Before computing a single historical value, we committed the tests our index had to pass — and the tests it had to fail — to a timestamped record. Then we ran them on 43 quarters of federal data. The contraction we set out to measure is real, large, and specific to young workers. The attribution everyone assumes — that it concentrates in AI-exposed occupations — failed our own falsification tests. We publish both findings with equal weight.
A lighthouse that misplaces the rocks is worse than no lighthouse at all. The first duty of the instrument is to constrain its own light.
VERDICT
Fail — as registered
Under the outcome rules we registered in advance, this result is a fail: the validation tests for exposure-concentrated, post-2022 onset did not pass. We committed to publishing that outcome with the same prominence a success would have received. This page is that commitment, kept.
V1 · FAIL0 of 3 high-exposure occupation groups show a ≥4-quarter, >2σ decline with onset inside 2022Q4–2024Q4. All three show far longer declines (16–19 quarters) — with onsets in 2021, before the window.
V2 · PASSMean cycle-adjusted entry-hiring deficit of 16.4 points across the high-exposure groups over 2023–2025 — more than three times the registered 5-point threshold. The contraction is large.
F1 · FAILOnly 1 of 5 low-exposure control groups stays quiet in the AI era. The contraction is nearly everywhere — and the only group whose collapse begins inside the registered window is transportation, a control.
F2 · PASSThe within-occupation placebo holds cleanly: 35–44-year-old hiring in the same occupations sits at or above baseline while entry hiring runs ~15% below. Whatever this is, it is specific to the young.
01
What the data shows
SOC
Occupation group
Cohort
σ
Longest run
Run onset
Cycle deficit (pp)
Entry deficit
35–44 deficit
15
Computer & Mathematical
validation
0.028
19q
2021-Q1
17.7
16.7%
−1.0%
27
Arts, Design & Media
validation
0.034
19q
2021-Q1
15.8
15.7%
+0.1%
43
Office & Admin Support
validation
0.034
16q
2021-Q4
15.7
11.4%
−4.8%
47
Construction & Extraction
control
0.013
0
—
−4.2
6.5%
10.3%
49
Installation & Repair
control
0.035
16q
2021-Q4
12.8
12.7%
+0.1%
31
Healthcare Support
control
0.028
15q
2022-Q1
10.9
−10.8%
−24.2%
35
Food Prep & Serving
control
0.078
17q
2021-Q3
21.3
17.6%
−4.0%
53
Transportation
control
0.038
10q
2023-Q2
13.5
6.7%
−7.5%
National series over a balanced 49-state panel (Alaska and Michigan excluded — their QWI publication terminated in 2016 and 2021 respectively; disclosed per Amendment 2). 43 usable quarters, 2015-Q1 → 2025-Q3. Negative deficits denote hiring above baseline. Healthcare support's "decline" is a ratio artifact — both cohorts grew, the older faster — visible here because we publish the components.
02
What we conclude — and what we don't
Confirmed: a large, sustained, youth-specific, cycle-robust contraction in entry-level hiring, beginning around 2021, across most of the U.S. economy. Young people are being hired roughly 15% below their own pre-2020 norm while older workers in the same occupations are not. This is the peril our index measures, and it is emphatically real.
Disproved (at this resolution, with this design): that the contraction is concentrated in AI-exposed occupations with onset following the release of ChatGPT. It is not concentrated, and it did not wait for ChatGPT.
Not resolved: what drives it. Three candidate explanations survive, and they are the subject of our next registered tests: (1) a genuine economy-wide entry-rung contraction predating generative AI; (2) demographic denominator drift — our registered index tracks hire counts, so a post-COVID change in the number of young workers seeking work would produce exactly this pattern; (3) resolution loss — our occupation bridge resolves 22 major groups, and the exposure contrast documented in firm-level microdata may live below that resolution.
02 · ADDENDUM
Where this sits in the literature (added 12 July 2026, after first publication; the results above are unchanged)
A literature sweep completed after this note first published shows the youth-hiring squeeze is documented in parts — by two camps whose claims contradict each other. The AI camp finds early-career decline concentrated in AI-exposed occupations (Brynjolfsson, Chandar & Chen, “Canaries in the Coal Mine,” Stanford Digital Economy Lab, 2025–26); Frank et al. (arXiv:2601.02554) date AI-exposed deterioration earlier, to roughly 2019. The cyclical camp argues the weakness is broad and demand-driven — Rodgers & Kassens, “It’s (Still) the Business Cycle” (Federal Reserve Bank of St. Louis, June 2026), and the Economic Policy Institute’s Class-of-2026 hires-rate series — but from national aggregates, without occupation-level identification.
Our design differs from both: occupation-general coverage including low-exposure trades, food, and transportation; onsets dated by the data (clustered 2021); a within-occupation older-cohort control; county-resolution administrative hires; and pre-registration, which has no analog in this literature. The result above is evidence against exposure-concentration at major-group resolution — consistent with the cyclical camp — while our registered Instrument B will test concentration at the finer occupation resolution where the AI camp’s strongest evidence lives. These camps are currently talking past each other; the instruments registered as round two are built to adjudicate between them, in public.
03
Why we're publishing a failure
Every claim this index will ever make rests on one thing: that its numbers cannot be spun — by counterparties, by markets, or by us. A pre-registered test you only publish when it flatters you is marketing. We registered the thresholds blind, amended them only for disclosed structural reasons before any result existed, ran the tests once, and published what came out. The instrument's first public act is constraining its own claims. That is the governance a settlement-grade index is supposed to have, demonstrated rather than promised.
Round two is being designed now and will be pre-registered before it runs: a rates-based variant with cohort denominators (testing the demographic hypothesis), an occupation-resolved corroborator (testing the resolution hypothesis), a trend-break design that treats the event date as three named hypotheses (GPT-3, Copilot, ChatGPT) rather than one assumption, and a continuous exposure gradient across all 22 occupation groups. Because we have now seen this data, those analyses will be labeled exploratory; confirmatory claims are reserved for the out-of-sample quarters that arrive with each monthly release, beginning August 2026.
04
The pre-registration, verbatim
The registered methodology, cohorts, thresholds, and all three amendments — each adopted before any post-2022 national result existed, each with its stated reason and blind-integrity disclosure.
Registered design (12 July 2026)
Entry cohort: ages 19–24 (Census QWI A02+A03), with 22–24 and 25–34 sub-series. Baseline: 2015–2019 same-quarter average per occupation group — the last window that is both pre-COVID and pre-generative-AI. Scale: fixed floor (1.0 = at/above own baseline; 0.0 = hiring ceased); the scale never rescales. Cycle adjustment: each group's entry ratio divided by the same group's 35–44 ratio — a general downturn nets out; a youth-specific collapse survives. σ estimated per group from 2015–2019 residuals only.
V1 — sustained decline (≥4 consecutive quarters >2σ below baseline), onset within 2022Q4–2024Q4, in the high-exposure cohort. V2 — mean cycle-adjusted deficit ≥5 points, 2023–2025. F1 — ≥4 of 5 low-exposure controls stay quiet. F2 — the 35–44 placebo deficit under half the entry deficit. Outcomes: pass / partial / fail, with fail requiring publication of the negative result.
Amendment 1 — cohort resolution (pre-results)
The occupation bridge resolves 22 SOC major groups; three of the five registered sub-major validation families share major 43 and are not separable. Validation cohort restated as majors {15, 27, 43}; V1 as ≥2 of 3 with major 43 required. Falsification cohort unchanged.
Amendment 2 — balanced panel (pre-results)
Alaska's QWI series terminated in 2016; Michigan's in 2021. A strict all-states rule let a dead series truncate the national window to 2016 (the first run's "fail" was an empty-window artifact; no post-2022 value was computed). National series restated over the 49 states with continuous publication (~97% of employment), identical membership every quarter; exclusions disclosed by name in every release.
Amendment 3 — falsification scope (pre-results)
F1's run search, left unscoped, would count 2020's pandemic collapse against the controls — failing the AI falsification test for COVID reasons. F1 restated to evaluate runs overlapping the AI era (2022Q4 onward); fully-resolved pandemic-era runs are out of scope; the complete run history remains published for every group.