Verification Is the Real Product in the AI Data Business

01What actually got priced

Valuing a data company by labor cannot explain the gap between $300 million and $3.2 billion. Fake expert tasks, drifting preferences, rubric holes, biased judges — and models finding holes in the scoring standard itself: every link can be gamed. The Rubric Dropout paper trained RL on the training judge's scores: they kept rising while the gold judge's rose and then fell, ResearchQA dropping 22 points.

The Meta-Scale deal split 400+ Hacker News comments into two camps: Scale as Filipino annotators wrapped in software, Meta buying Alexandr Wang; or Scale knowing which data and methods OpenAI and Anthropic used for reasoning RL, Meta buying that knowledge. The second is closer, plus one addition: the valuable part is the sandbox, test suites, meta-judges, the hack-to-zero mechanism, and its record of customer demand. Ask not how many people a vendor has, but how hard its verifiers are.

02People are leaving the training loop

Post-training's direction is clear: PPO-RLHF needed humans labeling preferences; DPO removed the reward model and the RL loop; GRPO removed the critic; RLVR replaced people with verifiers; the newest unsupervised RLVR has the model generate its own supervision. SC-GRPO already does token-level credit assignment from the model's own rollouts alone.

Two human roles remain. Build verifiers — ground truth for math, test suites for code, result-checkable environments for professional tasks; Anthropic is paying $167,000 for four months for this, and the posting requires tests models cannot game. Arbitrate the long tail machines cannot judge. Annotation keeps losing value; verifiers keep gaining it — teams still scaling annotation capacity should count how many quarters the premium lasts.

03The craft is in protocols, not headcount

OpenAI published the entire PRM800K protocol: a 75% agreement screen for annotators, 10-20% QC items per round, stepwise probing that stops at the first error, 10 rounds of active learning.

Frameworks can be copied; calibration cannot. One Chinese buyer purchased Yunnan dialect data from three vendors, and the tone and vocabulary labels disagreed — acceptance was impossible. Translating a spec is easy; getting three hundred annotators to judge consistently against one spec is the moat.

Two things matter above protocol. Effective tokens beat token counts: the D in scaling laws was never raw corpus volume — ten units of low-quality web text are worth less than one unit of high-quality code and math, so "not enough data" is often a wrong-ratio problem. Difficulty must match the model's level: 95-point questions for a 60-point model give no gradient; 30-point questions only align format. Adjust the mix weekly by 5-10%.

04Evals: the grading was never verified

A score is worth what it drives: release, procurement, compliance decisions. Braintrust is valued at $800 million; DeepEval charges $2,000 per month for exactly that.

Once a benchmark becomes a training set, it is dead. OpenAI stopped reporting SWE-bench Verified in February, with two stated reasons. Contamination: frontier models could reproduce original fix patches or recite task descriptions verbatim. Worse: of the 27.6% of tasks models most often fail, at least 59.4% of test cases are themselves defective, marking functionally correct submissions wrong.

The sharpest Hacker News comment: the headline blamed contamination, but the buried story was test design defects — a measure never calibrated against ground truth, used to report scores for two years. The "Verified" label came from professional developers screening each task by hand; even that review did not make the tests correct, and the Berkeley ten-line pytest hook that scored full marks on SWE-bench is still remembered.

Back to Scale: Google and OpenAI cut cooperation because a vendor who sets and grades your tasks, with half its equity in a competitor's hands, cannot receive your training direction. Leaderboards that sell data themselves are inherently untrustworthy; question-setting, grading, and training must be separate interests.

Eval standards are rising: fuzzy judgments decomposed into yes/no checks, human-machine agreement measuring the tool's own variance, Tool Head Accuracy tracked separately (a right answer through a wrong tool chain counts wrong), trajectory edit distance counting wasted steps. Within two years, trajectory-level, auditable, cost-aware evaluation becomes a hard enterprise requirement; text-answer-only tools depreciate.

05Three practical moves

Enter the verifier layer, not the annotation layer. Build test-suite ground truth for professional verticals: legal judgment chains, clinical treatment pathways, financial compliance checklists. Each environment built is an asset; two or three people deep in one vertical is enough.

Copy protocols, not scale. PRM800K's QC protocol, binary rubrics, human-machine agreement rates, and release gate thresholds are public and transferable to Chinese scenarios. The scarce thing is not the method; it is a team that executes every item.

Do not overtrust models making their own data. Synthetic data absorbs low-end demand, but the bottleneck is industry raw material, copyright and privacy handling, verifiable workflow construction. Pipelines produce format, not real-world dirty data and edge cases; experts plus real material stay irreplaceable upstream.

The industry looks labor-intensive and is actually trust-intensive plus protocol-intensive. A $3.2 billion company and a $30 million company differ not in headcount but in the verification chain behind every batch. People scaling annotation capacity today will look, five years out, like the people who scaled call centers.