Methodology

How this dataset was built, from which sources, with which decisions and which known weaknesses. Written so the numbers can be checked rather than taken.

Every figure quoted below is reproduced in data/out/*.json by the run that produced it; the commands to rebuild are in the README.

1. What the dataset is

A record of which tasks belong to which STEM occupations, which standardised activities those tasks are built from, and how exposed each of those is to being performed by AI — joined to employment and wages, and benchmarked against independently produced measures.

The three counts that differ from their headline are not losses to be explained away: 19 occupations are SOC "All Other" residuals with no task data anywhere in O*NET, and 105 tasks postdate the activity crosswalk (§3.2).

2. Sampling frame

The universe is every occupation O*NET classifies as STEM, taken from its own STEM filter rather than from a keyword rule or an SOC-range heuristic. This makes the boundary O*NET's editorial decision, which is citable, rather than ours, which would not be.

The filter has six top-level pages (?t=0 through ?t=5). Two of them split into sub-disciplines rendered as anchored sections within the same page (?t=1#f1 … #f4), not as separate URLs — a fact that has to be handled or two thirds of the category memberships are silently lost.

Membership is many-to-many: the individual category pages hold 302 rows against 287 unique occupations, because 15 occupations appear under more than one STEM discipline. Analyses should use occupation_stem_categories rather than assuming one category per occupation.

3. Sources

Row counts and SHA-256 digests for every file fetched are recorded per run in data/out/manifest.json.

3.1 Why the site and the bulk database

They are not redundant. The live site carries the STEM classification and the current task ratings; the bulk database carries the crosswalk the site does not publish. Both are needed and each is used only for what it alone provides.

3.2 The crosswalk the website does not expose

O*NET OnLine publishes an occupation's tasks and, separately, its detailed work activities. It does not publish which activities a given task maps to — the /link/moreinfo/task/<id> endpoint returns related occupations, not activities. That edge exists only in the bulk tasks_to_dwas.csv, so it is taken from there and joined on the O*NET task id, which the site exposes in each task's popup link.

Consequence: the crosswalk is fixed at release 31.0 while the site is live, so 105 of 5,717 tasks have no activity mapping — they were added to the site after the release was cut. That is a ceiling imposed by O*NET's publication cycle, not a scraping gap, and it is reported as such (98.2% coverage).

4. Extraction

4.1 Web scraping

287 occupation detail reports plus 6 category pages, fetched at 1.5 requests per second across 4 workers with jitter, an identifying User-Agent, exponential backoff honouring Retry-After, and robots.txt respected. The paths used are permitted. Every response is cached on disk by URL digest with its fetch time and content hash, so a rebuild needs no network at all.

Parsing is driven by O*NET's own data-title cell labels and data-text sort values rather than column position, so a re-ordered or re-styled table cannot silently scramble a column.

Two layouts exist. Occupations with incumbent survey ratings render tasks as a table; occupations without render the same section as a <ul> list with no importance or category. Both are parsed. Missing this costs six occupations their entire task list.

Self-check. O*NET prints the row count of each section ("… 16 displayed"). The build compares that figure against the rows it actually parsed and fails if they disagree. This is the check that catches a truncated table, and it is the only reason the list-layout omission was found.

Unrated scores are missing, not zero. Tasks O*NET has not rated display "Not available" behind a data-text="-2" sort sentinel. Parsed naively that sentinel enters the data as a score of −2 and drags every mean down. 264 of 5,717 tasks have no importance score and the field is left empty: 157 New, 102 from the six occupations with no incumbent survey, and 5 Supplemental tasks O*NET left unrated. A validation check rejects any rating outside 0–100.

4.2 Employment

The OEWS national cross-industry file, latest release discovered at runtime from the tables page. Only O_GROUP = "detailed" rows are used. BLS suppression markers (*, **, #, ~) are parsed as missing rather than as numbers.

4.3 Archived releases

Eight releases spanning 2015–2026, taken as db_<release>_text.zip. Availability is verified per release rather than assumed: db_20_0 does not exist (that series begins at 20.1) and an unpublished release is dropped from the span with a warning instead of failing the run.

4.4 Optional measurement files

Three files that are free but large, fetched by fetch-bulk --with-ratings. Every measure below degrades to what it did before if they are absent, so the pipeline runs without them.

Recurrence is the FT distribution collapsed to one number per task. The bands are named, not numbered — "yearly or less" through "hourly or more" — so each is mapped to an approximate rate per year and the logs are averaged. Averaging the band numbers instead would treat the step from monthly to weekly as equal to the step from daily to hourly, which is wrong by a factor of twenty in the quantity being measured. The per-year rates are ours, not O*NET's, and a test asserts the ranking does not depend on them.

Recurrence is measured and reported but deliberately not folded into tractability (§6.2). Against the three terms already on that axis it correlates at r = −0.53, and the components say why:

The most repetitive work in this corpus is the most physically embodied and the least exposed to language models. Emergency medicine physicians, physician assistants and orthodontists score highest because they repeat procedures; anthropologists and nuclear engineers score lowest because their work is rare. That is a result worth stating on its own, and it is also the reason not to average the term in: at r = −0.53 it cancels much of the existing signal, the axis spread falls by a third, and because TRACTABILITY_FLOOR and FRONTIER_K were calibrated against the wider distribution, 24 occupations shift out of "human held" and 24 into "handed off" — which would read as AI having taken over a quarter more of the corpus when nothing about the world had changed. Using it properly requires re-deriving the frontier, which is a judgment, so it is left undone and visible.

Measurement precision. Every descriptor table in this dataset carries n and standard_error and nothing read them: an importance rating from 4 incumbents was treated exactly like one from 225. onet_ratings.rating_precision now reports both. The result is reassuring rather than alarming — the median STEM occupation rests on 27 respondents and none falls below ten — but that was worth establishing rather than assuming.

5. Scoring

5.1 Unit of analysis

Scores are assigned to the 963 distinct activities, not the 5,717 tasks.

This is a substantive choice, not an economy. Rating the activity layer is six times less work, but more importantly it removes an artefact: the same activity cannot receive one score inside nursing and a different score inside engineering. Consistency comes from the structure rather than from the rater's discipline. Scores then propagate outward (§5.4).

5.2 Rubric

Seven dimensions, each 0–100, defined in onet_scraper/score.py and stamped on every row as rubric_version:

Plus a four-way verdict (largely_automatable / augmentable / resistant / human_anchored), a confidence level, and a one-sentence rationale naming the binding constraint.

The dimensions are deliberately separable. An activity can score high on llm_exposure and be human_anchored — drafting a legal opinion is largely language work that a named human must still own. 92 of the 963 activities score ≥70 on both. Collapsing these into a single "automation risk" number is the specific failure the rubric is built to avoid.

5.3 Model and parameters

Claude Opus 5, adaptive thinking, structured output validated against a schema, 12 activities per request, 81 requests, 0 failures, ≈ $8.20. Rubric 2026-09-16.1, fingerprint 49b87187f6f94ad8. Results are cached per activity keyed on the rubric version, so changing the rubric re-runs rather than mixing incomparable scores.

5.4 Propagation

• activity → task: unweighted mean over the activities the task maps to
• task → occupation: mean weighted by O*NET task importance, so an occupation is characterised by what matters in it rather than by its longest tail

Tasks with no scored activity are absent, not zero.

6. Derived measures

6.1 Susceptibility index

Not a mean of the seven dimensions. Two pairs are close to collinear in the scored data:

Averaging all seven would double-count "can a machine do it" and double-count "how much is at stake". Each pair collapses to one factor first, leaving:

• exposure = llm_exposure (physical embodiment is its mirror, not added again)
• anchoring = mean of stakes, interpersonal demand, judgment, physical embodiment

susceptibility = 50 + (exposure − anchoring) / 2, clamped to 0–100.

6.2 Handoff framework

occupation_handoff.csv maps the same dimensions onto the framework in Watson (2026), which scores work on tractability (can AI lead it) and resistance (will it be permitted to), and locates each unit on a six-stage scale of cognitive leadership. Stage is the lower of the capability ceiling and the permission ceiling, not an average — Watson's first four properties decide what AI can do, the last two what it is allowed to do.

Six of his eight properties are covered. Recurrence is now measured (§4.4) but deliberately not folded into tractability; feedback speed and clarity is not in O*NET at all and would need new scoring.

The frontier is calibrated, and the calibration now travels with the distribution. Watson's figure shows the curve's shape and no numbers, so the thresholds were chosen by eye against release 31.0: FRONTIER_K = 53, TRACTABILITY_FLOOR = 50, CROSSING_BAND = 5, plus three literals inside the watch-point rule. That made them quietly fragile. A threshold on an axis is meaningful only relative to the spread of that axis, and the axis is a mean of terms that can be added to or reweighted — so any change to the rubric would silently reclassify the corpus.

That is not hypothetical. Averaging recurrence into tractability narrows the axis by a third (sd 9.95 → 6.44), because recurrence runs against the other terms at r = −0.53. Under the fixed thresholds the result would have been reported as a large move toward automation. Under thresholds that track the axis, it moves the other way:

The stale constants did not merely exaggerate the effect, they inverted its direction: "handed off" rises from 36 to 60 under fixed thresholds and falls to 25 once the calibration adapts. Eighty-four occupations change class the first way, fifty-four the second.

Each constant is therefore expressed as a quantile of the observed distribution — the tractability floor at the 19.1st percentile, the frontier at the 38.6th percentile of the product T × R (which is what the curve is a threshold on), the crossing band at 0.391 standard deviations of resistance. The quantiles were obtained by inverting the original absolutes against release 31.0, so on that release they reproduce the old thresholds to four decimal places and not one classification changes. A test asserts that identity. The reparameterisation is not meant to move today's answer; it is meant to ensure tomorrow's answer moves for a reason.

One detail worth recording: the precision of the crossing-band coefficient is load-bearing. Architects sit at a frontier margin of +4.999362, within 0.0007 of the band edge, so rounding 0.39117889 to 0.391 moves them from "crossing now" to "human held". Their classification was never really determined by the data, and the same is presumably true of any occupation near the curve.

6.3 Employment weighting

O*NET reports at the 8-digit O*NET-SOC level; OEWS reports employment at the 6-digit SOC level. 37 SOC codes here contain more than one O*NET occupation. Attaching OEWS employment to each O*NET row counts the same workers repeatedly — 3.38 million registered nurses five times over.

So susceptibility is aggregated to the SOC level first, then employment is attached exactly once. susceptibility_sd_within_soc reports the spread across each SOC's constituent occupations, so a mean concealing real disagreement is visible. Where BLS publishes only a broad group (29-2011 and 29-2012 both appear as 29-2010) codes fall back to it — and because both then land in the same group, the fallback cannot double-count either.

6.4 Wage deciles

Employment-weighted, and an occupation's workers are split across bucket boundaries rather than assigned whole. Registered nurses alone are 16% of these workers — larger than a decile — so whole assignment produced buckets ranging from 0.4M to 3.8M and the word "decile" was not true. Each decile now holds 2,152,310 workers.

Occupation-level correlations and employment-weighted deciles answer different questions and point different ways here. Both are reported; quoting only one is a choice about which question to answer.

6.5 Transition pathways

A destination must clear three bars: sufficient activity overlap (cosine ≥ 0.12), meaningful relief (≥ 12 susceptibility points), and direction — the shared activities must include the destination's protected work, or the mover carries their exposure with them.

All three thresholds are arbitrary, so the stage sweeps them. The stranded count ranges 105 to 240 of 268 across the sweep: it is sensitive to how much relief is demanded and nearly insensitive to the overlap floor. The clustering of exposure is the finding; the number 161 is not.

6.6 Scenarios

task_fates.csv classifies every task as automated, augmented or unchanged under three assumption sets. These carry no dates and are not forecasts.

The mechanism is deliberately the variable Watson identifies as the real one. Capability is roughly fixed and already measured; what differs between scenarios is willingness to hand over accountability:

Tasks below the augment threshold are unchanged at every setting — for a nurse midwife, 7 of 21 tasks never move regardless of scenario, which is the point the chapter makes.

The fourth category, new tasks, is not modelled. It comes from O*NET's own emerging_tasks file — 121 statements flagged as new or revised work in these occupations — so the arrival of new work is observed rather than assumed.

Employment flows are summed per SOC code, once. The first implementation summed across O*NET occupations and reported 49.3 M workers against a true 21.5 M, because 37 SOC codes hold several O*NET occupations that each carry the same employment figure. The O*NET rows are collapsed to their SOC first; a test covers it.

6.7 Task churn

Task statements diffed across the eight archived releases. A task survives if its id persists or its normalised text matches one in the later release for the same occupation — ids are retired and reissued, and a reworded task is not a new one.

The 2019 SOC revision renumbered codes, so 93 of 268 occupations cannot be compared across the span at all. Those are reported as skipped, never counted as retired tasks.

The confound, stated plainly. O*NET re-surveys occupations on a rolling cycle, so an occupation whose tasks did not change may simply not have been looked at. Splitting on the Date column separates the two: 8.1% turnover among the 153 occupations re-surveyed since 2022, 2.4% among the 115 that were not. The unsplit 5.2% figure is diluted by occupations nobody checked, and no care with the diff fixes that.

6.8 National security matrix

security_matrix.csv places every occupation on three axes at once. The first two ask whether a handoff is attractive; the third asks whether it is reversible, which is the question the other measures in this dataset do not reach.

Unrated tasks still carry weight. O*NET does not rate task importance for every occupation: 159 of 5,612 tasks here carry none, and for six occupations — Cardiologists, Pediatric and Orthopedic Surgeons, Emergency Medical Technicians, Hydrologic Technicians, Health Information Technologists — every task is unrated. The first implementation skipped unrated tasks, which left those six with a zero denominator, returned 0.0 on both axes, and filed surgeons under "nothing is pushing this work toward AI". An unrated task now takes the mean importance of the rated tasks in the same occupation, and an occupation with nothing rated falls back to equal weighting — which says we do not know which of its tasks matters more, not that none of them matters. Radiologists, who had 13 of 30 tasks unrated, move cell as a result, so this is not only about the six.

Efficiency is deliberately not raw exposure. An occupation whose one exposed task carries 3% of its importance mass is not an efficiency opportunity, and averaging exposure across tasks would score it as though it were. Because efficiency reads the per-task verdicts from §6.6, it moves with the scenario: mean 42.6 at modest, 58.9 at substantial, 73.8 at extreme.

Why the risk weights exclude two dimensions we already have. This index does not reuse anchoring, which averages stakes, interpersonal demand, judgment and physical embodiment. Two of those do not belong in a security reading: interpersonal demand is a service-quality property — a call centre scores high and carries no strategic risk — and physical embodiment is a capability limit, not a consequence, so work a robot cannot reach is not thereby high-stakes. Including them produces a general "human-centred work" index; excluding them is what makes this a risk index. That single choice is the main thing to argue with here, and the weights themselves are a calibrated judgment, not a measurement.

Training depth comes from the Required Level of Education distribution (§4.4), averaged as years of schooling and rescaled so 10 years is 0 and post-doctoral training is 100. It yields 191 distinct values across 268 occupations. Job Zone is the fallback for the 44 occupations O*NET has not surveyed for education, and depth_source on every row records which was used. Scarcity is log-scaled employment — the corpus spans four orders of magnitude, and a linear read would call everything except registered nurses scarce. Isolation is the inverse of mean_similarity from the shared-activity network: where many neighbouring occupations share the work, people can cross-train in, and reconstitution is easier.

The cube is split at 50 on each axis into eight named cells:

(occupations / share of the 21,523,100 workers)

A correction, and why it is instructive. An earlier version of this section reported that no occupation anywhere in the corpus was a "reversible gamble", and read that empty cell as a finding: the reconstitution axis had a floor of 44.3, so nothing was genuinely easy to rebuild. That was an artefact of the measurement, not a fact about the workforce. Training depth came from Job Zone, which has five levels of which only three occur here, so it assigned the same depth to an ophthalmic medical technician and a clinical neuropsychologist. Replacing it with the Required Level of Education distribution (§4.4) — the same question at twelve categories instead of five — drops the floor to 33.2 and populates the cell with 21 occupations at the substantial scenario, 30 at the extreme. A coarse input had manufactured a finding, and the finding was about Job Zone.

The absolute split at 50 and the corpus median remain different questions, and the matrix page still offers both: the first asks whether an occupation is dangerous on an interpretable scale, the second which of these occupations is most dangerous relative to the others.

Employment partitions on a per-occupation share, not the SOC figure. §6.3 collapses O*NET occupations to SOC before summing, which is right for a total and wrong for a partition: 18 of the 37 multi-occupation SOCs have members that land in different cells. SOC 19-1029 splits four occupations across three cells, and crediting each cell the SOC's full 55,850 makes the cell shares add to 130%. Each SOC's employment is therefore divided evenly among its constituent occupations. The even split is an assumption — O*NET does not publish how a SOC's workers divide — but every slice then sums exactly to 21,523,100 however the reader cuts it. The alternative, classifying whole SOCs, would discard the occupation-level distinctions the matrix exists to show.

Aggregated to field, weighting each axis by employment:

Field is the leaf STEM discipline where an occupation has one, and the top-level role type for the 31 — teaching, management, sales — that do not. An occupation with several leaf memberships takes the lowest category id, so the assignment is deterministic across runs.

What this is not. O*NET carries no industry, clearance or criticality field, so nothing here identifies an occupation as defence-relevant. The matrix scores properties that make a handoff strategically dangerous, across the whole STEM corpus. Which fields matter is the reader's overlay, not our measurement — and the word "security" in the title describes the question being asked, not a classification this dataset is able to make.

7. Validation

Internal — 11 automated checks, in data/out/validation_report.json. Errors fail the run. They cover: every rostered occupation fetched; no swallowed failures; parsed rows equal the page's own declared count; every task carries an id; crosswalk ids resolve; activities present in the hierarchy; ratings inside 0–100; the wide table neither dropping nor duplicating a task; category membership complete. Table grain is asserted at write time, so a fan-out in a join raises instead of inflating counts.

Convergent. The index is computed only from the numeric dimensions, yet it orders the model's independently produced categorical verdict monotonically: largely_automatable 74.9 > augmentable 65.9 > resistant 43.6 > human_anchored 40.3. Re-checked every run, with a warning if monotonicity breaks.

External. Benchmarked against Eloundou et al. (2023), whose occ_level.csv is keyed to the same 8-digit O*NET-SOC codes — all 268 occupations matched, no crosswalk needed.

Both halves matter. Agreement with human experts at r ≈ 0.85 is the evidence the index measures what it claims. The weaker agreement with α is expected: α excludes tools and this rubric explicitly asks what a model could do given them.

The pre-LLM measures disagree, and should. Frey & Osborne give mathematicians a 4.7% probability of computerisation; this index ranks them third of 268. Those measures scored the routine/manual gradient; LLMs run the other way. A strong positive correlation there would have been the warning sign. These three expectations were written down as "moderate positive" before the test and the data refuted them; the code records that.

8. Limitations

1. The scores are model judgment, not survey data and not human expert judgment. They correlate with human ratings at r = 0.85, which is evidence of validity, not a substitute for it.
2. No reliability estimate exists. Nobody has re-run the scoring and measured agreement, so it is unknown whether the scores are a property of the work or of this model on this day. This is the largest open gap.
3. Chunk-order effects are untested. Activities were rated 12 per request; a score may depend partly on which others shared the request.
4. The unit is the task, not the decision. Watson's framework scores recurring decisions and is explicit that identifying which tasks are decisions is the novel work. That has not been done, so stage numbers are provisional.
5. Activity assignment is analyst-coded, not survey-measured. The network structure partly reflects O*NET's vocabulary choices; the 27% of activities unique to one occupation may be analyst granularity as much as real specialisation.
6. Exposure is not displacement. Nothing here measures what employers will do, what regulation will permit, or how fast anything diffuses.
7. O*NET's own update cadence bounds the churn result (§6.6).
8. STEM only. 287 occupations of O*NET's ~1,000. Nothing here generalises to the rest of the labour market.
9. Postsecondary teaching occupations are near-duplicates. O*NET gives all 25-xxxx occupations one boilerplate profile; Economics and Political Science Teachers score cosine 1.00. They dominate any similarity ranking and are excluded from network defaults.
10. Employment is a national cross-industry snapshot. No geography, no industry detail, no projections.

9. Reproducibility

Every response is cached by URL digest with its fetch time and SHA-256, so a rebuild is deterministic and needs no network:

python -m onet_scraper --offline --with-descriptors build

data/out/manifest.json records the tool version, source URLs, the O*NET release, per-file row counts and content hashes, run settings, and the validation summary. The scroller's network layout is seeded, so the drawing is identical on every run.

A full rebuild from scratch is python -m onet_scraper --refresh followed by the score, report, employment, validate-external, pathways, churn, figures, story and publish stages. 66 tests run in CI on Python 3.10 and 3.13.

One caveat on the recorded stats. http_stats in the manifest reads zero because the final build was served entirely from cache; it counts the requests that build made, not the requests that populated the cache.

10. Citation

If you use the dataset, cite the upstream sources — they did the expensive part.

O*NET data is provided by the U.S. Department of Labor, Employment and Training Administration, under CC BY 4.0. O*NET® is a trademark of USDOL/ETA. Employment and wage data from the U.S. Bureau of Labor Statistics, Occupational Employment and Wage Statistics (May 2025). Benchmark exposure ratings from Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2023), GPTs are GPTs. Handoff framework from Watson, P. (2026), Considering Handoffs of Cognitive Leadership from Humans to AI, Applied Emergence.

The code is MIT licensed. The derived tables carry their upstream terms. If you use the automation scores, cite the rubric version (2026-09-16.1) — they are not comparable across rubric changes.