How we score a resume

The existing exposure literature settles on occupations as its unit of analysis. Frey & Osborne (2017) rank 702 occupations by automatability. Eloundou, Manning, Mishkin & Rock (2023) compute three exposure measures (α, β, ζ) per SOC by aggregating 19,265 task-level judgments. Pew (Kochhar 2023) bins 41 O*Net work activities into three exposure levels and tallies workers by occupation. Massenkoff & McCrory (2026), the Anthropic Economic Index, publish an empirical R_o per SOC from actual Claude traffic. None of these give within-occupation resolution: two software developers at the same firm with different skill mixes collapse to one benchmark.

Korinek & Juelfs (2022) already argue, theoretically, that the right unit of analysis for the labor-saving-progress problem is the individual agent. Their Propositions 1 and 2 partition a plane indexed by worker productivity wⁱ and non-labor income Tⁱ into regions labeled Perish, Work, and Don't Work. Korinek & Stiglitz (2021) formalize inequality channels at the agent level via innovator rents and pecuniary externalities. Korinek & Suh (2024) write the labor market as a race between automation speed and capital accumulation across “atomistic tasks of varying complexity.” What the literature lacks is an operational individual-level index that can plug into those frameworks. LevelUp provides one.

We organise the theoretical foundation as three strands of ECON 4444 material and position every modelling choice inside one of them.

All six variables are defined once here and used consistently across the rest of the document.

Every analysis resolves to three values, each answering a different question. Job Fitness lives on [0, 100] (Dreyfus rubric scale); ρ_i and R_o both live on [0, 1] (adoption-rate scale) and are displayed as percentages so the Wedge can be read directly in percentage points.

Job Fitness uses the same Dreyfus 50-on-every-skill normalization at the skill level: an average worker maps to Fitness = 50. ρ_i, by contrast, is calibrated to R_o directly via the anchor property above; it lives on R_o's natural [0, 1] scale and is displayed as a percentage.

Three limiting cases are worth noting. (i) If T_it = 50 on every task, ρ_i = R_o (anchor property above). (ii) If every r̃_t in the occupation is equal, ρ_i reduces to that common penetration value regardless of the worker's skill mix. (iii) If the importance vector is a one-hot indicator on a single task, ρ_i collapses to r̃_t for that task.

The project's named output is the signed distance between individual exposure and the occupation benchmark. For any worker i in occupation o:

A Wedge of zero means the worker sits exactly at the occupation benchmark — and by the anchor property, this is mechanically true for the Dreyfus-50 average worker. A positive Wedge means her task-weighted skill mix is more AI-penetrated than the typical holder of her job title: the tasks she is strong on overlap more heavily with where Claude is already being used. A negative Wedge means the opposite: her portfolio is concentrated in tasks Claude has barely touched. Senior tradespeople are the archetypal negative-Wedge worker; junior knowledge workers are the archetypal positive-Wedge worker.

Under the task-based framing of Autor, Levy & Murnane (2003) and Acemoglu & Restrepo (2018), and the individual-agent heterogeneity Korinek & Juelfs (2022) insist on, the Wedge is the residual that occupation-level exposure indices cannot explain. Eloundou, Pew, and Anthropic all publish a single number per SOC. Two software developers can sit at W_i = +30 pp and W_i = −10 pp under the same SOC; that intra-occupation variance is the thing this project exists to surface. Live agent confirmed example: Korinek (Economists, SOC 19-3011) lands at F_i = 76.2, ρ_i = 0.157, Wedge = −8.5 pp.

The Wedge is diagnostic, not predictive. It does not forecast adoption timing or displacement probability. Under the CES framing Korinek & Stiglitz (Windfall Clause, 2024 revision) use, whether AI ends up labor-saving (σ > 1) or labor-complementary (σ < 1) is a design choice of the deployer. The Wedge only says which skill portfolios sit in the first path of exposure, conditional on Claude adoption continuing along observed patterns.

Per-skill grading uses a deterministic point-for-point rubric adapted from Dreyfus & Dreyfus (1986). Four components, each anchored in the industrial/organizational-psychology literature. Each resume bullet is classified into a band, and the arithmetic is reproducible from the band outputs.

Total. Add the four components and cap at 100. The per-skill breakdown is exposed in the dashboard as E · A · Ed · C so every number can be audited directly.

Each of the 35 O*Net skills is assigned a numeric anchor φ_s ∈ [0.05, 0.85] drawn from three sources, in priority order.

1. Eloundou et al. (2024), Table 5. Their OLS regression of occupation-level β exposure on normalized O*Net skill importance gives empirical coefficients for 11 of the 35 skills. These become our primary anchor for those skills.
2. Kochhar / Pew (2023). Pew groups the 35 O*Net skills into five families (Social, Fundamental, Analytical, Managerial, Mechanical) and separately bins the 41 O*Net work activities into high / medium / low exposure. We use the family groupings as a prior for the 24 skills Eloundou's regression does not cover, translating Pew's family-level ordering of AI exposure (Analytical > Fundamental > Managerial > Social > Mechanical) into our six-tier scale.
3. Massenkoff & McCrory (2026). Anthropic Economic Index empirical R_o values validate the floor (Appendix D's zero-coverage occupations — motorcycle mechanics, agricultural-equipment operators, cement masons, 31 others) and the ceiling (Computer Programmers, SOC 15-1251, at R_o = 0.745).

One attribution detail. Pew does not publish skill-level tiers. Pew publishes activity-level tiers and skill-family groupings. The six-tier skill system is LevelUp's synthesis, built on Eloundou's empirical coefficients, Pew's family structure, and Anthropic's observed usage.

Eloundou Table 5 β coefficients

Table 5 of Eloundou et al. (2024). Sign convention: positive β means the skill's importance correlates with higher occupation-level exposure. *** p < 0.001, ** p < 0.01.

Six-tier system

R_o is not a forecast, a risk score, or an exposure index in Eloundou's sense. It is an empirical fraction: how much of the representative O*Net task bundle for occupation o is observed in Claude traffic, weighted by each task's share of the occupation's work time. Massenkoff & McCrory (2026) — the Anthropic Economic Index — construct it in three stages.

Three-stage construction

Dataset specifics

Headline numbers we cite

Six numbered steps. The LLM appears only at parse + Dreyfus evidence classification; everything else is deterministic table lookups and arithmetic.

1. Parse the resume + pick the target SOC. Claude Haiku 4.5 extracts structured JSON from the PDF (name, skills, years of experience, education, certifications, job titles) and chooses the best-fitting SOC code from the full O*Net 27.2 taxonomy. The user can override.
2. O*Net 27.2 lookup. For the chosen SOC, we hit api-v2.onetcenter.org for the full 35-skill profile with importance ratings I_os. Covers all of the roughly 900 occupations O*Net publishes. Cached for 1 hour.
3. Dreyfus rubric → S_is. For every skill, the agent classifies resume evidence into the four-component rubric: Experience 35 pts (Hunter & Hunter 1984), Achievement 45 pts (Schmidt & Hunter 1998), Education 20 pts decayed (Roth et al. 1996), Certifications +10 pts. No evidence yields zeros (Level 1 Novice).
4. O*Net crosswalk → T_it. For each occupation task t, look up the DWAs covering it (Tasks → DWAs file), take the 5-segment DWA prefix, then map to skills via the Skills to Work Activities crosswalk. T_it is the unweighted mean of S_is over the skills mapped to task t. If crosswalk coverage falls below 5 matched tasks or 25% of the occupation's task list, the pipeline falls back to the skill-weighted E_i formulation (see §4).
5. Weighted aggregation → ρ_i. Pull task penetration r̃_t from task_penetration.csv (Anthropic Economic Index, keyed by task description text) and task importance w_t (O*Net IM rating, 1–5). Compute ρ_i = Σ_t w_t r̃_t T_it / Σ_t w_t T_it.
6. Wedge = ρ_i − R_o. Look up R_o for the SOC in Anthropic's occupation-level CSV; the Wedge is the signed difference, displayed in percentage points. The dashboard renders the three numbers, a per-skill table with E · A · Ed · C sub-scores, a radar of user vs. importance vs. task-penetration, and actionable recommendations.

• Resumes are lossy signals. Achievement-band classification is the highest-leverage node in the pipeline and the most subjective. Schmidt & Hunter (1998) report inter-rater ρ ≈ 0.70 for structured protocols; Claude is probably worse, especially at the exceptional-band boundary.
• R_o is point-in-time. The dataset reflects August–September 2025 usage. Anthropic's subsequent releases show R_o climbing for Computer & Mathematical occupations. We update the CSV when new releases ship.
• Crosswalk coverage is uneven. The O*Net Tasks → DWAs → Skills crosswalk was built to bridge cognitive work; it gets thin fast for physical-task occupations. Construction Laborers (47-2061) illustrates the failure mode — only 8 of 27 occupation tasks resolve to scored skills, so ρ_i would be computed from a tiny non-representative slice. The pipeline falls back to skill-weighted E_i below the 5-task / 25%-coverage threshold (§4), and the dashboard flags the fallback so users know which formulation produced their number.
• r̃_t joins are text-keyed. Anthropic's task_penetration.csv keys on task description text. O*Net occasionally rephrases tasks across vintages, so a small share of tasks fall out of the join when the wording changes. We hold O*Net 27.2 fixed to keep the join stable; future O*Net releases will require a re-key pass.
• Eloundou covers 11 of 35 skills. The Eloundou regression coefficients in §7 still inform the per-skill rubric's priors and the fallback E_i formulation. The remaining 24 rest on Pew skill-family priors and reasoned judgment.
• SOC vintage ambiguity. Anthropic's appendix does not specify 2018 vs. 2019 SOC. We assume 2018 based on O*Net taxonomy alignment. Minor mismatches are possible at the 6-digit level.
• Context ignored. The tool does not know the worker's geography, wage, employer size, or how a specific firm deploys AI. Mollick's (2024) “jagged frontier” applies: organizational factors routinely dominate technical feasibility.
• No welfare or policy output. v1 is a point estimate. Korinek & Suh (2024)'s scenario-planning framework — how fast, how aligned, how policy-responsive — lives above what this tool is designed to produce.