached our tracker in 1000 impressions.” An inconclusive use case gives the first check that has no result yet, for example “Waiting on the report_cadence check: The seller has not yet produced a decided reporting period.” Before any impression has reached Apostra’s tracker, the macro check waits with “No impression has reached our tracker yet, so there are no macro values to judge.” It judges macros from the first impression on.”

See each check as it lands

Each use case’s result on the agent read also carries checks: the checks that apply to it, in the order Apostra judges them, taken from its latest evaluation. Each check has: The checks that need only a few hours of serving are judged within hours: the buy goes live, your agent echoes our context, macros are replaced, clicks fire, spend stays within the budget, the pause, stop and resume work, quartiles fire, and live creative changes are taken. Only the reporting checks need the flight: whether reports arrive on time, whether they agree with our tracker, your ad server’s delivery and the settled report. expectedBy is the end of the three-day flight plus about 72 hours, the delay after which the settled report is due (see How long it takes). So six hours into a run, an agent that has passed every immediate check shows them as passed and the reporting checks as waiting with the date they are expected. The use case’s finding and the agent’s next step say it in one sentence, for example “7 of 10 checks passed on ctv_video; waiting on reporting until 9 November.” The count covers the checks that decide the use case. The sentence leaves out a use case that was skipped before a buy and one that waits for your agent to approve a certification creative. failedCheck and reason are unchanged. A skipped family never produced a buy. It is shown as inconclusive with its skip reason. It does not fail the use case, stops no running buys and cannot certify. A later run judges the family again. A family whose buy Apostra could not send yet is shown the same way until Apostra sends it.

Read a run and the calls Apostra made to your agent

Each live certification run is a session. Every use case result on the agent read (get with kind: "agent", in liveCertification.useCases) and every certification use case finding carries the sessionId of its run, when you can read it (see below). From your Agents account, read it with get and kind: "session", or list your runs with search and kind: "session". An Agents account reads only this session source, live_certification_run_v1, so search needs no filter.sessionSource. A run is not a Murph conversation: get with kind: "conversation" does not read it. The events of a run are its steps, in the order Apostra recorded them:
  • Run started. The products Apostra planned the run from. Apostra reads them from its own copy of your agent’s catalog, so this step has no call to your agent.
  • Buy sent. Apostra began sending the certification buy to your agent. It carries what Apostra sent and what your agent returned: the certification creative sync and the create, one call for each attempt. A buy sent again after a failed attempt is a new event with its own calls. When your agent answers a call asynchronously, the call still shows the request Apostra sent, such as the create_media_buy tool call, never the tasks/get or tasks/result poll, and its response is the last answer Apostra received: the task’s result once your agent finished it.
  • Buy created. Your agent accepted the buy. It carries what happens once the buy exists: later changes to it, such as the pause, and the callbacks your agent sends, such as a webhook.
  • Skipped. Apostra did not send a buy for the use case, and why. A family is skipped at most once, so this event appears once.
  • Agent answer. A request to your agent did not complete, with its status.
  • Evaluation. A result for the use case and, when it did not pass, every check that failed. When the context check failed, it carries the calls it judged: each call whose response did not hand back the context Apostra sent, or the get_media_buys read that returned the buy without its create context. An evaluation is a new event when the result changes.
An agent whose provider no administrator had claimed when the run started has no Session for it. A step that involved your agent carries its calls. Each call gives what Apostra sent, what your agent returned, the HTTP status, how long it took and the error. Apostra removes its own credentials and the values under keys that look like credentials, such as token, secret or password, and shows them as [REDACTED]. A step shows at most 50 calls. When a step had more calls than it shows, its event carries agentCallsTruncated: true. Apostra keeps the request and response bodies for 90 days; after that a call still shows its timing and status, with no body. A run is active while it is still collecting evidence, as When Apostra starts a run describes: while Apostra is sending one of its buys or waiting to, while a buy’s three-day flight is running, and after that until the buy has a passed or failed result, for at most 14 days after the buy was planned. A buy that already failed a check stays active while it flies, because Apostra keeps judging it. A run is closed once none of its buys is collecting: each was skipped, stopped or replaced, or has finished flying with a result. Use lifecycle on search to list one or the other. Beside the calls, a Session returns these facts about the run. Apostra holds them, and your agent did not receive all of them:
  • The run. Its id, your agent’s name, whether it is active or closed, and when it was created and last updated. search also gives the number of use cases that passed, failed, are inconclusive or are still pending, and how many product families they cover. A run tests a use case on one or more product families. The use case counts as failed when any of its families failed, as passed when none failed and one passed, as pending while a family has no result, and as inconclusive otherwise.
  • Run started. How many products Apostra read from its stored copy of your catalog, and the ids of the first 20.
  • Every step for a use case. The use case and its family key.
  • Buy created. The media buy id of the certification buy.
  • Skipped. Why Apostra did not send a buy, in the words of the skip reasons below, with the error code and message your agent returned when it returned one.
  • Agent answer. The status of the request that did not complete, with the error code and message.
  • Evaluation. The outcome (passed, failed, inconclusive, or recorded without a result) and every check that failed, or else the first check still waiting, with what Apostra saw. A failed context check names each key Apostra sent that your agent dropped or changed, and how many keys it added; the calls it carries hold the values. When Apostra placed the buy without a creative, the evaluation says so.
  • Stopped or replaced. That Apostra stopped the buy, or cancelled it because a newer run replaced it.
A Session does not return any other seller’s prices, buyers or settings, or anything else that was never sent to your agent. Apostra’s credentials are removed. Any member of the Agents account reads the runs of agents the account owns, whatever their role, whichever seller’s storefront the run used. A member reads exactly what is described above: the steps, the calls Apostra made to your agent, with credentials removed, and the run’s facts. No one changes or cancels a run from here. An Operator or Admin decides whether runs start, with Allow certification buys, and a Member cannot. The history starts when your organization took the agent. Moving the agent between accounts of one organization keeps it. A run that started before your organization took the agent, because Apostra staff assigned it to you or it moved from another organization, is not shown. Archiving an agent does not remove its certification sessions: you can still find and read them with search and get (kind session). The seller who connected your agent does not read these sessions.

See certification evidence on the Overview page

In an MCP host such as Claude or ChatGPT, ask to open your agent, or call open_agent_page with the agent’s agentId. The Overview page it opens carries the certification evidence, each item opening to its full detail:
  • Live certification. The latest result for each use case, named as the Use cases table names it, with its channel, when it was judged, the result of each check and, when it did not pass, the reason Apostra recorded. Apostra derives the use cases that apply from the agent’s products and capability declarations, so you do not choose them. Each one is marked as counting toward certification or not. Only the ones that count decide certification. A use case that does not count can still show Passed, and that pass certifies nothing.
  • Manual test runs. For a sales agent, the test plan (each test mode, whether it is available and why not) and the agent’s recent manual test runs, with the outcome, the failed stage, who is accountable and the next step. These are the runs someone starts on purpose. Live certification runs separately and does not create one, so an agent can have no manual test runs and still show live certification results. A manual test never certifies the agent.
  • AdCP 4.0 readiness and capability declarations. The same checks as Getting ready for AdCP 4.0 and the declarations the agent makes, with what Apostra saw and how to fix each. A declaration reads Declared or Eligible for testing: it says what the agent claims and that Apostra can test it, never that the agent passed.
  • Allow certification buys. Whether the owner has allowed certification buys. For an Operator or Admin of the Agents account, who can change the setting, this is a switch: turn it on or off and the page saves it with save_agent, then reads itself again, so what it shows is what Apostra saved. If Apostra refuses the change, the page shows Apostra’s message. Anyone else sees only whether it is on.
  • Exemption. When Apostra has exempted the agent from certification until a date, one line under the switch gives that date. Nothing is shown when there is no exemption or it has ended. See See whether Apostra exempted an agent from certification.
The page is a curated view of the same read get makes for the agent, shown to the same people who can read it. It leaves out internal fields, such as source identifiers and the links to a run’s diagnostics. Opening it changes nothing; its one write is the switch. The page is available in Agents accounts. The text Apostra writes for a result is in English whatever language the host uses. The Overview page is described in full under Open the Overview page.

Skip reasons

Skip reasons describe what stopped the buy before it was sent.

Your sandbox account

Your agent’s response to the buy

A creative sync failure is your agent’s only when the certification creative reached your agent. Apostra stops its own creative first when its creative sync does not reach the Storefront, when one of Apostra’s own checks refuses the creative, or when the AdCP library Apostra uses cannot shape it for the formats your product declares. The use case then reads creative_not_sent from that attempt on, with the cause Apostra recorded, and its finding says to wait for Apostra. Apostra staff are alerted. It does not replace an earlier pass of the use case, and a later run tests the family again. A refused buy is your agent’s only when the buy reached your agent. When the AdCP library Apostra uses refuses Apostra’s own request, or cannot shape the buy’s format choice for your agent, Apostra stops the buy before sending it. The use case then reads buy_never_sent, with the cause Apostra recorded, and its finding says to wait for Apostra. Apostra staff are alerted.

Budget

Agents that sell only through briefs

An agent that lists no wholesale products and declares the brief buying mode in media_buy.buying_modes is certified through a brief. Declaring media_buy.supports_proposals is not required: an agent that answers a brief with products and no proposal is certified the same way, and an agent that declares supports_proposals: true is certified through briefs too. Apostra’s standard briefs cover display with a tag, online video with a VAST tag, and linear TV with a format your agent supports. Each has a USD 10 budget (or its equivalent in your currency) and a three-day flight. Apostra buys a matching offer within the cap. Linear TV is judged on seller-reported evidence because the Apostra tracker cannot observe it. Each brief goes to your agent under the sandbox account your agent set up for the run, the same account Apostra’s test buy and creative use. Apostra sends a brief only for a channel your Storefront serves. A Storefront that serves display only is sent the display brief and no video or linear-TV brief. If your Storefront serves none of the covered channels, the run does not start and says so. If Apostra cannot start a run at all because buyers cannot see the Source, it says why: for example that the Storefront is not set up and what is missing, or that the Source is not mapped to an agent. Fix that first; certification reads the same view of the Source that buyers get.

Apostra’s side

The compact lifecycle

A product with no declared AdCP channels is skipped before any family is built (missing_channels). That is a finding with the products named, not a failure. Declare channels on the product, or applies_to_channels on each format option.

How certification is kept

Certification is continuous, and evidence has to stay fresh.
  • Fresh evidence. Most checks need a pass from the past seven days. The report check follows your reporting cadence instead, so a product that reports monthly stays certified between reports.
  • Real buys count. A real buy for a use case produces the same observations as a certification buy. While one is live, its traffic is the evidence. When none is, Apostra runs a synthetic buy on your sandbox account.
  • A failure blocks new buys only. When a check fails on live traffic, the use case cannot take new buys, and Apostra tells you what it saw and how to fix it. A failure never pauses a buy that is already running; the only pause Apostra sends is the one in its own certification buy. The use case recovers as soon as live traffic passes again, with no re-registration step. A use case with no family passing takes no new buys once your agent’s exemption ends; see The certification deadline.
Certification buys and creatives skip the Storefront’s own approval queues. Your agent and ad server still review them, and Apostra’s platform safety review still holds a buy it flags.

Capabilities required from 1 February 2027

Canonical creative formats, Reliable Reporting, the compact lifecycle and webhook signing become required on 1 February 2027. Canonical creative formats were due on 1 December 2026 until 15 November 2026, when Apostra moved them to the same day as the others. An agent that lacks one of them on its date loses its certification that day, including an agent certified earlier. An agent that speaks only AdCP 3.0 cannot offer canonical creative formats and must move to AdCP 3.1 or later. Nothing is enforced before a capability’s date. From its date, a pass does not count for an agent that lacks it, so the agent no longer completes the Successful transaction go-live step or lets another organization’s publisher connect, its Sources stop taking new buys, and it counts again once the agent has the capability. See Getting ready for AdCP 4.0 for the schedule and what to change.