> ## Documentation Index > Fetch the complete documentation index at: https://docs.apostra.com/llms.txt > Use this file to discover all available pages before exploring further. # Get products across storefronts > Use one canonical AdCP-shaped request to retrieve, evaluate, refine, and buy proposals across your connected storefronts `get_products` is the canonical product door for buyer agents. Send one AdCP-shaped request to selected storefronts—or all storefronts connected to an advertiser—and receive one bounded canonical page as `products[]` plus optional sibling `proposals[]`. Every identity remains qualified by its originating storefront. Apostra can optionally evaluate valid seller proposals before returning them. You supply plain-language operating instructions. A managed model chooses `accept`, `reject`, or `refine`, can attach buyer-owned enrichment, and can rank the accepted cohort across sellers. AdCP remains a one-buyer-to-one-seller protocol. Apostra is the many-to-one buyer layer: it fans out the canonical request, validates each seller independently, qualifies identities, and returns a bounded aggregate page. Sellers do not see one another or receive an Apostra-specific request. ## One protocol request, marketplace scale ```mermaid theme={null} flowchart LR B[Buyer agent] -->|canonical request + ext.interchange| I[Apostra get_products] I -->|1:1 AdCP get_products| S1[Storefront A] I -->|1:1 AdCP get_products| S2[Storefront B] I -->|1:1 AdCP get_products| SN[Storefront N] S1 --> I S2 --> I SN --> I I -->|bounded canonical page + qualified IDs| B ``` The scaling layer does four things without forking the protocol: * **Bilateral calls stay canonical.** Buying mode, brief, filters, proposals, and refine are ordinary AdCP semantics at each seller edge. * **Identity survives aggregation.** `sf1:` product IDs and `sfp1:` proposal IDs prevent cross-storefront collisions and round-trip into purchase. * **Slow sellers do not erase fast sellers.** Revisioned replacement snapshots expose seller progress while other agents remain pending. Candidates become selectable only after their complete seller/source identity is persisted. * **Responses stay agent-sized.** `pagination.max_results` bounds candidates; proposal graphs remain whole, and the cursor is frozen only when the execution is complete. ## Which sellers are searched for an advertiser Discovery for an advertiser searches only the media partners turned on for that advertiser: partners selected automatically for its markets and channels, partners included for the whole account, and partners turned **On** for that advertiser. Media partners that are off for the advertiser are not contacted and do not appear in `storefront_results[]`. They are outside the search, not failed sellers. When the request names no storefronts, `ext.interchange.storefronts_not_searched` lists the media partners left out for the advertiser: each one's `storefront_id`, `storefront_name` and `reason`. `NO_AUTOMATIC_MATCH` means it does not match the advertiser's markets and channels, `ADVERTISER_DISABLED` means it is turned off for the advertiser, `ACCOUNT_ALWAYS_EXCLUDE` means it is excluded for the whole account, and `HARD_INELIGIBILITY` means it cannot transact with you at all. At most 10 are listed; `total` counts every one. When you can turn any of them on, `suggestion` says how, once. The `get_products` text summary names them the same way. Each media partner appears once, even when it is reached through more than one sales agent. The list is reported only for an advertiser your account itself owns; for any other advertiser, such as one another organization delegated to yours, it is omitted rather than guessed. A media partner that is off also changes the rest of the response in these cases: * **You named it.** When the request names specific storefronts, or refines a product or proposal from one, a named media partner that is off returns a `storefront_results[]` row with status `completed`, no products, and an `outcome` of kind `action_required` with `reason_code` `advertiser_activation`. Its `message` says why it was not searched and its `suggestion` says where to turn it on. A partner turned off for the advertiser is turned on in the advertiser's view of Media Partners (open the media partner, choose **Controls**, set **Availability** to **On**). A partner excluded for the whole account is changed in the account view, from **Always exclude** to **Default** or **Always include**. * **No media partner is on.** When none is turned on for the advertiser, discovery searches nothing and `guidance` says so, with the same path to turn one on. * **Otherwise**, an empty result's `guidance` also reports how many media partners were left out because they are not turned on for the advertiser. A media partner that cannot transact at all is reported as `skipped` when you named it; turning it on does not change that. A media partner that is still finishing go-live is different. Some managed ad servers need the media partner to pass a no-spend certification, such as a test write to the ad server, before it takes its first booking. Until it passes, the media partner is searched like any other: its products appear in results, and you can add them to a buy you are planning. Creating or executing a media buy with it is refused with `ACCESS_DENIED` and a message saying the media partner has not finished going live. There is nothing to turn on; the media partner finishes the certification on its side, and booking opens once it passes. Saved results follow the same rule. When a media partner is turned off for the advertiser after a search, reading that saved discovery no longer returns its products; an empty page's `guidance` says how many saved products were removed for that reason. Adding one of its products to a buy, directly or by applying a proposal, is refused with `ACCESS_DENIED` and a message saying where to turn the media partner back on. ## Read each seller result honestly `ext.interchange.storefront_results[].outcome` is present when a seller was queried through its own agent, or when a seller you named was not searched. It is one buyer-safe terminal outcome: `returned_products` (with its count), `no_matching_products`, `no_inventory_for_request`, `action_required` (a named seller that is off for the advertiser, with its fix), `skipped` (with a safe reason), or `failed` (with a safe reason). A seller with no active wholesale inventory is `no_inventory_for_request`, not a failed seller and not a retry instruction. An empty seller result becomes `no_matching_products` only after that seller has completed without an error. A seller response with an error and no products is `failed` when no cached catalog can serve the request, even when the seller marks its task completed. An incomplete result remains `incomplete` and can carry a safe seller clarification instead. When one agent covers several sellers, `outcome` and `filter_result` are omitted because that agent-level result cannot be attributed honestly to an individual seller. Those rows carry product-derived counts and `property_coverage[]` only. Agent-level outcomes for that case are tracked in AI-10300. Returned products also produce `property_coverage[]`: each property domain and the number of returned products carrying it. Treat this as coverage of the seller response, not an exhaustive marketplace claim. Each seller row includes at most 12 property domains, with `property_coverage_total` and `property_coverage_truncated` when the buyer needs to recognise an abbreviated list. When a seller reports filter accounting or unfinished work, `filter_result` says whether no products matched, filters excluded every candidate, or the result is partial. Seller diagnostics and operational details are not exposed here. The advertiser's property exclusion list is applied after seller products are returned. `ext.interchange.excluded_product_count` is the total number removed, and each storefront row exposes its own `storefront_results[].excluded_product_count`. Each seller row also has a bounded `excluded_products[]` list with the removed product ID, name, and the matching domain or domains. These domains are the intersection with the buyer's own exclusion list, never the rest of that list. If `excluded_products_truncated` is true, use the count as the complete total and describe the listed products as a bounded sample. If exclusion removes every completed candidate, read the response guidance and report that **N products from M sellers were removed because they included properties on the advertiser's exclusion list**. If a legacy candidate has no seller identity, the total still includes it but no seller row does; report the removed-product count without inventing a seller count. Do not call either outcome "no matching inventory". `get_products` does not stop after the first 10 storefronts or choose a final cohort by response speed. When no storefront selector is supplied (or `storefronts: "all_connected"` is explicit), Apostra resolves every storefront in the buyer's eligible discovery scope and queries each backing sales agent once. An ordinary provisional revision contains only sellers that have answered so far and returns `pagination.has_more: false` with no `pagination.cursor`; poll the same `execution_id` with `since_revision` until `results_complete: true`. Only then follow `pagination.cursor` until `has_more: false` to consume the complete candidate set. This enforces the stable-cursor contract: incomplete ordinary results are replacement snapshots, not pages to append. `pagination.max_results` limits each completed response page, not the seller fan-out. The legacy Discover Products `groupLimit` and `groupOffset` fields do not apply to this surface. If the buyer's absolute wait budget ends while Apostra is still assembling the canonical page (for example, managed evaluation is still running), Apostra returns the latest persisted replacement snapshot when one exists. That page keeps the products and storefront results from sellers that already answered, sets `results_complete: false`, and names the remaining sellers as pending. The requested processing mode is `degraded`: evaluation requests mark `ext.interchange.evaluation.evaluated: false`, while screening requests mark `ext.interchange.screening.evaluated: false`; those candidates were passed through without the requested instructions. The usual `max_results` and its server-marked deadline-snapshot pagination cursor still apply to that replacement page. If no seller snapshot has landed yet, the same incomplete response is an empty page and says so without claiming that a seller failed to answer. For a pure pass-through storefront, a country filter also narrows the storefront's internal source fan-out after the staged routing flag is enabled. Apostra sends the request only to an active source whose own standard `primary_channels` and `primary_countries` match the requested coverage. A source with missing or malformed primary coverage is unknown and excluded from filtered fan-out after activation; before activation, legacy fan-out is unchanged. Declarations from two different sources are never combined to manufacture a match. An unfiltered request remains storefront-wide. Apostra rejects ambiguous seller-specific controls such as conditional catalog versions on the aggregate surface. Query that seller's AdCP endpoint directly when you need a seller-specific conditional request. ### Use canonical filter names Apostra rejects known cross-operation fields inside `filters` before contacting any seller. Use `filters.countries` to narrow products by country coverage. Put delivery targeting such as `geo_countries` under `targeting_overlay`, not `filters`. `excluded_content_categories` is not a supported `get_products` filter and must be omitted. Canonical filter fields from the AdCP versions supported by Apostra remain available. Put seller-specific filter criteria under a vendor namespace in `filters.ext`, for example `filters.ext.example_seller`. These unsupported fields return a validation error that names the field and how to correct it. ## Canonical options from older seller formats ### Buyers When a seller still uses named legacy formats, `get_products` can return the seller's explicit mapping as a URL-free canonical `format_options[]` entry. Buyer integrations do not need to change: existing canonical options and legacy `format_ids[]` keep their current shapes. When a buyer selects one of these canonical options, Apostra retains the seller's exact publisher-scoped route for later `create_media_buy`, `update_media_buy`, and assigned `sync_creatives` delivery. The route is never reconstructed from a similar format kind or ID. If the exact seller mapping is missing, malformed, or inconsistent with the returned option, delivery fails closed and the buyer should refresh products before retrying. ### Sellers Sellers opt in one format at a time through the catalog returned by `list_creative_formats`. Add a `canonical` declaration to the legacy row. If the option needs canonical parameters, include `canonical_parameters` on the same row and use the same kind: ```json theme={null} { "format_id": { "agent_url": "https://seller.example", "id": "seller_banner", "width": 300, "height": 250 }, "name": "Seller banner", "canonical": { "kind": "image" }, "canonical_parameters": { "format_kind": "image", "params": { "width": 300, "height": 250 } } } ``` Dimensions and duration in the canonical declaration must agree with the legacy format reference. Apostra ignores missing, invalid, conflicting, ambiguous, or unsupported mappings. If a returned product's older format remains unresolved, the seller result includes details in `projection.diagnostics`. Apostra never guesses a canonical kind from a legacy ID or asset list, and one seller's mapping cannot apply to another seller. Products may publish canonical `format_options[]` directly or retain exact legacy references from a shared format catalog. Use the catalog mapping to replace seller-specific formats that Apostra cannot interpret. ### Product-format resolvability When canonical-format compliance is enabled for a storefront, Apostra surfaces a seller-visible blocker only when a format cannot be mapped to an interpretable declaration. Exact legacy `{agent_url, id}` references remain supported without a sunset, for both existing and newly connected sources. Direct URL-free `format_options[]` are accepted but are not required yet. A source with no persisted product-format evidence receives an evidence-pending blocker until its initial catalog sync records declarations. Catalog age is a separate freshness concern: an interpretable declaration does not become noncompliant when its cache entry ages. An unresolved declaration remains withheld until corrected. For a sales agent on the Agent-supplied path, each live `get_products` response is a brief-specific observation, not a complete catalog snapshot. The newest response for each public or buyer-account context replaces that context's previous compliance observation, so products omitted from a later brief do not keep the source withheld. A buyer that tries to purchase a previously returned product still has that exact product's declaration checked before the buy is forwarded. This policy audits active external third-party sales-agent sources; managed and modular inventory sources use their own readiness contracts. A canonical-format blocker withholds the affected source from discovery and media-buy admission until the seller replaces the unresolved custom reference with an exact shared-catalog reference or an interpretable custom canonical declaration. Other healthy sources in the storefront continue selling. If no other source can serve, the storefront remains transaction-ineligible. ## Choose `get_products` or Discover Products Both surfaces share the same seller fan-out implementation. They are different buyer projections, not different integrations. | Use | `get_products` | Discover Products | | - | - | - | | Primary contract | Canonical AdCP products and proposals, plus `ext.interchange` orchestration | Apostra discovery-session view | | Best for | Buyer agents, protocol-compatible workflows, named/all storefront queries, seller proposals, screening/refine, direct purchase continuation | Existing discovery-session integrations that need grouped pagination, summary/budget context, browse, or SSE | | Intelligence | Optional absolute proposal screen: accept, reject, or seller refine | The same optional screen; no separate ranking brain | | Result shape | Flat, bounded `products[]` and optional `proposals[]`, with per-storefront progress metadata | `productGroups[]`, discovery summaries, budget context, and plans | | Durable continuation | `execution_id` → `create_media_buys` → DRAFT campaign cart | `discoveryId` → legacy selection endpoints or the same campaign reconciliation path | Start with `get_products` for new agent integrations. Use [Discover Products](/v2/buyer/discovery/tasks/discover-products) when you need its legacy grouped/session presentation or need to continue an existing `discoveryId` workflow. Calling both for the same request is normally unnecessary; Discover Products projects the same raw seller snapshot rather than running another ranking implementation. ## Surfaces * REST: `POST /api/v2/buyer/products/query` * Buyer MCP: `get_products` * Buyer `api_call`: operation `get_products` The purchasable continuation is: * REST: `POST /api/v2/buyer/media-buys/batch` * Buyer MCP: `create_media_buys` * Buyer `api_call`: operation `create_media_buys` ## Continue the selected product query When the buyer selects returned proposals or products, reuse the exact `ext.interchange.execution_id` and storefront-qualified IDs from that `get_products` result. Do not call `get_products` or Discover Products again merely to attach, re-fetch, or reconstruct the selection. A new discovery is a different snapshot: its candidates may differ or be empty, and it does not identify the query the buyer selected from. The documented legacy recovery for an execution that predates its durable accepted-proposal set still requires a fresh `get_products` call; treat that response as a new selection and ask the buyer to select from it again. If an agent host offers the typed `create_media_buys` tool, call it directly. On a reduced or shared tool catalog where that typed tool is absent but `buyer_api_call` is available, use `buyer_api_call` with `operation: "create_media_buys"`. Both tool routes invoke the same continuation contract. REST clients call the batch endpoint above. If the conversation no longer contains the execution ID or a qualified selected ID, report the missing handle and offer a new discovery as recovery; do not describe that new query as the original attachment. ## Retrieval without evaluation Omit `ext.interchange.evaluation` to receive valid seller responses without a managed evaluation: ```json theme={null} { "buying_mode": "wholesale", "filters": {}, "pagination": { "max_results": 50 }, "ext": { "interchange": { "advertiser_id": 123, "storefronts": "all_connected", "wait_mode": "quick" } } } ``` Protocol validation always runs. Malformed seller responses, broken proposal allocations, and invalid identities are rejected before evaluation and do not consume evaluation work. ### Include products from verified AdCP 2.x sellers Some sellers still run a verified AdCP 2.x response contract, which predates the current required `reporting_capabilities` block. Their otherwise valid products are excluded from the default response because Apostra does not guess reporting cadence, delay, metrics, or webhook support. REST and MCP callers can request the separate compatibility response by adding the following field to `get_products`: ```json theme={null} { "buying_mode": "wholesale", "filters": {}, "ext": { "interchange": { "advertiser_id": 123, "response_contract": "interchange_product_compatibility_v2" } } } ``` The response repeats that contract at `ext.interchange.response_contract`. A projected 2.x product omits `reporting_capabilities` and explains the omission explicitly: ```json theme={null} { "ext": { "interchange": { "compatibility": { "status": "projected", "source_contract": "adcp_2_x", "source_adcp_version": "2.5", "unavailable_fields": ["reporting_capabilities"] } } } } ``` Apostra accepts this projection only when protocol negotiation provides a non-synthetic, seller-declared 2.x version. Unknown, conflicting, synthetic, or current-version responses still fail closed. The exact seller product payload—not the buyer-facing projection—is retained for any later purchase. Omit `response_contract` to keep the original response schema, where every product requires `reporting_capabilities`. Existing REST and MCP integrations therefore need no changes. Sellers can remove the compatibility marker from future discovery by upgrading their declared contract and returning the current reporting block; no buyer-side account or purchase setting changes. ### Request a smaller product projection Use the canonical `fields` array when the next step needs only part of each product. Apostra validates the complete seller response before applying the projection, then returns only the requested product fields. The qualified `product_id` and `ext.interchange` provenance remain present so the result can still be refined or passed to `create_media_buys`. ```json theme={null} { "buying_mode": "wholesale", "filters": {}, "fields": ["product_id", "name", "pricing_options", "delivery_type"], "pagination": { "max_results": 10 }, "ext": { "interchange": { "advertiser_id": 123, "storefronts": "all_connected" } } } ``` For this request, unrequested product blocks such as `format_options` and `reporting_capabilities` are omitted. Omit `fields` when you need the complete canonical product objects. ## Evaluate and negotiate proposals Add evaluation instructions when you want Apostra to remove unsuitable proposals, enrich accepted results, negotiate promising proposals, and optionally rank the accepted cohort before it reaches your agent: ```json theme={null} { "buying_mode": "brief", "brief": "Premium UK sports video with measurable reach", "filters": { "countries": ["GB"] }, "pagination": { "max_results": 50 }, "ext": { "interchange": { "advertiser_id": 123, "storefronts": "all_connected", "evaluation": { "instructions": "Accept proposals that fit the brief and include measurable reach. If CPM is above the stated budget, ask for a lower-cost option. Reject proposals without usable measurement.", "max_refinement_rounds": 1, "max_evaluation_passes": 20, "enrichment": true, "ranking": { "instructions": "Prefer stronger measurement and efficient expected reach." } } } } } ``` The first evaluation pass is absolute: * **Accept** returns the seller's unchanged proposal and referenced products, with optional buyer-owned enrichment. * **Reject** suppresses the proposal from the normal result. * **Refine** sends a canonical AdCP refine instruction back to that proposal's seller, then evaluates the seller's revised response again. Apostra never rewrites seller pricing, allocations, availability, or terms. Any changed proposal must come back from the seller. Autonomous refinement is bounded by `max_refinement_rounds`, repeated-request detection, the request deadline, and `max_evaluation_passes`. A proposal still requesting refinement at the limit is rejected rather than silently accepted. When `ranking` is present, a separate comparative pass scores accepted candidates under its optional instructions. The response is ordered by score and includes a rank and rationale. Provisional rankings replace earlier snapshots as sellers settle; the final completed revision is frozen. The platform selects the managed models. This version does not expose provider model names, customer executable code, or a model picker. `ext.interchange.screening` remains a deprecated compatibility alias for accept, reject, and refine. It does not enable enrichment or comparative ranking. Do not send `screening` and `evaluation` together. ## Discover Products compatibility `discover_products` is a convenience view over the same seller fan-out and evaluation engine. Passing its optional `evaluation` field enables the same disposition, enrichment, refinement, and ranking behavior, projected into the Discover response shape. Without `evaluation`, Discover preserves its existing enrichment, relevance, ordering, and refinement behavior. Existing Discover clients do not need to migrate to keep their current behavior. ## Proposals and products Canonical AdCP proposals reference sibling products through allocation `product_id` values; they do not duplicate complete product objects. Evaluation resolves that graph automatically. A top-level product that is not allocated by any proposal remains a proposal-less candidate, principally for wholesale catalogs. `pagination.max_results` bounds the combined number of proposals and proposal-less products. Products referenced by a paged proposal accompany that proposal and do not consume another candidate slot, so the proposal graph is never split across pages. Apostra accepts at most 50 allocated products in one proposal. A response contains at most 100 total proposal and product objects; when complete proposal graphs reach that limit, the page returns fewer proposal candidates and the cursor continues from the next whole proposal. Both IDs are opaque and storefront-qualified: ```json theme={null} { "products": [ { "product_id": "sf1:42:...", "ext": { "interchange": { "storefront_id": "42", "source_product_id": "seller-product-7" } } } ], "proposals": [ { "proposal_id": "sfp1:42:...", "allocations": [{ "product_id": "sf1:42:..." }], "ext": { "interchange": { "storefront_id": "42", "source_proposal_id": "seller-proposal-9", "evaluation": { "disposition": "accept", "evaluated": true, "confidence": 0.86, "reasons": ["Fits the brief and measurement requirements"], "enrichment": { "summary": "Premium UK sports video with measurable reach.", "strengths": ["Strong measurement"], "concerns": [], "labels": ["video", "sports"] }, "ranking": { "rank": 1, "score": 0.91, "rationale": "Best combination of measurement and reach.", "provisional": false } } } } } ] } ``` Pass qualified IDs back unchanged in refine and media-buy creation calls. When more than one seller or storefront is in scope, Apostra rejects an item-scoped refine instruction that uses an unqualified product or proposal ID. Request-scoped refine instructions still apply across the request. Two storefronts can expose the same upstream product and therefore carry the same `ext.interchange.source_product_id`, including the same `wh:` value. Treat the top-level qualified `product_id` as the identity, and use the `storefront_id` and `storefront_name` from that same product when displaying provenance. Never merge or relabel results by `source_product_id` alone. ## Stage or buy the result The `execution_id` in the response is also the durable search/refinement record used by `create_media_buys`. Choose whole proposals and/or proposal-less products from that execution. The operation stages them on a DRAFT campaign—the campaign is the shopping cart and the parent of every media buy. Only proposals in the execution's durable accepted set can be selected. Run `get_products` again before purchasing from an older execution that predates that set. Use an existing campaign cart: ```json theme={null} { "productQueryId": "query-id", "campaign": { "campaignId": "campaign_123" }, "mode": "stage", "selections": [ { "kind": "proposal", "proposalId": "sfp1:42:...", "totalBudget": 40000 }, { "kind": "product", "productId": "sf1:57:...", "pricingOptionId": "cpm-standard", "budget": 10000, "bidPrice": 8.5 } ] } ``` Or provide `campaign.create` with the same required fields as a discovery-mode campaign (`advertiserId`, `name`, `flightDates`, and `budget`). Apostra creates the DRAFT campaign and returns its ID. Media buys are never standalone, even when the caller did not create a campaign first. * `mode: "stage"` updates the cart and prepares DRAFT media buys without contacting sellers. * `mode: "execute"` performs the same preparation, then sends each resulting buy to its originating storefront through ordinary bilateral AdCP `create_media_buy` calls. * `replace: true` replaces the cart selection before applying this batch; otherwise selections merge idempotently. Use `mode: "execute"` when the selected result can be submitted immediately. When you need to inspect or customize per-buy creatives, dates, pacing, or optimization goals, use `mode: "stage"`, make those changes on the returned DRAFT media buys, then call `execute_campaign` or repeat `create_media_buys` with the same campaign and selections using `mode: "execute"`. Execution submits the prepared DRAFT directly; it does not rebuild it from the product query. One marketplace call may reach several sellers, but the seller calls are not one atomic transaction. The response reports `mediaBuysExecuted`, `success`, and per-buy `errors`. Retry the same request after a partial failure; completed buys are not dispatched again. A completed product query accepts only an unchanged retry of its recorded batch. To change the selection after execution, start a new `get_products` query. See [Create media buys from a product query](/v2/buyer/discovery/tasks/create-media-buys) for the complete request and response contract. ## Progressive responses Inside Murph, the runtime retains the original execution and request arguments, reads cumulative results with `since_revision: 0`, and follows producer-issued frozen cursors before attaching a selection. Its bounded wait can end with an incomplete result; that does not mean the query found no products. See [Murph continuation and attachment](/v2/buyer/discovery/tasks/browse-products#response) for the wait limit, confirmed retries, and recovery behavior. Seller responses arrive independently. Each response is a replacement snapshot: * `revision` increases as sellers settle or a refinement produces a new version; * `provisional: true` means replace your previous displayed result; * `pending_agents` identifies sellers still working; * fresh provisional revisions may contain seller progress but no candidates while typed product identity is still being persisted; * proposals being refined are withheld; * ordinary incomplete responses set `pagination.has_more: false` and issue no cursor; poll the `execution_id` with `since_revision` instead. No stable cursor is issued until `results_complete: true`; * `storefront_results[].message` preserves a seller's explanation for a successful empty response, including when readiness prevents that storefront from transacting; do not interpret that case as “no matching inventory”; * `storefront_results[].error: "product_persistence_incomplete"` means the storefront answered, but one or more candidates could not be persisted with a complete seller/source identity. Those candidates and any proposals that allocate to them are withheld. Safely persisted candidates from the same execution are still returned; start a new query to retry the withheld inventory. If no safe candidate remains, the request returns a retryable service-unavailable error instead of an authoritative empty catalog; * the final cursor is scoped to the execution, revision, evaluation instructions, and offsets. Response speed affects which sellers are visible in an early provisional revision, not which sellers are retained in the completed execution. Without comparative ranking, completed products use deterministic storefront-qualified identity ordering; when `evaluation.ranking` is present, the accepted cohort is ordered by that ranking instead. Poll with the returned execution identity: ```json theme={null} { "buying_mode": "brief", "ext": { "interchange": { "execution_id": "query-id", "since_revision": 3 } } } ``` Do not append provisional pages. Replace the prior page with the new revision. Evaluation is idempotent by candidate version and instruction fingerprint. Polling an unchanged revision does not rerun or remeter it. A seller's changed proposal is evaluated as a new candidate version. ## Evaluation metadata and cost The response reports evaluation separately from seller progress: ```json theme={null} { "ext": { "interchange": { "evaluation": { "enabled": true, "status": "complete", "model_class": "fast", "evaluation_passes": 2, "ranking_passes": 1, "candidates_evaluated": 17, "candidates_ranked": 8, "accepted": 8, "rejected": 7, "refining": 0, "refinement_limit_reached": 2, "estimated_ius": 3, "charged_ius": 0, "billing_status": "calibrating" } } } } ``` One **Proposal Evaluation Pass** evaluates a bounded batch of at most ten candidates. One **Comparative Proposal Ranking Pass** ranks at most 100 accepted candidates. We measure these value units rather than provider tokens. Both are currently calibrating, so `charged_ius` remains `0`. Retrieval without evaluation reports `billing_status: "included"`. If managed evaluation is unavailable, valid candidates pass through and the evaluation status becomes `degraded`; an infrastructure failure does not silently reject seller supply. Every pass-through candidate is explicitly marked `ext.interchange.evaluation.evaluated: false` (with `confidence: 0`), and the response `guidance` says so. Check `evaluated` before treating an accept as a judgment: `evaluated: false` means your evaluation instructions were **not** applied to that candidate — re-run the request or apply your own screening before buying on it. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.