Most AI-and-labor numbers measure one of two things and rarely say which: what could be automated, or what demonstrably was. DMX publishes both, separately, and publishes the distance between them. The demand leg is demand for hours, in the same unit as supply.
Status: The demand leg follows the supply launch. D-R launches on public sources DMX captures itself; D-X and DRR are research series. Nothing here has a public track record yet, and this page describes the specification in the methodology paper.
Two versions of A are maintained and published apart. A-R, realized, feeds D-R; A-X, exposure-implied, feeds D-X. They are never averaged or blended into a single headline, because realized and potential displacement are different kinds of evidence and an index that blurs them cannot be audited.
A-R admits only employer-attributed displacement, converted to hours under published inclusion rules written to be applied without judgement: WARN Act filings and SEC filings that name AI or automation as a cause, corroborated where required against payroll-scale evidence, dated to the effective separation date, subject to a 90-day seasoning window and a 25 percent reversal test, aggregated over a trailing twelve months, and recorded in a public event log. If it cannot be attributed, it does not enter.
At launch A-R is sourced from DMX’s own capture of state WARN filings — 51 jurisdictions surveyed, most under automated weekly capture with the notice documents themselves archived, since attribution language lives in the documents — and from SEC EDGAR. No licensed announcement database is used at launch, and that absence is disclosed in a per-print source-coverage statement.
A-X estimates potential displacement: for each occupation, a task-level exposure share, times a measured adoption rate, times that occupation’s share of hours. Exposure uses a published task rubric applied to the O*NET task file, re-rated annually by DMX with disclosed frontier-model raters under a published aggregation rule; adoption is a composition of a worker survey (level by occupation group), an observed-usage index (within-group shape) and BLS employment weights, cross-checked against Census firm-level data.
Two properties are stated without apology. Exposure measures potential, not realized displacement. And A-X systematically overstates hour removal, because automated tasks partially reallocate within roles rather than disappearing. It is published anyway, because potential displacement is the quantity forward-looking users need — and because pairing it with A-R turns its bias into information. D-X is research-tier only and never enters settlement.
DRR answers the question every institutional user of AI-labor analysis is asking: how much of theoretical exposure is landing as realized, attributed displacement? Rising toward one, exposure is converting into outcomes. Persistently low, the story is reallocation, augmentation or attribution lag — a different world to underwrite. DRR is a standalone licensable research series with its own history and revision log; it is not a settlement object.
No causal identification of AI’s effect on employment, anywhere. A-R is an attribution measure under published rules; A-X is an exposure model. The composites are defined over those published constructions, not over an unobservable true effect. Firms have incentives to cite AI when cutting costs for unrelated reasons, and to avoid citing it when displacement is real; A-R therefore carries bias in both directions, mitigated by corroboration and public logging and disclosed in the paper.
The methodology paper is available on request; the FAQ answers the common ones on coverage, revisions and what each tier permits.