Creative work and taste

The lane is chosen, so the question is what is true about it. Mercor already sells the premium version at $150–250/hr, the arena layer holds 6,047,075 image votes collected free, the data costs $0.91 a comparison, and a 10–25x pricing contradiction sits unresolved at the centre. Three things are genuinely in the client's favour and one of them is a property right no US competitor can hold.

Taste has no oracle, so inter-rater agreement is the manufactured one — which makes the panel, not the file, the product. Build a calibrated, named, consented panel and sell the sealed evaluation content that falls out of it; run it first in product and UI design on de-novo briefs; and remember that the corpus is a $1M line item a lab insources in one meeting.

The dangerous competitor is not Contra. It is the company whose founder already sold to foundation labs, whose fourth and fifth hires train models, whose raters nominate each other — and which has published a research agenda promising to build the one asset Contra actually owns.

The $60M ARR figure is self-reported, has no consumer revenue behind it, and is contradicted by the round it was announced with — a company at $60M does not raise a $7.9M seed from Index. What is real is 5.3M people voting for free, which is Pick-a-Pic commercialised.

Six million image votes collected free, $100M annualised on 28 people, and 80% of the preference data retained as a strategic asset. The specialist loses this fight on volume permanently — and Yupp is the corpse that proves both that generic preference is unsellable and that the demand side is thinner than the supply story implies.

The most rigorous public human-judged design benchmark was built by an expert-data company a year before Contra Labs existed, from 194 hand-picked experts who appear to have been unpaid. Publishing a benchmark for credibility is not a novel move here — it is the category's standard customer-acquisition tactic.

Mercor already sells the exact product — narrated senior design reasoning — at $150–250/hr, from Pentagram and Wolff Olins pedigree, to an unnamed frontier client. Contra's 'up to $100/hr' is the middle of the band, not the top of it.

Adobe writes the recruiting pitch for you: an undisclosed bonus paid at its discretion, no opt-out, and a payout floor cut from $25 to $1. Meanwhile Behance, Dribbble and 99designs sit on more creative supply than every funded competitor combined, with no announced AI-data programme between them.

One lab has a disclosed procurement function and it covers Sora — but the word 'aesthetic' appears in none of its 753 postings. Google discloses 366,569 ratings and never a price. And across the whole literature the same split recurs: designers write the prompts, a general crowd supplies the judgement.

The 'probable first customer' hypothesis, tested against 1,562 job postings and largely failed: one dedicated AI-quality-evaluation role, zero named vendors, and $31.6B of vibe-coding valuation employing nobody to evaluate design quality. Canva is the exception, says so in writing, and has already built its own supply.

$50–300M of externally-purchased spend, central $100–150M, built from a $6–10B expert-data market with a $3.8B verified floor times a 1–3% creative share. The uncomfortable half: published practice pays $0–16/hour for aesthetic judgement, and a field-defining benchmark cost $13,433.55.

Freelancers bill 22.4 hours a week against a 50–60% utilisation benchmark, so hours 23–40 earn zero — which means the correct price for a bench hour is at market, not above it. Recruit the majority-self-employed segments, not the highest-paid ones, and expect to clear at $60–85/hr with no premium for creative expertise.

Roughly 800–1,000 named, peer-selected senior creatives are published with title and country every year by the award bodies, at zero cost to identify — and ADC's discipline taxonomy is your segment taxonomy. But awards are a 20% sample, so use the registers as the seed and never as the funnel.

Stated attitude is 81%, revealed behaviour is 10%. The Sora revolt is the strongest evidence that creatives object to being unpaid, uncredited and approval-gated rather than to evaluating AI output — every demand was a term sheet, not a principle. But 51% object to who profits, not to being underpaid, and money does not fix that.

There is a published pass mark: human experts hit 68.9% on the Visual Aesthetic Benchmark where the best frontier model manages 26.5%, and direct ranking yields substantially higher inter-annotator agreement than score-derived ranking. Collect rankings, screen at roughly 69% agreement with expert consensus, and treat peer nomination as a seeding mechanism rather than a scaling one.

The competitor published its own cost base: $90/hr, 10 designers, 14,400 comparisons, about $13,000 — roughly $0.91 per pairwise judgement. Extrapolated, a frontier-grade 1.17M-comparison corpus costs about $1M in direct labour, which is a build-versus-buy decision a lab makes in one meeting. Everything defensible in the taxonomy sits above that line.

The narrated professional-creative-trajectory category has exactly one public occupant and its largest dataset is four trajectories. Synthetic costs $0.55 against $400–1,500 for a human one — a thousandfold ratio you must justify with an ablation you run yourself — and trajectories recorded during real client briefs are a liability, not a differentiator.

The popular version — that model rankings invert between ideation, mockup and refinement — rests on 28 evaluators and expires with each model release. The generalisable half is that inter-rater agreement rises monotonically with phase, 0.345 to 0.549 on ad images, which is a structural claim about creative work and yields a pricing schedule nobody else has.

Enormous headroom in taste, none in compositionality: the best frontier model scores 26.5% on the Visual Aesthetic Benchmark against 68.9% for human experts, while GenEval drifted to 17.7% absolute error and had to be replaced. Half the design benchmarks that matter were built by companies selling into the same market.

In security a perfect oracle makes a product and no oracle makes a permanent human service. Aesthetic judgement has no oracle at all — which is why inter-rater agreement has to be manufactured into one. A single opinion has no error structure; a calibrated panel does, and that is the only part of the stack a buyer cannot rebuild from the delivered file.

Article 11 of the Database Directive restricts the sui generis right to EU nationals, residents and companies — so Contra, Taste Labs, AfterQuery and Mercor structurally cannot hold it over their own corpora. But BHB v William Hill excludes investment in creating data, which is exactly what a commissioned dataset is. The right is available and the default posture forfeits it, which makes separate ledgers for verification and presentation an operating decision to take in month one, not year three.

A rating is not a work in the EU — no free creative choice — so contract, trade secret and the database right do all the work. The decided cases point the same way: Bartz priced acquisition at about $3,000 a work while calling training itself exceedingly transformative, which means provenance rather than use is what generates buyer liability. And two European courts gave opposite answers on whether weights reproduce works within five days of each other.

Everyone cites §29 UrhG's non-assignability, which is handled by drafting an exclusive licence. The clause that actually bites is §40a: a lump-sum exclusive grant becomes non-exclusive by operation of law after ten years, and a per-brief flat fee is a lump sum. The fix is a small ongoing royalty, which also defuses §32a — and the wider finding is that Contra's creator-friendly terms are simply what German and French law compel, so a European operator faces no trade-off between attractive terms and compliant ones.

Enforcement powers over Article 53 went live on 2 August 2026 with fines to 3% of global turnover or €15m, and compliance is visibly patchy — Anthropic, Mistral and xAI substituted prose for the template, and Microsoft's Phi-4 summary failed an independent quality assessment. A commissioned corpus is the easiest line a lab will ever write into that template. And since no lab has published data-supplier requirements for creative content, the first credible published provenance-and-consent standard sets it.

A sequence with a disproof attached to every step: seed from the published award juror rosters but never treat them as the funnel, ask the attitudinal question before the skills test, screen at roughly 69% agreement using rankings rather than scores, put the §40a royalty clause in the agreement before the first German contributor signs, split the verification and presentation ledgers in month one, and run the synthetic-versus-human ablation yourself rather than asking a buyer to run it.

Ranked by how much the answer moves the decision. The top item is that nobody has ever asked creatives directly whether selling judgement differs from selling work — the load-bearing assumption of the entire business, closable for a few thousand dollars with a segmented 300-respondent survey. The second sets the whole cost structure and no competitor has published it.

The prose companion to the people graph. Six conversations come first, led by the only person found anywhere who has already done the hard job at a frontier lab — recruiting sceptical professional creatives with a real budget. The structural finding is that sixteen of seventy people have no verified edge to anyone, almost all practitioners and critics, because there is no warm path from the AI industry into the design community.