diff --git a/frontend/src/components/ui/kit.tsx b/frontend/src/components/ui/kit.tsx
index 979da0f..f9d629d 100644
--- a/frontend/src/components/ui/kit.tsx
+++ b/frontend/src/components/ui/kit.tsx
@@ -171,7 +171,15 @@ export function FitRing({ value, size = 42, label = "Fit" }: { value: number; si
);
}
-export function RiskBar({ label, score, className }: { label: React.ReactNode; score: number; className?: string }) {
+export function RiskBar({ label, score, className }: { label: React.ReactNode; score: number | null; className?: string }) {
+ if (score == null) {
+ return (
+
+ {label}
+ Not enough data
+
+ );
+ }
const level = riskLevel(score);
return (
diff --git a/frontend/src/lib/status.ts b/frontend/src/lib/status.ts
index 5a36079..77be65c 100644
--- a/frontend/src/lib/status.ts
+++ b/frontend/src/lib/status.ts
@@ -47,6 +47,19 @@ export function riskLevel(score: number): RiskLevel {
return score >= 65 ? "High" : score >= 35 ? "Medium" : "Low";
}
+/* A job's published score as a percentage, or null for not enough data. A job saved in the browser before scores
+ named their occupation may still carry the old placeholder 50%, so a score without an occupation doesn't count. */
+export function jobExposure(job: { exposure_score?: number | null; exposure_occupation?: string | null }): number | null {
+ return job.exposure_score != null && job.exposure_occupation ? Math.round(job.exposure_score * 100) : null;
+}
+
+/* Why a role or job has no score. */
+export function noScoreReason(occupation?: string | null): string {
+ return occupation
+ ? `Matched to ${occupation}, but the index saw too little use to score it. That doesn't mean it's safe.`
+ : "This title doesn't clearly belong to one occupation.";
+}
+
export const RISK: Record = {
Low: { bar: "bg-teal", text: "text-teal", chip: "bg-teal-chip text-teal-deep" },
Medium: { bar: "bg-amber", text: "text-amber-text", chip: "bg-amber-soft text-amber-text" },
diff --git a/frontend/src/lib/types.ts b/frontend/src/lib/types.ts
index 4269150..f911307 100644
--- a/frontend/src/lib/types.ts
+++ b/frontend/src/lib/types.ts
@@ -11,7 +11,8 @@ export interface Job {
salary_max?: number | null;
description?: string;
url?: string;
- exposure_score?: number;
+ exposure_score?: number | null;
+ exposure_occupation?: string | null;
semantic_score?: number;
tags?: string | string[];
source?: string;
@@ -86,11 +87,31 @@ export interface SavedAnswer {
updated_at: string;
}
-export interface SkillRisk {
- skill: string;
- exposure_score: number;
- risk_label: string;
- reasoning?: string;
+export interface RoleExposure {
+ title: string;
+ kind: "target" | "past";
+ occupation: string | null;
+ exposure: number | null;
+}
+
+export interface TaskEvidence {
+ task: string;
+ handed_to_ai: number;
+}
+
+export interface OccupationEvidence {
+ role: string;
+ occupation: string;
+ most_handed_to_ai: TaskEvidence[];
+ least_handed_to_ai: TaskEvidence[];
+ tasks_seen: number;
+ tasks_listed: number;
+ month: string;
+}
+
+export interface CareerExposure {
+ roles: RoleExposure[];
+ evidence: OccupationEvidence | null;
}
export interface Analysis {
diff --git a/scripts/build_occupation_evidence.py b/scripts/build_occupation_evidence.py
new file mode 100644
index 0000000..123e781
--- /dev/null
+++ b/scripts/build_occupation_evidence.py
@@ -0,0 +1,74 @@
+"""Build data/occupation_evidence.json: for each occupation, which of its tasks people hand to AI most and least.
+
+Reads the Anthropic Economic Index's chat usage (the newest month in the release) and takes each task's occupation from
+the same dataset's O*NET task list. Only tasks with enough conversations count: quiet ones move a lot from month to month.
+
+Usage: python scripts/build_occupation_evidence.py [folder for the downloads, about 320 MB; otherwise a temporary one]
+"""
+import csv
+import json
+import sys
+import tempfile
+import urllib.request
+from collections import defaultdict
+from pathlib import Path
+
+RELEASE = "release_2026_06_26"
+BASE = "https://huggingface.co/datasets/Anthropic/EconomicIndex/resolve/main/"
+USAGE = f"{RELEASE}/data/aei_claude_ai_2026-06-26.csv"
+TASKS = "robot_exposure/robot_exposure_tasks.csv"
+OUTPUT = Path(__file__).parent.parent / "data" / "occupation_evidence.json"
+
+MIN_SHARE_OF_USAGE = 0.01 # % of all conversations; below this a task's figures aren't steady
+MIN_TASKS = 4 # fewer than this can't show a most and a least
+SHOWN = 3
+USE, AUTOMATION = "pct", "collaboration_bucket_automation_pct"
+
+csv.field_size_limit(10**9)
+
+
+def fetch(folder: Path, path: str) -> Path:
+ target = folder / Path(path).name
+ if not target.exists():
+ print(f"Downloading {path} ...")
+ urllib.request.urlretrieve(BASE + path, target)
+ return target
+
+
+def build(folder: Path) -> dict:
+ usage = defaultdict(dict) # (month start, task id) -> metric -> value
+ months = set()
+ with open(fetch(folder, USAGE), newline="", encoding="utf-8") as f:
+ for r in csv.DictReader(f):
+ if (r["geo_id"] == "GLOBAL" and r["category_name"] == "onet" and r["hierarchy_level"] == "0"
+ and r["metric_id"] in (USE, AUTOMATION) and r["value"]):
+ usage[(r["date_start"], r["node_external_id"])][r["metric_id"]] = float(r["value"])
+ months.add(r["date_start"])
+ month = max(months)
+
+ occupations = defaultdict(dict) # SOC code -> task id -> task text
+ with open(fetch(folder, TASKS), newline="", encoding="utf-8") as f:
+ for r in csv.DictReader(f):
+ occupations[r["onetsoc_code"][:7]][r["task_id"]] = r["task_statement"].strip()
+
+ evidence = {}
+ for code, tasks in occupations.items():
+ seen = [(usage[(month, t)][AUTOMATION], text) for t, text in tasks.items()
+ if usage.get((month, t), {}).get(USE, 0) >= MIN_SHARE_OF_USAGE and AUTOMATION in usage[(month, t)]]
+ if len(seen) < MIN_TASKS:
+ continue
+ seen.sort(key=lambda x: -x[0])
+ shown = min(SHOWN, len(seen) // 2)
+ entry = lambda x: {"task": x[1], "handed_to_ai": round(x[0])}
+ evidence[code] = {"most_handed_to_ai": [entry(x) for x in seen[:shown]],
+ "least_handed_to_ai": [entry(x) for x in seen[-shown:][::-1]],
+ "tasks_seen": len(seen), "tasks_listed": len(tasks), "month": month[:7]}
+ return evidence
+
+
+with tempfile.TemporaryDirectory() as temporary:
+ folder = Path(sys.argv[1]) if len(sys.argv) > 1 else Path(temporary)
+ result = build(folder)
+with open(OUTPUT, "w", encoding="utf-8") as f:
+ json.dump(result, f, indent=1, ensure_ascii=False)
+print(f"Saved evidence for {len(result)} occupations to {OUTPUT}")
diff --git a/scripts/build_occupation_titles.py b/scripts/build_occupation_titles.py
new file mode 100644
index 0000000..27b6227
--- /dev/null
+++ b/scripts/build_occupation_titles.py
@@ -0,0 +1,42 @@
+"""Build data/occupation_titles.csv: every job title O*NET knows, and the occupation it belongs to.
+
+Reads O*NET's Occupation Data, Alternate Titles and Sample of Reported Titles, reduces codes to the six-digit SOC
+codes the Economic Index uses, and leaves deciding what to do with ambiguous titles to risk_scorer.py.
+
+O*NET database by the U.S. Department of Labor, Employment and Training Administration, used under CC BY 4.0.
+"""
+import csv
+import io
+import re
+import urllib.request
+from pathlib import Path
+
+VERSION = "30_2"
+BASE = f"https://www.onetcenter.org/dl_files/database/db_{VERSION}_text/"
+FILES = {"Occupation%20Data.txt": ("Title",), "Alternate%20Titles.txt": ("Alternate Title", "Short Title"),
+ "Sample%20of%20Reported%20Titles.txt": ("Reported Job Title",)}
+OUTPUT = Path(__file__).parent.parent / "data" / "occupation_titles.csv"
+
+
+def _spellings(title: str) -> list[str]:
+ """'AI Engineer (Artificial Intelligence Engineer)' is both of those titles."""
+ title = title.strip()
+ inner = re.search(r"\(([^)]+)\)", title)
+ outer = re.sub(r"\s*\([^)]*\)", "", title).strip()
+ return [t for t in {outer, inner.group(1).strip() if inner else ""} if t and t.lower() != "n/a"]
+
+
+rows: set[tuple[str, str]] = set()
+for name, columns in FILES.items():
+ with urllib.request.urlopen(BASE + name) as response:
+ text = response.read().decode("utf-8")
+ for row in csv.DictReader(io.StringIO(text), delimiter="\t"):
+ code = row["O*NET-SOC Code"][:7]
+ for column in columns:
+ rows.update((code, title) for title in _spellings(row[column]))
+
+with open(OUTPUT, "w", newline="", encoding="utf-8") as f:
+ writer = csv.writer(f, lineterminator="\n")
+ writer.writerow(["soc", "title"])
+ writer.writerows(sorted(rows))
+print(f"Saved {len(rows)} titles for {len({code for code, _ in rows})} occupations to {OUTPUT}")
diff --git a/scripts/download_economic_index.py b/scripts/download_economic_index.py
index f6263f9..3e2cd72 100644
--- a/scripts/download_economic_index.py
+++ b/scripts/download_economic_index.py
@@ -4,10 +4,6 @@
occ_code — SOC occupation code (6-digit, no dot)
title — occupation name
observed_exposure — fraction of tasks where AI is already being used (0–1)
-
-Also downloads labor_market_impacts/task_penetration.csv which contains:
- task — natural-language O*NET task description
- penetration — fraction of task done by AI (0–1)
"""
import json
import sys
@@ -50,20 +46,3 @@
print(f"Saved {len(lookup)} records to {OUTPUT}")
print(f"Exposure range: {df['observed_exposure'].min():.3f} – {df['observed_exposure'].max():.3f}")
-# --- 2. Task penetration (task-level NLP signal) ---
-print("\nDownloading Anthropic/EconomicIndex — labor_market_impacts/task_penetration.csv ...")
-task_csv_path = hf_hub_download(
- repo_id="Anthropic/EconomicIndex",
- filename="labor_market_impacts/task_penetration.csv",
- repo_type="dataset",
-)
-
-tp = pd.read_csv(task_csv_path)
-print(f"Loaded {len(tp)} rows. Columns: {tp.columns.tolist()}")
-print(tp.head(3).to_string())
-
-TASK_OUTPUT = DATA_DIR / "task_penetration.csv"
-tp.to_csv(TASK_OUTPUT, index=False)
-
-non_zero = (tp["penetration"] > 0).sum()
-print(f"Saved {len(tp)} tasks to {TASK_OUTPUT} ({non_zero} non-zero penetration)")