# The AI Exposure Score Reference, Each Measure Paired With Its Demand Figure

*Take any AI risk figure for a target occupation, name the measure behind it, map it to a SOC code, and pair it with the demand number it cannot replace.*

- Canonical URL: https://www.refolk.ai/candidates/guides/ai-exposure-score-reference
- Pillar: Reading the market
- Format: Reference
- Published: 2026-11-08
- Last reviewed: 2026-11-08
- Reading time: 21 min
- Keywords: is my job at risk from AI, BLS AI exposure categories meaning, AI exposure score by occupation, AI exposure vs job growth projections, how to read AI job risk calculator, which jobs are safe from AI to retrain into

## Key takeaways

- An AI exposure score measures how much of an occupation's work AI could do or is doing, and BLS states its own exposure categories do not predict wages, AI adoption, or the probability that workers will be replaced.
- In Anthropic's data the Computer and Math group shows 94% theoretical coverage but only 33% observed coverage, so a theoretical index tells you nothing about current adoption.
- Anthropic's regression found that each 10 percentage points of observed coverage lowers the BLS growth projection by about 0.6 points, which bends growth toward flat rather than collapse.
- Occupation-level growth can hide an entry-level hiring drop: the Canaries study puts the employment gap for 22-25 year olds in the most exposed occupations at 19% while older age groups show no comparable decline.
- In Refolk's index, US profiles titled Software Engineer or Developer at entry level outnumber those titled Programmer or Computer Programmer by 15.8x, so someone searching 'programmer risk' may be reading the wrong SOC code.
- Before an AI exposure figure moves a target, pair it with the same SOC code's BLS projected change and annual openings, and with the realized early-career quintile if you are entering at entry level.

This reference is for anyone deciding whether to keep aiming at, or retrain into, an occupation when every site quotes a different AI risk number for it. It gives each measure family its own row: what it scores, on which occupation code, what it proves, how it misleads, and the one demand figure you must read beside it before the number moves your target list.

## What does an AI exposure score actually measure?

An AI exposure score measures how much of an occupation's tasks AI could do, or is being used to do. It does not measure how many people employers will hire into that occupation.

Every figure you meet sits at one layer of a stack. The top layer asks whether AI is capable of a task. Lower layers ask whether people actually use AI for it, whether AI does the task or helps a person do it, what official projections expect for headcount, and what has already happened to hiring. The layers disagree by design, and most scary single-number scores come from reading a top-layer figure as if it were a bottom-layer one.

The gap between layers is large. In Anthropic's published usage study, the Computer and Math occupation group shows 94% theoretical coverage but only 33% observed coverage. Coverage here means the share of an occupation's tasks that AI touches. Both numbers are correct; they answer different questions.

#### The five layers an AI risk number can come from

1. **Theoretical capability** - Could AI do this task? Task, ability, and patent-based indices
2. **Observed usage** - Is AI being used for it? Chatbot usage logs mapped to tasks
3. **Automation vs augmentation** - Does AI perform the task, or assist a person doing it?
4. **Projected demand** - BLS projected employment change and annual openings
5. **Realized hiring** - Payroll and Census employment by age and exposure

*An exposure figure from a top layer cannot answer a question that only a bottom layer measures.*

BLS is blunt about this. Its own exposure categories "do not predict wages, productivity, adoption of AI or the probability that workers will be replaced." If the agency that built the classification says it is not a replacement forecast, no site repackaging it can turn it into one.

> **Rule:** Exposure never moves a target alone
>
> Read every exposure figure beside a demand number for the same occupation code. If you cannot find the demand number, the exposure figure does not get a vote.

## The measure families, one row each

There are six families of AI exposure measure in common use: the BLS categories, theoretical indices, observed-usage indices, the ILO generative AI index, realized employment studies, and LLM-scored calculators. Each scores a different unit on a different scale, and only some separate automation from augmentation.

Automation means AI completes the task itself. Augmentation means AI helps a person complete it. The distinction matters because the Canaries study, covered below, finds employment declines where AI automates and growth where it augments.

| Measure | Unit and code | Scale | Automation split |
|---|---|---|---|
| BLS categories | Occupation, SOC (831) | 4 categories | No |
| ILO index | Task to ISCO-08 4-digit | 0-1 mean, Gradients 1-4 | No |
| Copilot applicability | O*NET IWA to occupation | Score, groups about 0.26-0.29 | Goal vs AI action |
| Anthropic observed | Task, O*NET/SOC | % coverage | Automation weighted higher |
| Karpathy LLM score | 342 OOH occupations | 0-10 | Merged |

### BLS four-category relative exposure

BLS combined five external sources into a four-category classification for every detailed occupation it projects, 831 in all. Three sources are theoretical and compare AI capabilities with occupational abilities or tasks. Two are observed and are built partly on usage logs from Claude and Microsoft Copilot. BLS took theoretical and observed median percentile ranks, then used a clustering algorithm to group occupations into four categories. The BLS page names the five sources; read them there rather than from a summary.

- **Proves:** where an occupation ranks against all other projected occupations, blending capability and usage.
- **Misleads when:** read as a forecast. 206 occupations sit in the Very High tier, and a third-party tally (not BLS) counts most of them as still growing.
- **Watch for:** imputation. Of 4,155 occupation-source combinations, 211 were imputed, affecting 75 occupations.
- **Read beside:** BLS projected change and annual openings for the same SOC code, from the same release.

### Task-based theoretical indices

The Eloundou et al. measure scores each task or detailed work activity (DWA, an O*NET description of a specific work action) for whether a large language model could speed it up, then aggregates to the occupation. Its GPT-4 β variant is the one the Canaries study uses to sort occupations.

- **Proves:** potential. What share of the work a capable model could touch.
- **Misleads when:** read as adoption. Theoretical coverage of 94% coexists with 33% observed in the same group.
- **Read beside:** an observed-usage figure, then a demand figure.

### Ability-based and patent-based indices

The Felten AI Occupational Exposure score (AIOE) is ability-based: it matches AI capabilities to the human abilities an occupation requires. The Webb index is patent-based. I did not find a reliable public statement of the Webb index's unit and scale, so verify those against the source paper before quoting it.

- **Proves:** another theoretical angle on potential.
- **Misleads when:** treated as interchangeable with task-based scores. The Eloundou authors report that the AIOE is, for the most part, not correlated with their measures.
- **Read beside:** a task-based or observed figure, then demand.

### Observed-usage indices

Two published studies map real chatbot conversations onto occupational tasks.

- **Copilot applicability score:** built from 200,000 Bing Copilot conversations over 9 months, scored at the O*NET intermediate work activity (IWA, a grouping of related DWAs) level. It scores the user's goal and the AI's action separately, which approximates an assist versus perform split. Computer and Mathematical and Sales both average 0.29; Office and Administrative Support averages 0.26.
- **Anthropic observed exposure:** combines Claude usage with Eloundou task ratings. When Claude completes a task autonomously it counts more heavily; when it assists a human it counts with less weight. Computer programmers show 74.5% observed coverage and data entry keyers 67%.

- **Proves:** what people are actually using AI for, today, on one or two products.
- **Misleads when:** read as economy-wide adoption. Each reflects one product's user base.
- **Read beside:** BLS projected change, and the automation versus augmentation split each one reports.

### ILO generative AI index

The ILO index rates 2,861 tasks, using ratings from 1,640 surveyed workers for 52,558 data points, and scores occupations at the ISCO-08 4-digit level (the international occupation classification). It sorts occupations into four gradients using the mean and standard deviation of their task scores. The mean automation score across all occupations is 0.29 in the update, against 0.30 in the prior version. The ILO estimates one in four workers globally are exposed and 3.3% of global employment falls in Gradient 4, the highest.

- **Proves:** automation potential on an international scale, useful if you are comparing countries.
- **Misleads when:** applied to a US title without translating ISCO to SOC.
- **Read beside:** the national demand projection for the matched occupation.

### Realized early-career employment

The Canaries study uses payroll records to compare employment by age across occupations sorted into exposure quintiles (fifths). It is the only family that measures what has already happened to hiring. Details are in the early-career section below.

- **Proves:** a realized drop in employment for 22-25 year olds in the most exposed occupations.
- **Misleads when:** generalized to all ages, or read as layoffs. The effect runs through reduced hiring, not separations.
- **Read beside:** BLS projected change, so you see both the occupation total and the entry-level slice.

### LLM-scored and commercial calculators

Karpathy's visualizer scores 342 Occupational Outlook Handbook occupations on a 0-10 scale using an LLM (Gemini Flash). The score "considers both direct automation and indirect effects," so it puts automation and augmentation on one axis. The repository calls itself a development tool, not a paper, though it does publish its prompt. Derived visualizers report an unweighted average of 5.3 out of 10 and a workforce-weighted average of 4.9. How free commercial calculators build their scores is not publicly established.

- **Proves:** how one model, given one prompt, rates an occupation.
- **Misleads when:** read as research. A different prompt gives a different number.
- **Read beside:** everything above. It should never be the deciding figure.

### What each family proves, at a glance

| Family | What it proves | Read beside it |
|---|---|---|
| BLS category | Relative rank, blended | Projected change and openings |
| Theoretical index | Potential, not use | Observed usage, then demand |
| Observed usage | Current use on one product | Projected change, auto split |
| ILO index | Global automation potential | National projection via crosswalk |
| Realized early-career | Hiring drop for 22-25s | Occupation-level projection |
| LLM calculator | One prompt's opinion | All of the above |

Two numbers that look different are often the same ordering on a different scale. One cross-index study found Spearman rank correlations of 0.92 between two US-based indices and 0.81 to 0.83 between each of them and the ILO index. Rank correlation measures whether two lists put occupations in the same order, regardless of units. The agreement is not universal, given the weak Eloundou-AIOE relationship, so before concluding two sources disagree, check whether they rank your occupation in the same fifth.

## Which demand number belongs beside each score?

Three demand numbers do work that no exposure score can: BLS projected employment change, BLS annual openings, and realized early-career employment by exposure quintile. Pull the first two for every target, and add the third if you are entering at entry level.

- **Projected employment change:** the percentage change BLS expects in total jobs for the occupation over its ten-year projection window. It tells you the direction of the occupation.
- **Annual openings:** the number of positions BLS expects to come open each year. Openings count jobs that come open as people leave, not just net growth, so they can stay large while total employment declines.
- **Realized early-career employment:** what has already happened to hiring of 22-25 year olds in occupations at that exposure level. It catches what the occupation total hides.

BLS frames its exposure categories as a complement to its projections. Most explainers lead with the exposure score; I lead with demand, because that is the order the agency that publishes both implies.

How much does exposure bend growth? Anthropic ran an occupation-level regression, weighted by employment, of observed coverage against BLS projections.

**-0.6 pts - Drop in BLS projected growth for every 10 points of observed AI coverage**

Anthropic's employment-weighted regression, run on the 2024-34 projections rather than the 2025-35 release.

Do the arithmetic and the picture is calmer than the headlines. At 75% coverage, that coefficient implies roughly 4.5 points less projected growth than an occupation with zero coverage (my derivation, not a published figure). Total US employment is projected to grow only 3.5% over the 2025-35 window, from 170.3 million to 176.2 million. So a heavily exposed occupation is typically looking at flat growth, not collapse. Exposure changes the slope, not the sign, for most occupations.

The sign does flip for some groups. Office and Administrative Support is projected at -4%, or 752,100 fewer jobs, while Computer and Mathematical occupations are projected at +7%. Among individual high-exposure occupations, data scientists are projected at +35% and information security analysts at +21%; financial analysts and computer network architects are both in the Very High tier.

#### Exposure against demand for one target occupation

Horizontal axis runs from Low AI exposure to High AI exposure. Vertical axis runs from Shrinking projected demand to Growing projected demand.

| Quadrant | What it means |
| --- | --- |
| Low exposure, growing | Aim. Confirm openings are large enough to matter |
| High exposure, growing | Keep. Aim at augmented, judgment-heavy tasks |
| Low exposure, shrinking | A demand problem, not an AI problem. Check openings before dropping |
| High exposure, shrinking | Adjust to a senior or tacit-knowledge variant, or drop |

*Exposure tells you how the work will change; projected demand tells you whether the jobs will be there.*

## How do I map my job title to the code a score used?

Find the occupation code the measure actually scored, not the title on your business card. Use O*NET OnLine's search or Crosswalk Search to land on one 6-digit SOC code, then confirm its task list matches what you do.

The code systems in play:

- **SOC:** the US Standard Occupational Classification, a 6-digit code. BLS categories and projections use it.
- **O*NET-SOC:** the SOC code plus a .00 extension, or .01, .02 and so on for more detailed occupations. Task-based and observed indices score at this level. BLS maps its projections onto O*NET codes using employment weights.
- **ISCO-08:** the international classification the ILO index uses, at 4 digits.
- **2010 SOC:** the vintage the Canaries study maps its 7,000+ standardized job titles onto. Check that your code has not changed between vintages.

Some codes cannot be scored at all. In one study using the Current Population Survey, thirteen Census occupation codes had no scores: twelve were "All Other" catch-all categories and one was military, all lacking O*NET task-level data. How often a crosswalk moves a title into a different exposure bucket is not publicly established, so treat any title that spans two codes as two readings, not one.

Titles mislead in practice. In Refolk's index, US profiles titled "Programmer" or "Computer Programmer" turn out to be a mixed bag.

**15.8x - Entry-level US software engineer and developer profiles per programmer-titled profile, in Refolk's index**

420,976 entry-level Software Engineer or Developer profiles against 26,586 Programmer or Computer Programmer profiles at all levels.

| Title in Refolk's index | US profiles | BLS projection for the SOC | Vintage |
|---|---|---|---|
| Software Engineer/Developer, entry level | 420,976 | Developers +17.9% | 2023-33 |
| Programmer/Computer Programmer, all levels | 26,586 | Programmers -6% | 2024-34 |
| Ratio | 15.8x | n/a | n/a |

The computer programmers code is the one with high observed coverage (74.5%) and a shrinking projection. It is held by relatively few people. Someone typing "is my job at risk from AI" with "programmer" in mind is often reading the wrong code. The ratio actually understates the gap, because the developer row counts entry level only. I could not produce a seniority-matched pair, and the projection vintages in that table are mixed; pull current values for both codes from the BLS file before using them.

The title itself is also noisy. In a 25-profile sample of programmer-titled profiles, only 12 read plain "Programmer." Others included "Music Producer," "Programmer, Short Films," "Staff DevOps Engineer," and "Cyber Security Engineer." A music programmer and a DevOps engineer share a word, not an occupation code.

## What does the realized early-career data show?

Realized data shows a hiring drop concentrated on workers aged roughly 22 to 25 in the most AI-exposed occupations, with no comparable decline in older age groups. It is the one family that measures what has happened rather than what could.

| Study | Data | Ages | Gap | Vintage |
|---|---|---|---|---|
| Canaries, first wave | Payroll | 22-25 | 13% (15% revised) | 2025 |
| Canaries, update | Payroll | 22-25 | 19% | June 2026 |
| Tucker | Census | 22-24 | 12.4% | Mid-2025 |
| Facius and Iacono | Nordic registers | 22-25 | No clear gap | Spring 2025 |

What these figures establish:

- **The effect is age-specific.** The Canaries gap for 22-25 year olds widened from 15% at the earlier data vintage to 19% in the update, while other age groups show no comparable decline.
- **It runs through hiring.** Employers cut entry hiring rather than separating existing staff, so occupation totals and pay data barely register it.
- **Experience protects in some occupations.** Employment increased among experienced workers in occupations that rely more on tacit knowledge, meaning know-how learned on the job rather than written down.
- **Automation and augmentation diverge.** Declines show up where AI automates and growth where it augments, which a merged score cannot show.

The Census replication by Tucker works at the industry level, not occupation: in the most exposed fifth of industries, hires of 22-24 year olds fell about 9% right after ChatGPT's release, driving a 12.4% relative fall in early-career employment by mid-2025.

Limits to state plainly. The exact age at which the effect disappears is not publicly established; the Canaries dashboard shows 26-30 year olds with modest gains or slight declines. A Nordic register study found no clear gap. And replications on survey data can be thin: the monthly survey sample for young software developers can be as small as 26 people. If you are over 30 or switching at a senior level, this family matters less to you; if you are entering at entry level, it may matter more than any projection.

## Step-by-step: reading one AI risk figure for one target occupation

Run these seven steps for each occupation on your target list. The first four take about 20 minutes per occupation, and the order puts the code and demand ahead of any judgment.

#### Reading one AI exposure figure

1. **Identify the measure family behind the number** - Find the source footnote and record the source index, the scale, and the vintage. Done when the figure is labelled BLS category, theoretical, observed, ILO, realized, or LLM/editorial.
2. **Map the job title to one code** - Use O*NET OnLine search or Crosswalk Search to find the SOC, O*NET-SOC, or ISCO code. Done when one 6-digit SOC code is chosen and any two-code or "All Other" title is flagged.
3. **Look up the BLS exposure category** - Find the SOC code in the BLS AI exposure file. Done when the category is recorded along with whether any of its inputs were imputed.
4. **Pull the demand pair** - From the same BLS file, record projected change and annual openings for the code. Done when both sit beside the category; openings can be large even when employment declines.
5. **Check realized early-career data** - If you are under 30 or entering at entry level, find the occupation's quintile on the Canaries dashboard. Done when the quintile is recorded.
6. **Check automation versus augmentation** - Look at the Anthropic or Copilot split for the occupation. Done when you can say whether AI mostly performs or mostly assists the work.
7. **Decide keep, adjust, or drop** - Choose to keep the target, adjust it toward a senior or tacit-knowledge variant, or drop it. Done when the decision and the demand figure that drove it are written down.

Write each reading down in the same shape so you can compare occupations side by side.

**Exposure reading record, one occupation**

```
Target title as I would say it:
SOC code chosen:            Second code if title spans two:
"All Other" code (ends in 9)? Y / N     Inputs imputed per BLS? Y / N
Figure that prompted this:   ______   Source:   ______   Scale:   ______   Vintage: ______
Measure family: BLS / theoretical / observed / ILO / realized / LLM-editorial
BLS exposure category:
Projected employment change:          Annual openings:
Early-career quintile (if under 30 or entry level):
Automation or augmentation lean:
Decision: KEEP / ADJUST to __________________ / DROP
The one demand figure that drove the decision:
```

*One record per target occupation. Leave a field blank rather than guess, and fill vintage for every figure.*

An "adjust" is the most common outcome for high-exposure, growing occupations, and it usually means pointing your resume at the senior, judgment-heavy, or tacit-knowledge version of the same role. That shift is mostly a rewrite of how your existing experience is framed. [Refolk](/candidates) writes your resume from your own history and scores how well you fit postings in the adjusted target, which tells you whether the adjustment is a retitle you can make now or a retrain you need first.

## How do AI exposure readings go wrong?

Most bad readings come from one of seven errors: treating a category as a forecast, a theoretical score as adoption, the wrong code as yours, a catch-all code as scored, an occupation total as an entry-level signal, an old figure as current, or an LLM rating as research. Each has a specific check.

> **Watch out:** The most expensive mistake
>
> Dropping a growing occupation because of a Very High category. Data scientists sit among highly exposed occupations and are projected to grow 35%; a category read alone would have taken them off your list.

### 1. A BLS category read as a forecast

- **What it looks like:** "Very High" read as "shrinking."
- **Reality:** a third-party tally, not BLS, puts projected growth for the Very High tier at +2.0% and counts 141 of its 206 occupations as still growing. Treat that tally as indicative and verify your own code.
- **Check:** read projected change and openings beside the category, every time.

### 2. A theoretical index read as adoption

- **What it looks like:** a 94% coverage figure quoted as "AI already does 94% of this job."
- **Reality:** the same group shows 33% observed.
- **Check:** does the source use usage logs? If not, it measures potential only.

### 3. Title-to-SOC mismatch

- **What it looks like:** you search your title, get a frightening score, and never check what code it came from.
- **Reality:** in Refolk's sample, fewer than half of programmer-titled profiles were plain programmers; music and film programmers were in the mix.
- **Check:** compare the O*NET task list for the code against your actual duties before accepting it.

### 4. "All Other" codes

- **What it looks like:** a confident score for a catch-all occupation.
- **Reality:** these codes lack O*NET task data, so scores are imputed or missing.
- **Check:** does the SOC code end in 9, and does BLS flag the occupation as imputed?

### 5. Occupation growth hiding an entry-level drop

- **What it looks like:** "Software is still growing, so I am fine as a new graduate."
- **Reality:** software occupations can show positive projected growth while the 22-25 employment gap in highly exposed occupations is 19%.
- **Check:** the Canaries quintile, if you are entering at entry level.

### 6. Vintage drift

- **What it looks like:** a figure quoted without a date, or two figures from different releases compared as if they match.
- **Reality:** Anthropic's growth correlation used the 2024-34 projections, and the Canaries headline moved from 13% to 19% between versions.
- **Check:** write down the vintage of every figure, and pair figures only within the same release.

### 7. LLM-scored 0-10 calculators

- **What it looks like:** a single 0-10 score with no source.
- **Reality:** the Karpathy score merges direct automation with indirect headcount effects, and a different prompt gives a different number.
- **Check:** look for a named source index, a published prompt or rubric, a task-time weighting, and a code vintage. Missing any of them means editorial.

> Most different AI risk numbers are the same ranking on a different scale, read against the wrong code.

## How do I keep the reading current?

Re-run the procedure whenever a source publishes a new release, and pair figures only within the same release. These measures are revised: BLS publishes new ten-year projections periodically, the Canaries dashboard updates its gap, and observed-usage studies are refreshed as usage logs grow.

Practical habits:

- **Re-check on release, not on headlines.** When BLS issues new projections, re-pull projected change and openings for every code on your list from the BLS file itself, not from coverage of it.
- **Re-check the early-career quintile before you apply.** The realized gap has already moved once between versions; read the dashboard's current figure.
- **Keep the record.** The vintage field in each reading record is what lets you tell a real change from a revised method.
- **Revisit the code when your title changes.** A promotion or lateral move can put you on a different SOC code with a different reading.

Once your target list settles, the remaining work is aiming each application at it. If your list now leans toward augmented or senior variants, Refolk tailors the resume and cover letter to each posting in that adjusted target, so the shift shows up in every application, not only in your notes.

#### Before an exposure figure changes your target list

- [ ] The figure is labelled with its measure family, scale, and vintage.
- [ ] One 6-digit SOC code is chosen, and its O*NET task list matches your actual duties.
- [ ] The code is not an "All Other" code, or you have noted that its score is imputed or missing.
- [ ] The BLS exposure category is recorded with its imputation flag.
- [ ] Projected employment change and annual openings for the same code are recorded from the same release.
- [ ] If you are under 30 or entering at entry level, the Canaries quintile is recorded.
- [ ] You can say whether AI mostly performs or mostly assists the occupation's work.
- [ ] No figure in the record came from an LLM calculator without a published prompt or source index.
- [ ] The keep, adjust, or drop decision names the one demand figure that drove it.

## Frequently asked questions

### Is my job at risk from AI if the BLS puts it in the Very High exposure category?

Not necessarily. The BLS category ranks relative exposure and BLS says it does not predict replacement, wages, or adoption. Data scientists and information security analysts sit among highly exposed occupations yet are projected to grow 35% and 21% under the 2025-35 projections. Read the category beside projected change and annual openings for the same SOC code, and check the early-career quintile if you are entering at entry level.

### What do the BLS AI exposure categories mean?

They are four groups of relative AI exposure covering 831 occupations. BLS took percentile ranks from three theoretical measures and two observed-usage measures, then used a clustering algorithm to sort occupations into four categories. Some inputs were imputed: 211 of 4,155 occupation-source combinations, affecting 75 occupations. The category tells you where an occupation ranks against others, not whether it will shrink.

### Why do different AI job risk calculators give my job different numbers?

Mostly because they use different scales and different questions. One study found rank correlations of 0.81 to 0.92 between major indices, so many disagreements are the same ordering expressed as 0-1, 0-10, percentages, or categories. Real disagreement comes from theoretical versus observed measures, from merging automation with augmentation, and from scoring a different occupation code than the one you hold.

### Which jobs are safe from AI to retrain into?

No exposure index can answer that alone, because exposure does not measure hiring demand. Shortlist occupations with growing BLS projected change and large annual openings first, then use exposure to check whether AI mostly augments the work. Canaries found employment growth for experienced workers in occupations relying on tacit knowledge, which points toward roles where judgment and context matter more than routine output.

### How do I read an AI job risk calculator that does not cite a source?

Treat it as editorial until it shows four things: a named source index, a published prompt or rubric, a task-time weighting, and a code vintage. How free commercial calculators build their scores is not publicly established. If the score comes from an LLM rating occupations on a 0-10 scale, a different prompt would give a different number, so do not let it move a target.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/candidates/guides/ai-exposure-score-reference*
