--- published: true layout: post title: 'Advertising: A Bundle And The Field That Does Not Exist' image: https://kinlane-images.s3.amazonaws.com/apievangelist/api-evangelist-images/advertising-a-bundle-and-the-field-that-does-not-exist.png date: 2026-11-18 author: Kin Lane tags: - Advertising - OpenRTB - AdCOM - Structured Outputs - JSON Schema - OpenLint - Inference --- I've been building bundles that pair a JSON Schema with OpenLint-format rules, after finding out that [the decoder only pours part of the bedrock](https://apievangelist.com/2026/11/02/the-decoder-only-pours-part-of-the-bedrock/). When I [looked at advertising](https://apievangelist.com/2026/10/28/advertising-the-shape-of-the-answer/) I used an example brief: "CTV, US only, 30 second video, no gambling adjacency, private deal with this publisher." I said constrain the decoder to OpenRTB and that brief can only become a bid request an exchange will parse. Building it, I found that one of those five phrases doesn't have an obvious place to go. ## The answer I picked Turn that brief into an OpenRTB 2.6 video deal request. The [specification](https://github.com/InteractiveAdvertisingBureau/openrtb2.x/blob/main/2.6.md) lives as Markdown tables, and nobody publishes a JSON Schema of it that I could find, so I wrote one from the tables for only the fields this brief needs. This is my subset, and no exchange has agreed to it. ```yaml bundle: id: advertising-video-deal-request version: 0.1.0 status: illustrative # an OpenRTB 2.6 subset of ours, not an exchange's supported subset answers: Turn a campaign brief into an OpenRTB 2.6 video deal request derived_from: - https://github.com/InteractiveAdvertisingBureau/openrtb2.x/blob/main/2.6.md - https://github.com/InteractiveAdvertisingBureau/AdCOM engine: runtime: deepinfer tested_against: {xgrammar: 0.2.2, strict_mode: true} so_args: {schema_type: json} sampling: {temperature: 0, max_tokens: 1024, seq_seed: 20261118} env: {DEEPINFER_BATCH_INVARIANT: 1, DEEPINFER_ENABLE_MTP_VARLEN: 0} input: context: [brief, deals, mimes] rules: deal-on-file: description: A private deal id has to exist before a request can name it given: $.context.deals severity: error then: {function: length, functionOptions: {min: 1}} brief-present: description: There is nothing to translate without a brief given: $.context severity: error then: {field: brief, function: truthy} shape: type: object additionalProperties: false required: [id, imp, tmax, device] properties: id: {type: string, minLength: 1, maxLength: 64} tmax: {type: integer, minimum: 50, maximum: 1000} site: type: object additionalProperties: false required: [domain] properties: {domain: {type: string}} app: type: object additionalProperties: false required: [bundle] properties: {bundle: {type: string}} device: type: object additionalProperties: false required: [devicetype, geo] properties: devicetype: {enum: [3, 7]} # AdCOM: Connected TV, Set Top Box geo: type: object additionalProperties: false required: [country] properties: country: {type: string, pattern: "^[A-Z]{3}$"} imp: type: array minItems: 1 maxItems: 1 items: type: object additionalProperties: false required: [id, video, pmp] properties: id: {const: "1"} video: type: object additionalProperties: false required: [mimes, minduration, maxduration] properties: mimes: type: array minItems: 1 items: {enum: "${context.mimes}"} minduration: {type: integer, minimum: 0, maximum: 300} maxduration: {type: integer, minimum: 1, maximum: 300} pmp: type: object additionalProperties: false required: [private_auction, deals] properties: private_auction: {enum: [0, 1]} deals: type: array minItems: 1 items: type: object additionalProperties: false required: [id, bidfloor, bidfloorcur] properties: id: {enum: "${context.deals}"} bidfloor: {type: number, minimum: 0} bidfloorcur: {enum: [USD]} output: rules: site-or-app: description: House rule. OpenRTB says site and app each apply only to their own inventory given: $.answer severity: error then: function: xor functionOptions: {properties: [site, app]} durations-in-order: description: minduration cannot exceed maxduration given: $ severity: error then: function: notAfter functionOptions: {earlier: answer.imp.0.video.minduration, later: answer.imp.0.video.maxduration} us-only-means-usa: description: House check for this brief, geo must be the ISO alpha-3 code given: $.answer.device.geo.country severity: warn then: {function: pattern, functionOptions: {match: "^USA$"}} on_failure: error: {retry: 2, change: [seq_seed, append_rule_messages_to_prompt], then: send_to_trader} warn: {action: flag_for_review} ``` ## The phrase with nowhere to go Here's how the brief maps. "CTV" becomes `device.devicetype` 3, from [AdCOM's](https://github.com/InteractiveAdvertisingBureau/AdCOM) device type list. "US only" becomes `device.geo.country`, and OpenRTB wants ISO alpha-3, so it's `USA`, not `US`. A "30 second video" becomes `video.maxduration`. The private deal becomes `pmp.deals`, with `private_auction` at 1. Then there's "no gambling adjacency." My first instinct was the `bcat` field, the block list, along with `cattax`. But the spec defines `bcat` as "blocked advertiser categories." That means it keeps gambling *advertisers* out. My brief means something else: *my* ad shouldn't run next to gambling *content*. Those are different things pointing in opposite directions, and the two fields sound close enough that a model, or a person, could easily confuse them. So I left it out, because I couldn't find a field in the parts of the spec I read that says what the brief means in a bid request. In practice that kind of adjacency is often handled by a deal's terms or by verification vendors, and I'd want someone who trades this to correct me. But think about what a decoder would've done. If I'd given it `bcat`, it would have filled it in with confident valid-looking codes, the exchange would have parsed it, and the constraint the buyer asked for wouldn't have been applied. **A valid answer to a question the shape can't express** is the failure that worries me most in this whole series, and no rule can see it, because the document is correct. The runner needs a way to say "this brief asked for something I can't represent." I don't have that in the bundle yet. A model constrained to a shape has no way to tell you that the shape is the wrong one, unless the shape gives it a place to say so. ## Where the other design choices are **Only real deals.** `deals[].id` binds from `context.deals`. A model can't invent a deal ID, which would otherwise parse and then simply never match. **`site` or `app`, as a rule.** OpenRTB says each applies only to its own kind of inventory. The natural JSON Schema is `oneOf`, and the decoder doesn't enforce that as exclusive. So it's the linter's core `xor` function, and no custom code is needed. I've labeled it a house rule, since the text stops short of a formal ban on sending both. **`minduration` under `maxduration`.** Two fields that have to agree with each other, so that belongs in a rule. **The floor is a guess, and that's a problem.** Nothing in the brief gives a `bidfloor`, yet the shape requires it, so a model will produce one. In a real deal that number comes from the deal itself and should be bound as context. I kept it in to show this exact trap: **a required field forces a confident answer**, whether the input supports one or not. ## The second answer: classification The other job I described was classifying content into the IAB taxonomy. There, the shape is mostly one enum, so I tested just the enum. I loaded all 704 IDs from [Content Taxonomy 3.1](https://iabtechlab.com/standards/content-taxonomy/) and used `cattax`, which AdCOM's list has as 9 for version 3.1: ```yaml shape: type: object additionalProperties: false required: [cattax, cat] properties: cattax: {const: 9} cat: type: array minItems: 1 maxItems: 3 items: {enum: "${context.taxonomy_3_1_ids}"} # all 704 ``` It compiled in about 4 milliseconds against DeepSeek's real tokenizer, so size isn't the concern. `181`, Casinos & Gambling, was accepted. `IAB9-7`, a deprecated-style 1.0 code, was refused, and so was a `cattax` of 1. That confirms what I said last time, with the same caveat as everywhere. A model can reach any of the 704, including the wrong one, and with an enum this size picking the wrong one is the whole risk. ## How it runs as part of inference 1. **Input rules run first.** No brief or no deals on file stops the request. 2. **The shape is bound and compiled**, in about 150 milliseconds on my laptop against DeepSeek's real tokenizer. 3. **The decoder generates.** 4. **Output rules run, then a trader reviews anything that failed.** This all happens at planning time, and that's a point I should have been more careful about last time. I said an exchange answers bid requests by the billions inside a `tmax` budget and that batch invariance matters there. It does, for anything that's actually in that path. But a large model isn't sitting in a 200 millisecond auction loop building requests. The realistic place for this is ahead of time, setting up a deal or a line item, or tagging a page in advance. The determinism argument holds up there, and so does the audit argument, without pretending this is what runs in the auction. ## What the fixtures showed These are hand-written documents testing each path, not model output. - A clean CTV deal request passed everything. - One with both a `site` and an `app`, and a 60 second minimum over a 30 second maximum, **was accepted by the decoder**. The rules flagged both. - A deal ID not on file was refused by the decoder. - An empty brief and no deals were stopped by the input rules. ## What's still missing Every exchange implements its own subset of OpenRTB, and I said last time that they describe it in tables. That's the piece a bundle really needs, because passing my subset says nothing about whether TripleLift or Magnite would accept the request. A bundle per exchange is what you'd actually want, and it's what I'd ask them to publish. VAST is XML and isn't in here at all. I went in expecting advertising to be about scale. What I got instead was a good reminder that a shape can only hold the questions it was built to ask. The rules did their job on what the bid request can express. The buyer's real concern fell through a gap that was in the shape from the start, and nothing in the toolchain noticed.