The data thesis
The governance commitments come first
Customer context is the customer's asset. That is the platform's core promise, and it is senior to everything below. Nothing a customer captures or builds on their decks is used as training data - not by default, not quietly, not in aggregate - unless they explicitly opt in to a clearly named research pool. As of this writing no such pool exists, no customer data exists, and the only corpus of the kind this page describes is Deckwright's own: the captures and pipeline records produced building Deckwright with Deckwright. If the thesis below is ever pursued, it will be pursued on consented data or not at all.
The premise: what an MMC payload actually contains
A Multi-Modal Capture is not a screen recording and not a chat log. It is a set of discrete, timestamped signal streams - voice, cursor kinetics, gesture, dwell and hesitation, canvas mutations, runtime and console events - fused on one clock. A human working on the Bench is not merely interacting with a space; they are creating context and calling it out at the same time: naming the thing while pointing at it, describing the behavior while demonstrating its timing, correcting the result while reacting to it. Deixis resolves, because the system knows where "there" was pointing when "put that there" was said.
And because capture is fused to a governed production pipeline, the record does not end at expression. A captured intent flows into specification, build, independent review, and verification against a predeclared test. The unit this produces is a verified intent trajectory: intent, action, and outcome - with the outcome graded against a contract fixed before the work began.
Why that data class is scarce
- Internet video has planetary scale and no intent: the world changes on screen and no one says what anyone was trying to do.
- Robot teleoperation has intent and no scale - and almost no natural language riding along with the action.
- Annotation pipelines manufacture grounding after the fact, at cost, by workers reconstructing what a stranger probably meant.
A verified intent trajectory has all three properties at once: purposeful action, contemporaneous natural-language grounding, and a ground-truth outcome label. It is produced as exhaust from work someone was doing anyway, inside a tool that pays for itself. To exist at all, it requires a capture layer fused to an execution engine with verification gates - which is a precise description of Deckwright's architecture and not a common one.
The claim, sized honestly
World models pretrain on internet-scale corpora. No plausible volume of MMC data competes with that, and this page does not claim Deckwright will train a world model. The claim is the layer above: small, dense, verified-intent data is the shape of what grounds and aligns large pretrained systems - the decisive thin layer, not the bulk. The nearest analogy is preference data: a corpus thousands of times smaller than pretraining that nonetheless determined how the resulting systems behave. Verified intent trajectories are a candidate for that role in spatial, creative, and agentic domains: how humans actually direct work in space, labeled by their own voices, graded by real outcomes.
The jump: depth and 3D
On a 2D canvas, this is interface telemetry with intent. The architecture, however, is sensor-agnostic: streams on one clock. Adding depth capture and 3D creation - LiDAR, volumetric canvases, room-scale gesture - adds streams, not a new system. At that point the data class changes name: from instrumented interface use to embodied intent trajectories - purposeful, narrated, verified action in three-dimensional space. That is the gap between what video corpora have and what robotics needs, and nothing at scale currently occupies it.
The consumer path: a pool that is consented by design
The Bench now ships on iPad as a released surface - the same canvas, with Apple Pencil as a first-class capture stream: stroke, pressure, tilt, and timing fused on the same clock as voice and canvas. That release opens a path this page would otherwise leave abstract: an App Store edition for casual creators - talk, sketch, and generate on a hosted bench - offered as a small monthly subscription, or free for users who explicitly opt in to a clearly named training pool. That pool would be the first of the kind the governance commitments above describe, and it is deliberately consumer-shaped: exploratory creation, never production work, never a customer's deck.
The asymmetry is the point. Developer data is encumbered - employer IP, confidentiality, code that is itself the customer's asset - which is why the commitments above exist and why they hold. Casual exploration carries none of those stakes, and it produces the same scarce shape: sketch and voice as contemporaneous grounding, generation as the action, and the accept-redraw-iterate loop as a weak outcome label on every trajectory. Volume where consent is easy; structure where the platform already lives. The two moats stop competing: customer context stays the customer's, and the model layer learns from a pool built to be consented from the first install.
Downstream, that corpus is aimed at exactly the "decisive thin layer" claim above: how people direct spatial and creative work - narrated in their own voices, in pencil and gesture, graded by their own reactions - as grounding and alignment data for the domains where that layer is missing. Status: strategy, stated in public for the same reason as the rest of this page - so it can be judged, and so the consent-first shape of it is on the record before it matters.
Why this page exists
Deckwright's business does not depend on this thesis: the platform is sold as a context development platform, and its defensibility rests on the capture layer and on each customer's own accumulated context - which, per the commitments above, stays theirs. But an evaluator asking "what does this become at scale" deserves the candid answer: a consensual accumulation of the scarcest data class in AI, growing as exhaust from a tool people use for its own sake. We state it as a thesis, in public, so it can be judged - and so our governance position on it is on the record before it matters.