Every figure carries four things, or it does not publish.
Its population, its basis, its definition and its read time. A figure missing any of the four is a draft, not a claim.
Verellax publishes company-named findings, sells reports priced on their defensibility, and licenses derived intelligence to institutions. Every one of those motions survives exactly one question: how do you know? This page is the standing answer, cut from the internal master. One methodology, three rooms.
Two AI populations exist and are named on every figure: roles that mention AI anywhere in the specification (12,547 live) and AI-central roles, where AI is core or supporting the work as written (5,674 live). They are never interchanged.
What does the instrument measure?
Verellax measures the distance between the AI a company describes and the knowledge work it has designed. The unit of analysis is the job specification, read as an organisational disclosure. When a company writes a job spec it discloses, usually without intending to, how it distributes tasks, who decides what, how it relates human judgement to machine output, and whether it is building a new organisation or refilling an old one.
The instrument answers one question per role: as written, does this role define how decisions divide between the person and the AI system? It does not read what happens after the hire, and it never attributes a state of mind, a policy, or an absence to an organisation. It describes documents.
Where do the signals come from?
Three records per company, each with its own provenance. The distance between the first record and the second is the say-do gap. The third record is how velocity is read.
How much of what is collected is published?
Job specifications are perishable: a posting removed from a board can never be re-collected. Capture is wide, daily and free, banking raw text durably across every wired company. Extraction drains from the store on a per-cohort weekly cadence, applying geography and function filters so analysis spend lands only on in-scope roles. Out-of-scope captures are retained, a banked reserve that can be activated per market without re-collection.
Career sites are re-read on a schedule of seven to thirty days by source; a specification whose source falls outside its window loses live status until the next read, so the live count is a claim about sources read, not sources listed.
Raw counts are never presented as coverage. The gap between raw and canonical is mostly out-of-geography capture retained by design.
How is the work divided?
Every specification is decomposed into four layers of work across fourteen sub-categories and 830 canonical skills: execution (technical, functional, domain), coordination (orchestration, alignment, enablement), judgement (operational, strategic, evaluative, taste) and AI-adjacent (tool use, tool building, governance). Each spec carries a ratio composition across the four layers, weighted by responsibility described rather than by skill count.
The decomposition gives each role a shape and each seniority a curve: execution falls from 65.1 per cent at entry to 21.9 per cent at C-suite, while judgement climbs. The execution layer is the layer a model absorbs first, which is why execution load is a load-bearing measure. Mapping completeness is reported per layer, never as one blended number.
The AI-adjacent layer is the weakest and it is the one carrying the thesis. Until coverage is flat, skills figures are composition claims, never trends.
What are the four tests?
Every specification is scored against four boolean tests during extraction. The composite runs zero to four. A role scoring zero says nothing, as written, about how decisions divide between the person and the system. The composite measures whether an organisation is designing the human-machine relationship or hiring around it.
Every test carries a recorded definition. The current definition anchors the overrule test to the AI system itself rather than to any escalation clause, which narrows the population and lowers the share. A change of definition that moves a figure is disclosed on the corrections page as a dated note. No trend is drawn from the share until the base has a year of history.
When does a cell publish?
Sectors classify; cohorts benchmark. A cohort is a named membership list independent of sector labels. A company cell publishes at 8 or more specifications. A sector or cohort cell publishes at 8 or more members. Below the floor a cell renders as counts beside shares with the label directional, and never supplies a headline, an annotation or a readiness tier. Floors are a founder decision; no process moves them. The Index and every public page report by sector. Cohorts are named only where the company has bought its name in.
Every published figure is a row in the claims register: the claim text, its source query, its population key, its basis, its definition, its floor rule, its value, its n, the companies behind it and its read time. An artefact renders from one call that refreshes every row in one transaction at one timestamp, and that timestamp is printed beside every figure. Companies are named only in the positive, only above floor, and only with notice. Deficit findings publish at cohort level, without exception. If a figure is wrong it is corrected in public with the query shown.
The regulated sector designs best, and the designed minority exists in every cohort, so the deficit is a choice.
What are the limits?
A method that never fails is a method nobody checked. The August mapping acceptance check failed at eleven disagreements per hundred and forced a correction pass that killed a better-sounding headline; the failure and the correction are both on the record.