D. Dynamic Maps

These movies show how posting-derived task, skill, and AI signals move across occupational space in the NLx corpus. They are exploratory visualizations of change at work: useful for seeing structure, diffusion, and bridges across SOC groups, but not a substitute for pair-level validation or a causal model of occupational change.

Reading The Movies

Figure DM-1. Static Guide to the Dynamic Occupation Maps

Static guide. LLM-validated task-pair map, no rim

Occupations are arranged around the circle by SOC family, and chords connect occupational groups with shared or moving posting-derived signals. The no-rim view keeps attention on cross-occupation links rather than marginal bars.

Two quantities are doing different jobs here:

• RCA > 1 means a task, skill, or AI code is overrepresented for an occupation relative to the corpus baseline. It is a feature-level screen: this occupation uses the signal more than expected.
• Relatedness is a pair-level score. Two occupations are more related when they share more RCA-screened feature evidence, scaled so broad generic features do not dominate every edge.

In short, RCA says "this occupation lights up for this signal." Relatedness says "these two occupations share a meaningful bundle of overrepresented signals."

From Statements To Maps

The same mapping problem appears across tasks, skills, and emerging technologies. For STEM, for example, the dashboard uses LLM-as-judge review to identify a bundle of task statements that should be treated as STEM-related. Once a bundle of task, skill, or AI statements is identified, postings can show where those statements appear across the occupation structure.

Related-occupation maps use a second step. For each occupation j and task or skill statement s, define an RCA-screened evidence weight:

RCA(j,s) = (mentions of s in j / postings in j) / (mentions of s in all postings / all postings)

e(j,s) = mentions of s in j / postings in j, if RCA(j,s) >= 1 and mentions(j,s) >= 50
       = 0, otherwise

rho(j,j') = sum_s e(j,s) * e(j',s) / max(sum_s e(j,s), sum_s e(j',s))

The denominator makes the measure conservative: a smaller occupation does not become highly related to a larger one merely because its few statements are a subset of the larger occupation's profile. The displayed relatedness movies keep the top partners for each occupation. This frees the evidence from a single canonical O*NET task home and lets postings reveal many-to-many occupational structure found in the data.

LLM-Validated Task Pairs

The LLM-validated task-pair view is closest to the occupation-task-pairs analysis. It is narrower than an all-task movie because the displayed links have passed the review procedure used for the public task-pair evidence. It is still a posting-derived map: a link can reflect genuine task sharing, taxonomy narrowness, noisy occupational coding, or shifts in which employers appear in the corpus.

Related Occupation Maps

Related-occupation movies use shared RCA-screened evidence to place occupation pairs in a common map. The goal is not to say that one occupation becomes another. The goal is to show which occupations repeatedly share distinctive posting signals.

Interpretation Limits

These maps are best read as exploratory measurement tools. Three limitations matter most.

First, occupational coding is imperfect. SOC or OCC6 assignment error can create false bridges or hide real ones, especially when postings are vague, multi-role, or mapped from employer titles into standard occupation codes.

Second, the NLx corpus changes over time in ways that are not fully observed. Employer participation, job-board coverage, posting composition, and occupational mix can all shift. A new chord can reflect changing demand, changing coverage, or both.

Third, broad dictionaries behave differently. ESCO skills and lenient AI-adjacent labels are useful for seeing wide neighborhoods, while LLM-validated tasks and strict AI are cleaner for pair-level interpretation. These movies are retained as exploratory visualizations of change at work rather than as a full edge-audit page.

Evidence Path

LLM-validated task-pair evidence is documented in A New Map of Work and Agreement And Audit Trail. AI-score validation is summarized in Finding Emerging Skills.

Audit Downloads