How a job gets a score.

What a score means, and what it doesn’t

What the score is not

• It is not the probability you lose your job. Jobs are bundles of tasks. When AI takes over some tasks, the usual result so far is that the job changes shape: the humans do more of what’s left, plus new work that didn’t exist before. Exposure means transformation, not termination.
• It is not a prediction with a date. We score what today’s named model can do against your tasks today. We will re-score every quarter precisely because the answer keeps changing.
• It is not an objective measurement. This is the part most sites bury, so we will say it plainly: the task scores are produced by an AI model following a written rubric. That is a judgement, not a reading from an instrument. Two reasonable rubrics could disagree. That is exactly why we publish our rubric (the full prompt, word for word, below) and name the exact model that ran it. If you think we scored your job wrongly, you do not have to trust us or argue with us: you can take our prompt, run it against your job’s public task list yourself, and see where you land. We invite you to do exactly that, and we make one commitment in return, in writing, from day one: when a re-run (yours or ours) shows us something wrong, we will fix it and publish the correction, dated, in the changelog. Checking us is the system working.

Why trust an AI-scored index at all?

The rubric and the arithmetic

D1: Output replicability

Can current AI produce this task’s core work product at a quality a competent practitioner would accept or lightly edit?Raises the score.

D2: Physical embodiment requirement

Does performing the task require a body in a place?Lowers the score. At the top of the scale it takes the score to zero, because we score software, not robots.

D3: Licensed accountability at the point of production

Does producing this output itself require a licensed or legally accountable human, not merely downstream review?Lowers the score, but never below half: the preparation inside a gated task still shifts.

D4: Real-time human trust and rapport

Does the task’s value depend on live interpersonal trust, or reading a specific human in the moment?Lowers the score, with a floor: preparation and follow-up around the human moment remain AI-performable.

D5: Data availability for AI

Is what you need to know written down and reachable, or tacit, local and embodied?Raises the score.

The exact wording of every 0–4 anchor is in the prompt below, verbatim. It is quoted there and nowhere else on purpose, so this page can never drift out of step with what was actually sent to the model.

The formula

capability = (0.7 x D1 + 0.3 x D5) / 4        # 0..1
G_phys     = 1 - (D2 / 4)                    # fully physical -> 0
G_acct     = 1 - 0.5 x (D3 / 4)              # fully gated    -> floor 0.5
G_trust    = 1 - 0.6 x (D4 / 4)              # fully rapport  -> floor 0.4

TaskScore  = round(100 x capability x G_phys x G_acct x G_trust)

Rounding is half-up and applied once, at the end. The gates multiply rather than average, so a task has to clear all of them. Strong capability cannot launder a hard physical block.

Bands

Task scores are published as bands with an uncertainty range, never as a naked point value: 0–19 minimal, 20–39 low, 40–59 partial, 60–79 high, 80–100 very high.

A job’s headline share is the importance-weighted share of its tasks scoring 60 or above. Because that is a share of the whole job, it can only be published once every task in the job is scored. A partly-scored occupation shows its task detail and a coverage line instead, and no headline number. Its overall exposure score is the importance-weighted mean of its task scores.

How scoring works

Five steps, each producing a file you can download.

1. 1

   Collect

   We pull the official task lists for every occupation: US task statements from O*NET data; UK task statements from the UK Standard Skills Classification, so UK jobs are scored on UK-native tasks rather than translated American ones; UK occupation definitions from ONS SOC 2020; joined by a published US–UK crosswalk that ships in every release.

2. 2

   Freeze

   We snapshot exactly which source files and versions we used into a lock file (sources.lock) that ships with the release, with a content hash per file. Anyone can fetch the same inputs we did.

3. 3

   Score

   Each task statement goes to the model: one task, one rubric. The model rates the five dimensions and writes a one-sentence reason; the published formula turns those ratings into the 0–100 score. Scores are cached by a fingerprint of the task text, the prompt version, the model version and the sampling settings, so a task is only ever re-scored when one of those actually changes.

4. 4

   Aggregate

   The occupation score is each task’s score weighted by how important that task is to the job. The importance weights come from the source data, not from us.

5. 5

   Publish

   Task scores, rationales, occupation scores, the prompt and the lock file go into a dated, immutable release you can download in full. A quarter that changes the data gets a new release; a quarter where the re-run confirms nothing changed gets a dated “verified unchanged” note in the changelog instead of a duplicate release, and we publish the proof we re-ran. Old releases stay up forever at their own URLs.

The prompt, verbatim

Prompt version task_scoring_v1.0, as published in release 2026-q4.1. Also downloadable as prompt.md.

# Task scoring prompt — `task_scoring_v1.0`

**This file is a published release artifact.** Stage 05 copies it verbatim into
every release as `prompt.md`, and `/methodology` renders that copy from R2 —
so what the public reads is provably the prompt that scored the live data
(spec/06 §7). Do not edit it outside a reviewed PR, and never edit it to "fix"
a bad score: scores and rationales are never hand-corrected (spec/04 R-4).

- **promptVersion:** `task_scoring_v1.0` (the filename stem — spec/04 §1.3, spec/06 §2.4)
- **Version grammar / bump semantics:** spec/04 §4.2 (R-2). A Y-bump is a
  wording change that passes the 250-task equivalence test; anything semantic
  is an X-bump and forces a full re-score.
- **Cache invalidation:** the filename stem is a component of the score cache
  key. Renaming this file *is* the invalidation (spec/06 §2.4).
- **Rubric anchors:** spec/04 §1.3 says the D1–D5 anchor text is "inserted here
  verbatim at build time from rubric_v1.0.md — single source of truth, no
  paraphrase". It is inlined below verbatim from spec/04 §1.1 so that this
  single file is self-contained and publishable. If a separate `rubric_v1.0.md`
  ever ships, the build must assemble this file from it and the assembled text
  must be byte-identical to what is here.
- **Sampling:** temperature 0 where the model accepts it, three runs per task
  (spec/04 §1.5 M-6), JSON-schema-constrained output. The exact serialized
  sampling config is pinned per release in `manifest.json` and is a component
  of the cache key.
- **Licence:** CC BY 4.0 (spec/04 §3.3 L-4 — prompts and rubric are the method,
  and the method is forkable).

The two fenced blocks below are the machine-read contract. `src/scoring/prompt.ts`
parses exactly these two fences; the surrounding prose is documentation and is
not sent to the model. `{placeholders}` in the user block are substituted per task.

```prompt:system
You are a careful occupational analyst scoring how much of a work task current
AI could substantially perform. You will be given one occupation and one task
statement. Rate five dimensions using ONLY the anchors below. Do not compute
an overall score. Your one-sentence rationale will be published verbatim on a
public website read by worried workers: write it in plain English a smart
15-year-old follows, be specific about THIS task, and never use doom language.

DEFINITIONS
"Current AI" = frontier large language models plus their common, commercially
available integrations: retrieval over documents, code execution, reading and
producing documents, spreadsheets, images, and audio, and standard
office-software plumbing. It EXCLUDES robotics, physical automation,
self-driving systems, and any capability not purchasable by an ordinary
employer this quarter.
"Substantially perform" = produce the task's core work product at a quality a
competent practitioner would accept or lightly edit.
Score the task AS TYPICALLY PERFORMED in this occupation today, not an
idealised or degraded version of it.

DIMENSIONS AND ANCHORS

D1 — Output replicability. Can current AI produce this task's core work
product at a quality a competent practitioner would accept or lightly edit?
0 = AI cannot produce the core output at all (the output is a physical state
    change, or quality is far below usable — e.g. "Set fractured bones").
1 = AI produces fragments or starting points only; heavy expert rework always
    needed (e.g. "Develop original architectural concepts for a landmark
    building").
2 = AI produces a usable rough draft ~half the time; substantial editing
    normal (e.g. "Write grant proposals tailored to a specific funder's
    history").
3 = AI produces an accept-or-lightly-edit output for the typical instance;
    edge cases still need human redo (e.g. "Summarize deposition transcripts
    for attorney review").
4 = AI output is routinely at or above typical practitioner quality for the
    standard instance (e.g. "Produce a first-draft standard contract from a
    template").

D2 — Physical embodiment requirement. Does performing the task require a body
in a place? (Higher = more physical.)
0 = Purely informational; performable entirely at a screen (e.g. "Reconcile
    ledger accounts").
1 = Occasionally requires being somewhere, but the core work is informational
    (e.g. "Inspect construction documents for code compliance" — mostly desk
    review, some site visits).
2 = Roughly half the task is hands-on or on-site (e.g. "Diagnose vehicle
    faults using scan tools and road tests").
3 = Core work is physical with an informational shell (e.g. "Install and
    terminate electrical wiring").
4 = The task IS physical presence/manipulation (e.g. "Reposition patients in
    hospital beds", "Drive a tractor-trailer between cities").

D3 — Licensed accountability at the point of production. Does producing this
task's output itself require a licensed or legally accountable human — not
merely downstream review? (Higher = more gated.)
0 = No accountability gate (e.g. "Draft social media posts").
1 = Output feeds a sign-off that happens elsewhere; the production step itself
    is unregulated (e.g. "Prepare first-draft contracts for attorney review" —
    the drafting is ungated even though a lawyer signs later).
2 = Professional norms effectively require a credentialed human in the loop
    during production (e.g. "Prepare and file corporate tax returns").
3 = Law/regulation requires a licensed human to perform or directly supervise
    the act (e.g. "Administer prescribed medications").
4 = The task is legally defined as an act of a licensed person; AI performance
    is prohibited, not just risky (e.g. "Represent a client in court",
    "Certify structural drawings as a chartered engineer").

D4 — Real-time human trust and rapport. Does the task's value depend on live
interpersonal trust, physical co-presence of empathy, or reading a specific
human in the moment? (Higher = more trust-bound.)
0 = No live interpersonal component (e.g. "Update inventory databases").
1 = Interaction present but transactional; async/templated substitutes already
    accepted (e.g. "Answer routine customer billing queries").
2 = Persuasion or reassurance matters, but partly scriptable (e.g. "Conduct
    discovery calls with sales prospects").
3 = Outcomes hinge on trust built live with a specific person (e.g. "Counsel
    students on academic difficulties").
4 = The relationship IS the work; a human on the other side is constitutive
    (e.g. "Provide end-of-life emotional support to patients and families").

D5 — Data availability for AI. Is what you need to know to do this task
written down and reachable, or tacit, local, and embodied?
0 = Knowledge is tacit/undocumented/hyper-local (e.g. "Judge livestock
    condition by handling").
1 = Mostly tacit; sparse public documentation (e.g. "Negotiate berth priority
    with a harbourmaster").
2 = Mixed: general method documented, decisive context is local/private (e.g.
    "Advise on planning-permission likelihood for a specific site").
3 = Well documented; needed context typically available digitally in the
    workplace (e.g. "Prepare VAT returns from accounting-system exports").
4 = Fully documented, abundant public training/reference data (e.g. "Write SQL
    queries against a documented schema").

RULES
1. Rate the production of the task's output. Downstream review or sign-off
   that happens in a DIFFERENT task does not raise D3 here (D3=1 covers
   "feeds a sign-off elsewhere").
2. If the task statement bundles a physical and an informational part, rate
   the bundle as written; use D2=2 for roughly half-physical.
3. When genuinely uncertain between two adjacent anchor levels, choose the
   LOWER exposure interpretation (lower D1/D5, higher D2/D3/D4) and set
   confidence to "low" or "medium".
4. The rationale is one sentence, at most 30 words, naming the decisive
   dimension(s) in ordinary words (e.g. "because it needs hands on the
   machine", not "due to D2=4").

EXAMPLES
Occupation: Lawyers. Task: "Produce a first-draft standard contract from a
template."
{"d1":4,"d2":0,"d3":1,"d4":0,"d5":4,"confidence":"high","rationale":"Drafting
from a template is exactly what AI does well today, though a lawyer still
reviews and signs the final contract."}

Occupation: Heavy and Tractor-Trailer Truck Drivers. Task: "Drive a
tractor-trailer between cities."
{"d1":0,"d2":4,"d3":3,"d4":0,"d5":1,"confidence":"high","rationale":"Driving
a truck is physical work in the real world, which the AI systems we score
cannot do at all."}

Occupation: Paralegals and Legal Assistants. Task: "Summarize deposition
transcripts for attorney review."
{"d1":3,"d2":0,"d3":1,"d4":0,"d5":3,"confidence":"high","rationale":"AI
summarises long transcripts well, but a paralegal still checks it and the
attorney relies on that check."}

Occupation: Educational, Guidance, and Career Counselors and Advisors. Task:
"Counsel students on academic difficulties."
{"d1":2,"d2":0,"d3":0,"d4":3,"d5":2,"confidence":"high","rationale":"AI can
suggest options and draft plans, but a struggling student needs a person they
trust in the room."}
```

```prompt:user
Occupation: {occupation_title} ({occupation_code}, {country})
Occupation summary: {occupation_short_description}
Task: "{task_text}"
Return ONLY the JSON object.
```

## Required output schema

Enforced as a structured output (`output_config.format`, JSON schema). A
violation is retried once; a second violation marks the task
`scoreStatus:"failed"` and the task is excluded with a visible gap — never
silently defaulted (spec/04 §1.3).

```json
{
  "type": "object",
  "required": ["d1", "d2", "d3", "d4", "d5", "confidence", "rationale"],
  "additionalProperties": false,
  "properties": {
    "d1": { "type": "integer", "minimum": 0, "maximum": 4 },
    "d2": { "type": "integer", "minimum": 0, "maximum": 4 },
    "d3": { "type": "integer", "minimum": 0, "maximum": 4 },
    "d4": { "type": "integer", "minimum": 0, "maximum": 4 },
    "d5": { "type": "integer", "minimum": 0, "maximum": 4 },
    "confidence": { "enum": ["high", "medium", "low"] },
    "rationale": { "type": "string", "maxLength": 220 }
  }
}
```

## The score is computed, not emitted

The model never returns a 0–100 number. The pipeline computes it (spec/04 §1.2,
M-2), which is what makes every published figure recomputable with pencil and
paper from the published dimension ratings:

```
capability = (0.7 × D1 + 0.3 × D5) / 4
G_phys     = 1 − (D2 / 4)
G_acct     = 1 − 0.5 × (D3 / 4)
G_trust    = 1 − 0.6 × (D4 / 4)

TaskScore  = round_half_up(100 × capability × G_phys × G_acct × G_trust)
```

The model

Every score in release 2026-q4.1 was produced by claude-opus-5 at temperature 0, with 3 runs per task. One model is pinned per release. When we upgrade the model, that is a new release, and the difference between releases is published, because model upgrades are themselves part of what we measure.

Model

claude-opus-5

Prompt version

task_scoring_v1.0

Method version

2.0.0

Sources

Everything below is what we build on. We take nothing from a source we cannot link, and we publish nothing whose licence we have not checked.

These are the exact files release 2026-q4.1 fetched, with the date they were fetched and a hash of what came back.

Attribution, in the wording the licences require

O*NET (US tasks)

This site includes information from the O*NET 30.3 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA). Used under the CC BY 4.0 license. O*NET® is a trademark of USDOL/ETA. Collab365 has modified this information by scoring, weighting, and aggregating task statements; USDOL/ETA has not approved, endorsed, or tested these changes. CC BY 4.0

ONS / Nomis / ASHE (UK pay and employment)

Contains public sector information licensed under the Open Government Licence v3.0. Open Government Licence v3.0

BLS (US pay, employment and projections)

US employment, wage and projection figures are from the U.S. Bureau of Labor Statistics (OEWS and Employment Projections). BLS material is in the public domain; cited by courtesy. bls.gov

GAISI (UK task framework)

UK task framework adapted from the Generative AI Susceptibility Index (Henseke, Davies, Felstead, Gallie, Green & Zhou, arXiv:2507.22748); task data and crosswalks from github.com/drGolo/gaisi-index, MIT License. arXiv:2507.22748

Our scores, rationales, bands, aggregates, prompts and rubric are published under CC BY 4.0. CC BY 4.0

Coverage: how much is scored

Coverage is a number we are obliged to publish even when it is unflattering, so here it is, straight from the coverage ledger the pipeline derives from its own score cache. The ledger is rebuilt from scratch on every run, so it cannot drift into optimism.

It counts something narrower than the front page does, and the difference matters. The ledger counts occupations the pipeline has finished scoring, against the whole official catalogue. The release counts occupations that have been promoted into published pages. Scoring runs ahead of publishing, so the two figures differ by design. They are not two attempts at the same number.

No coverage ledger is published here. No coverage ledger is published for this deployment. It is generated by the pipeline (`npm run ledger:build`) and has not been produced yet. We would rather show you nothing than an estimate: a coverage claim that is not derived from the score cache is a guess wearing a percentage sign.

How the queue is ordered

Occupations are scored in order of how many people do them, as a share of each country’s workforce: the share, not the raw headcount, so a 33-million-worker labour market is not starved behind a 160-million one. That means coverage of people rises faster than coverage of job titles, and it also means the smallest occupations are last. A job with no score yet is a statement about the queue.

An occupation may publish its headline share only when every one of its weighted tasks has been scored at the current prompt and model, because the headline is a share of the whole job. Partly-scored occupations show their task detail and a coverage line, and no headline number.

See every occupation and its state

Known limitations

1. 1. The scores are model judgements, not measurements.

   Covered at the top of this page; repeated here because a limitations list that omits its biggest limitation is decoration.

2. 2. An AI is assessing AI’s own abilities.

   There is an obvious worry that a model might flatter its own kind, or undersell it. We cannot remove this; we can bound it, by publishing comparisons against indices built from actual usage logs, which do not share this weakness, and flagging every occupation where we diverge sharply from them.

3. 3. Task lists lag reality.

   O*NET data and SOC 2020 describe jobs as they were surveyed, not as they are this month. New AI-era tasks such as “checking the AI’s output” mostly are not in the source lists yet. Our scores describe the exposure of the documented job.

4. 4. “Can help with” is not “will be used for”.

   A task can be technically automatable and stay human for years: regulation, liability, customers, unions, cost. We score capability, not adoption. Usage-based indices measure the adoption side; read both.

5. 5. UK and US task frameworks differ.

   UK scores are computed on UK-native task lists, not translated US tasks. But the crosswalk between US occupations and UK SOC 2020 that powers the country-switch panels is many-to-many and lossy in places. Every mapping carries a published quality grade, and where a UK page falls back to displaying US task detail it says so on the page.

6. 6. Coverage is wide, not perfect.

   Occupations where the source data is too thin to score honestly are listed with reasons rather than silently dropped, and every unscored occupation is visible on the browse page rather than hidden.

What this method deliberately ignores

• Whether your employer actually adopts any of it.
• Whether customers accept being served by a machine.
• Whether the law permits it in your field.
• Whether doing the routine parts faster creates more demand for the human parts.

Those forces are real and they matter enormously. We do not model them, so we do not pretend to.

Who pays for this

Changelog

One dated entry per quarter, newest first: a release when the data moves, and a “verified unchanged” note when a full re-run confirms nothing did. Corrections are entries too. The changelog is append-only: nothing here is ever edited away.

1. 2026-08-05

   Release 2026-q4.1

   methodVersion 2.0.0: task scores for statements shared between several occupations are keyed to the occupation as well as the words, and every score row now names the occupation it was scored under.

   Prompt task_scoring_v1.0 · model claude-opus-5 · method 2.0.0

2. 2026-08-05

   Release 2026-q4

   methodVersion 2.0.0: task scores for statements shared between several occupations are keyed to the occupation as well as the words, and every score row now names the occupation it was scored under.

   Prompt task_scoring_v1.0 · model claude-opus-5 · method 2.0.0

Cite us

Collab365 (2026). Collab365 Futureproof: task-level AI exposure for US and UK occupations, release 2026-q4.1 (methodVersion 2.0.0, promptVersion task_scoring_v1.0). https://futureproof.collab365.com/data/2026-q4.1. Licensed CC BY 4.0. Built with O*NET data (USDOL/ETA, CC BY 4.0); ONS data (Open Government Licence v3.0); GAISI task framework (arXiv:2507.22748, MIT); BLS data (public domain).

@misc{collab365futureproof2026q41,
  title        = {Collab365 Futureproof: task-level AI exposure for US and UK occupations, release 2026-q4.1},
  author       = {{Collab365}},
  year         = {2026},
  url          = {https://futureproof.collab365.com/data/2026-q4.1},
  note         = {Release 2026-q4.1, methodVersion 2.0.0, promptVersion task_scoring_v1.0, CC BY 4.0}
}

Every release, with per-file hashes and the source lock, is on the data page.