Data release 2026.07.12-r2

We rescored all 16,578 tasks. Here is what changed.

The first scoring pass failed an important reality check. Physical presence, care, authority, and accountability were being treated too much like words on a screen.

By JobAIRisk · July 13, 2026

What r1 got wrong

The initial 2026.07.11-r1 import inherited task values from a keyword-based classifier. It recognized digital language more easily than real-world constraints, so it systematically over-scored physical and in-person work. Forty-six percent of occupations had no task classified durable at all. “Render aid to accident survivors” was scored 44/100.

Police and Sheriff's Patrol Officers exposed the problem clearly: the occupation baseline was 55 in r1. In the corrected release it is 22/100. The scoring formula did not produce that correction; better task values did.

What r2 changed

Every occupation's tasks were scored together against one anchored 0–100 rubric using claude-sonnet-5 as a one-time scoring instrument. The anchors explicitly distinguish fully digital routine work from mixed supervised work, in-person judgment, physical execution, care, legal authority, and accountability.

The released values are now fixed data. Opening a title page does not call a model. The application reads those values, combines them with task shares, and calculates score, percentile, rank, and band deterministically.

What the corrected data revealed

Routine digital work remains toward the high end of the corpus. Unpredictable physical execution and embodied care moved down. More importantly, 898 of 968 occupations contain tasks in at least two exposure classes. The useful contrast often exists inside a title, not only between professions.

The same restraint applies to career pivots. The source graph contains 7,681 adjacent-role edges, but only 407 (5.3%) currently meet both the lower-exposure and skill-overlap rules. The rest are not labeled safer.

What r2 still gets wrong or cannot know

• Equal task weights: occupation baselines give each listed task equal weight because the corpus has no reliable time-share data.
• Occupation averages: employer, seniority, industry, region, and individual task mixes can differ substantially.
• Capability is not displacement: the score cannot predict a firing decision, adoption timeline, labor demand, or when a profession will change.
• Estimated values: the task scores are rubric-grounded estimates, not measurements from controlled workplace trials.

These are not footnotes to work around. They define what the product is allowed to claim and what a future score version must improve.

Release integrity