Miju LabsEverything else in the artefact taxonomy is a judgement about a finished output. A trajectory is the only unit that captures the process that produced it — and it is the only thing in this market that a lab cannot commission from a crowd platform, scrape from the web, or synthesise for pennies. It is also the thing Contra buried under "preference pairs" on its own homepage, and the thing Taste Labs has publicly announced it will build.
What a trajectory actually is
The Premiere Pro dataset card is the clearest public specification of the unit anywhere: 234 steps across 4 trajectories, sessions of 111–245 minutes, recorded on macOS as professional editors built vertical social reels. Per-step schema: trajectory_uuid, session_uuid, image (screenshot), thought, action_type, tool_call as structured JSON, execution_paths (MCP tool / keyboard shortcut / menu path) and preferred_execution (HF card).
Two design decisions in that schema carry the entire commercial argument.
The thought field derives from the editor's spoken narration, explicitly distinguished on the card from "model-synthesized rationales." The card's own claim: "The screenshots, the recorded action, and the pointer coordinates are the editor's real execution." You cannot synthesise a professional's mouse path through Premiere, and you cannot get their reasoning without them talking while they work.
execution_paths plus preferred_execution encodes the fact that expertise is partly about route choice. A professional knows there are three ways to ripple-delete and reaches for one of them without thinking. A trajectory that records only the resulting state throws that away; a trajectory that records the alternatives and which was chosen is teaching a model something about fluency, not just about outcomes.
Contra's public trajectory sets, by size: descript-video-editing (803 rows), photoshop-creative-design (294), gemini-creative-campaign (266), premiere-video-editing (234), firefly-creative-campaign (137) (HF org). Note the naming convention — these are named for the tool the designer used, not for a customer. Do not misread gemini- and firefly- as Google or Adobe deals.
The category has one occupant
This is the most defensible finding in the product research, and it deserves to be stated flatly.
Beyond Contra's eight Hugging Face datasets, there is no comparable public corpus of professional Figma, Photoshop or Premiere sessions with narrated intent. As of August 2026 the narrated-professional-creative-trajectory category has one public occupant, and its largest published dataset is four trajectories.
What exists adjacent to it, and why none of it competes:
General computer use. AgentNet / OpenCUA is the largest open human-demonstration corpus: 22.6K human-annotated computer-use tasks across Windows, macOS and Ubuntu and 200+ applications, collected with a purpose-built cross-platform tool capturing screen recordings, mouse/keyboard signals and accessibility trees, post-processed into state-action pairs with reflective chain-of-thought. Licence MIT (arXiv 2508.09123; HF). It is a hundred times bigger than anything Contra has and it is general — booking flights, filling forms, navigating settings. It is not professional craft work and it carries no narrated design intent. Neither the paper nor the card discloses annotator recruitment or pay.
Environments, not corpora. OSWorld, WebArena, Mind2Web and Windows Agent Arena are evaluation harnesses with small human-verified task sets. They do not compete with a trajectory product; they create demand for one.
Code trajectories, which are abundant — and the reason matters. SWE-bench-derived agent rollouts, open agent-scaffold logs and synthetic execution traces exist in volume and cost almost nothing, because code has an oracle: tests pass or they fail, so trajectories can be filtered automatically. There is no automated filter for a good Figma session. That is the structural reason creative trajectories are scarce and will stay scarce — and it is the same argument, from the other end, as The oracle problem and the logic-bug wedge in security, where the absence of an automatic verifier is precisely what makes human work non-substitutable.
The thousandfold price ratio you have to defend
AgentTrek synthesises GUI trajectories by replaying web tutorials and reports "a cost of just $0.55 per high-quality trajectory without human annotators" (arXiv 2412.09605). Its framing of the problem is the honest statement of the market: existing approaches "rely on expensive human annotation, making them unsustainable at scale."
A human trajectory costs $400–$1,500 all-in [WEAK — my planning band]. That is a ~1,000× ratio.
Build the human figure from the ground up. A 111–245 minute session at Contra's published "up to $100/hr" is $185–$410 of contributor time (contralabs.com/jobs). Then add: screen-and-audio capture tooling; a narration protocol and a practice run, because narrating while working is a learned skill and first sessions are usually unusable; privacy review — Contra's own card records screenshots reviewed to mask "the screen-recorder window and its live camera, account avatars and names, profile photos"; action extraction and alignment of narration to steps; and QA. Contributor time is under half the cost.
The entire commercial case for the human version rests on one claim: that the narrated thought and the preferred_execution field contain information the synthetic pipeline cannot recover. That claim is probably true — a tutorial replay has no access to why an editor rejected the first three grades — but a buyer looking at a 1,000× ratio will make you prove it, and the proof is an ablation: a model trained with and without your trajectories.
Budget to run that ablation yourself. Asking the buyer to run it means asking them to spend engineering time proving your pitch, at the exact moment they are deciding whether to spend $1M on the comparison corpus instead. It is the single highest-leverage piece of first-party research on the product side, and it doubles as a paper.
The constraint that shapes the product
Contra's explicit differentiator is that these are recordings of real client briefs — professionals doing work they were already doing, rather than performing a task for a data vendor (HF card). It is a good marketing line and a bad legal position.
A trajectory recorded during a live client engagement captures the client's brief, brand assets, unreleased campaign and possibly their customer data, on screen, in frame. The contributor almost certainly does not have the right to license any of it. And the injured party — the client — has no contract with you, so there is nothing to negotiate and no cap to rely on. Consent from the contributor does not cure it; the contributor is not the rights-holder.
Contra's card records that "editors participated with consent" and that screenshots were masked for recorder windows, avatars and profile photos (HF card). That handles the contributor's privacy. It does not touch the client's confidential material, which is the larger exposure and the one that survives redaction badly — you cannot mask a campaign you are editing.
So de-novo briefs are not a nice-to-have alternative to real client work. They are the only structurally safe version of the trajectory product at scale. A brief written for the purpose, with assets you commissioned or licensed, recorded by a professional who is being paid for that session and nothing else, produces a clean chain of title on every frame. It costs more per trajectory — you pay for the brief as well as the session — and it loses the "this is real work" line. It gains the ability to sell the corpus to a regulated buyer, to warrant provenance, and to survive diligence. Copyright is the weakest thing you own and The clause that expires your corpus in year ten carry the drafting.
The secondary effect is that de-novo briefs solve the contamination problem at the same time: an original brief that was never web-published is provably absent from pre-training, which is a property the buyer increasingly wants and cannot verify any other way.
Who else is in the lane
Two developments say the lane is not empty for long.
Mercor is already selling this. Its Senior Design Expert posting — $150–$250/hr, 4–5 hours per task, portfolio required — states that "the client wants to capture how expert designers think: how you approach a problem, what separates strong craft from weak," with sessions structured and recorded remotely (Mercor). That is narrated expert design reasoning, sourced from Pentagram/Wolff Olins-calibre designers, at the top of the price range, for an unnamed frontier client. The lane is occupied by the biggest player in the category, with the good version of the product — see The generalists in creative.
Taste Labs has published the roadmap. Its Requests for Research page lists "design intent inference from edit sequences," "design history/versioning beyond code diffs," "agent process fingerprinting" and creativity evaluation of process rather than outputs (tastelabs.com). That is the trajectory asset, announced in a research agenda months before shipping. Taste has four engineering/ML roles out of eight open; Contra has five open roles and zero ML engineers.
The read: trajectories are the right product and the window is measured in quarters, not years. Whoever gets a few hundred clean, de-novo, narrated professional sessions into a lab's training run first owns the reference implementation — and that is the argument for sequencing them ahead of comparison volume, which the arenas will always win.
No public ablation exists showing what narrated human creative trajectories buy over synthetic ones. The 1,000× price ratio is documented on both sides; the value ratio is asserted on one.
No comparable corpus discloses annotator pay — not AgentNet, not any of the environment projects — so there is no external check on the $400–$1,500 band beyond arithmetic on Contra's published hourly rate. And no vendor has published a per-trajectory sale price, so the gross margin on the most defensible unit in the taxonomy is entirely unobserved.