Format documentation
An opt-in, fully skippable ad format for television, in two placements: the pause the viewer chooses and the break the show takes. This page documents how it works end to end, in the open, so anyone in the ecosystem can evaluate it, build against it, or hold us to it.
Version 0.10 draft / Published August 25, 2026 / Status: reference implementation in development
Linear television trained viewers to endure advertising, because it never had a moment when attention was available and nothing was playing, and it standardized ad lengths because the break architecture demanded it, not because attention does. Intermission is built on the opposite premises: the ad plays only when invited, and it lasts exactly as long as it deserves to.
The card presents an option, not a takeover. Playback begins only on an explicit viewer action. If the viewer declines, the card cost them nothing: silent motion, no sound, no interruption.
Once playing, the ad is skippable at any moment with no penalty and no gate. There is no forced duration and no countdown to a skip button. And the skip is never waste: it is the format's correction signal, a labeled no that refines the audience the way an opt-in refines it as a yes.
The creator's video runs whole, at its source length, never cut to fit a slot. Long pieces carry chapters the viewer can move through; the work itself is untouched. The format defines no maximum duration. Where long stops paying is discovered by the yield, not decreed by the spec.
The unit of measurement is the opt-in and the chosen seconds that follow it. Both are explicit, volunteered, and reported in place of exposure counts.
The card is a 1920 by 1080 composition with exactly one moving element: the teaser, a short silent loop of the advertiser's real social video. The silence costs the creative nothing, because creator video is already made to be watched without sound: captions burned in, the story told visually, the gist legible on mute. The format ports that video in its native mode, into contexts that are sound-off by nature. The shape ports just as cleanly: a 9:16 clip seats into the center of a 16:9 screen almost exactly, which is why Composition A exists at all. The thirds are what vertical video naturally leaves behind on a television. Everything around the teaser is static and fully branded, rendered from the advertiser's catalog data, logo, colors, and type. Motion never competes with motion.
The teaser and the feature it opens are separate creative decisions. The teaser's job is the opt-in; the feature's job is the watch. They need not come from the same file or share a shape: a 9:16 creator clip can tease a 16:9 feature, the way a trailer is built from the moments that best sum up the film rather than from its opening scene. The feature plays full screen at its own aspect, whole and unedited. The teaser remains the only excerpt the format ever makes, and it can be cut from the feature, from another of the advertiser's videos, or chosen by the engine as the strongest tease this product has. The engine learns the best teaser and the best feature independently, which is why an advertiser benefits from supplying more than one video per product.
Each composition ships in two variants: one carrying its own play affordance for placements where the player honors it, and one reserving clean space for platforms that composite their own interactive chrome. Composition choice is set per advertiser or per product, or left to the optimization engine, which treats composition as a tested variable like any other.
The viewer pauses their content and the card renders on the paused frame. An opt-in plays the feature, skippable from the first frame, chaptered when long, closing back to the paused show on skip, completion, or resume. The viewer's show is exactly where they left it.
The pause placement carries no incentive mechanics of any kind. Pressing play costs the viewer their own time and buys them nothing except the content of the ad. For that reason, pause opt-ins are the format's cleanest signal: an opt-in here is active reconsideration or pure interest, with nothing else mixed in.
The break opens with The Invitation: a six second unit built from the same card language, product, silent loop, play affordance. Unpressed, it completes as a standard bumper and the break proceeds normally. Pressed, the feature plays in place of scheduled spots, whole and skippable.
The yield condition, stated plainlyEvery break the format touches has a baseline: what the publisher's standard fill would have earned for that time, computed from the publisher's own rates at decision time. Chosen playback accrues revenue against that baseline per chosen second. The break ends when its yield is met, whether that takes fewer seconds than the scheduled break or more. If a play ends short of baseline, remaining time is filled to the baseline by other units. An opt-in, by itself, never shortens a break: only watched seconds can complete a break early, and unwatched ones cannot. Skipping the chosen video returns the viewer to the remainder of the break, not to content.
What this means for the signalBreak opt-ins carry a second motive the pause cannot: choosing what fills the break. The format's measurement treats the two placements as distinct signal tiers rather than pretending they are the same. Watch depth within a chosen play is voluntary in both placements and is reported identically. A play that runs past the scheduled break length is the strongest signal the format produces: the viewer paid net time from their own evening to keep watching a product ad.
ScopeVariable length breaks require the player or stitcher to own the session clock, which is the norm in on-demand and most free streaming. Live and linear breaks with hard out-times constrain the play to the remaining break, with the unbounded experience reserved for the pause.
Attention isn't limited by time. It's limited by interest.
The format's two most distinctive properties, no maximum duration and skippability at every second, are one mechanism. Unbounded length is only honest because exit is instant and free. Instant exit is only meaningful because there is no clock running it out. Each makes the other safe, and together they produce something no fixed-length format can: an audience that re-qualifies itself continuously. The viewer still watching at second ninety is not a targeting estimate; they are ninety consecutive decisions not to leave. A completed three minute play is not one impression. It is one hundred eighty acts of self-selection.
The true-upSelection begins from priors: match provider input, context, and the household's own history. Priors are guesses, and in every other television format they are also the final answer, because a non-skippable impression returns no verdict on whether the guess was right. Here, every play grades its own targeting. The opt-in is a labeled yes. The skip is a labeled no: this living room, this product, not now, flowing into the household representation exactly as opt-ins do. The skip is the true-up. Audiences reconcile against observed choice continuously, so targeting error decays with volume instead of compounding with it.
The economics of the exitUnder chosen-seconds pricing, the skip is also the advertiser's protection. Spend ends the moment interest does. Nothing in the format monetizes a viewer who has stopped watching, which is the property that makes an unbounded ad safe to buy, not only safe to watch.
Most advertising resolves identity and creative inside a real-time auction measured in milliseconds. Both Intermission placements invert this. A viewing session announces itself when content starts; the pause or the break arrives minutes later. Phase one, at session start: a match provider resolves the household to recent product interest, selection runs at leisure, the decision is written to a session cache with a freshness window, and card assets prefetch. Phase two, at the pause or the pod: a cache read. The card renders instantly, with no auction latency in the viewer's moment.
Selection is a learned matching system, structurally similar to the two-tower retrieval architectures behind modern content feeds. One representation is maintained per product, derived from catalog data and the video content itself. One is maintained per household, derived entirely from chosen attention: opt-ins, chosen seconds, skips, and completions. Matching scores are similarity between the two, recomputed as events arrive, and consumed by the phase-one decision described above.
Two auxiliary inputs feed the household side without defining it: session context, such as content genre and daypart, and consideration signal delivered through the match provider role. Both matter most when a household's own event history is thin, and matter less as it grows. The system's targeting asymptotically depends on one thing: what living rooms choose to watch.
Intermission does not own identity, by design. Matching a browsing session to a viewing household is performed by a match provider: an identity vendor, supply platform, or commerce network that already holds both sides of the join. The role is a neutral interface, open to any qualified implementer. At session start the match provider posts a session token, the matched advertiser, product identifiers with confidence scores, and a freshness window. The format server never receives user-level identity, only opaque session tokens and the engagement events the viewer generates. Merchant-side capture is consent-aware and uses hashed identifiers only. Sensitive categories never produce product-level cards.
Intermission speaks the tracking vocabulary CTV players already implement. The format's defining interactions have had names in the VAST standard for years. No proprietary SDK is required to count what matters.
Event names per the IAB Tech Lab VAST 4.x tracking vocabulary. Intermission adopts the standard vocabulary as implemented; this documentation is independent of, and not affiliated with or endorsed by, IAB Tech Lab.
A stable HTTPS URL per creative version: composed MP4 and poster, CDN-served, long-cached. The lowest-friction handoff, and a telemetry-blind one.
The same content packaged as an HLS manifest. Required for feature playback, where duration is unbounded and chapters are manifest markers.
A VAST 4.x document referencing the renditions with the event vocabulary baked in. The integration that makes the format measurable, which is the point.
Second placement added, The Break, with The Invitation unit, the yield condition, signal tiers, and live/linear scope. Unbounded tenet rewritten: source-length playback, no cuts, no defined maximum. Chosen seconds adopted as the unit of measurement. Selection documented as a learned matching system. New section: the skip as the format's correction signal and the pairing of skippability with unbounded length. The teaser and the feature defined as separate creative units with independent sources and aspect ratios; silent-native and vertical-native rationale documented.
Initial publication: position, card compositions A and B, the opt-in interaction, two-phase runtime, match provider role, event vocabulary, delivery surfaces.
Implementation notes for players, stitchers, and match providers; card asset specifications; the reference implementation.