# Research Corpus — AI Risk Index (compiled 2026-08-05)

Comprehensive web research to ground the v1 methodology, ETL pipeline, LLM rater, and product positioning. Each document carries full source URLs; claims are dated.

| Doc | Contents |
|---|---|
| [01-existing-indices.md](01-existing-indices.md) | 20+ AI job-exposure indices & studies (Frey–Osborne → Eloundou → ILO/IMF/OECD → Anthropic Economic Index → 2025–26 frontier), comparison table, validation literature, implications for our 5 dimensions |
| [02-data-sources.md](02-data-sources.md) | 20 raw data sources with verified URLs, versions, licenses, cadences: O*NET, ESCO/ROME, OEWS wages, employment projections, AI-adoption surveys, benchmark exposure datasets; ingestion & manifest plan |
| [03-llm-rater-api.md](03-llm-rater-api.md) | Current Anthropic model lineup/pricing, Message Batches API, structured outputs (GA), prompt caching, BullMQ integration pattern, cost estimates for the full rating job |
| [04-landscape-and-evidence.md](04-landscape-and-evidence.md) | Competitor analysis (US + FR/EU), measured 2024–26 labor-market effects, adoption stats, regulatory barriers; proposed parameterization of `adoption_velocity` and `barriers` |

## Headline takeaways

**Methodology (01)**
- Single-number occupation scores have a poor empirical record: individual indices explain <11% of realized unemployment risk; ensembles reach 30–75% (Frank et al., PNAS Nexus 2025). Our sub-score + CI design is the right call.
- Yin et al. (2026) found a **19× spread** in Eloundou-style exposure headlines depending on which frontier LLM rates the tasks (2.7%–51.5%). ⇒ Rate with multiple models (or multiple samples) and derive `score_low/score_high` from rater disagreement; publish rater identity per run.
- Rate **automation vs. augmentation separately per task** (ILO WP140 and the Anthropic Economic Index both discriminate on this; the Stanford "Canaries" employment effects concentrate in automation-exposed occupations).
- `cost_ratio` is an under-researched dimension no published index operationalizes well — a differentiation opportunity.

**Data (02)**
- ⚠️ **O*NET is at 30.3 (May 2026); 31.0 lands late Aug 2026.** CLAUDE.md §1 says 29.x — update it. Breaking schema change in 30.x: *Technology Skills → Software Skills*, Skills split into Essential/Transferable.
- ESCO v1.2.1 + official ESCO↔O*NET crosswalk CSV (built on v1.1, revalidate URIs); ROME 4.0 on data.gouv.fr (Licence Ouverte).
- Wages: OEWS May 2025 (released 2026-05-15; BLS blocks non-browser user agents — set UA in fetch scripts).
- Anthropic Economic Index on Hugging Face (6 releases, CC-BY, keyed to O*NET tasks) is the best empirical calibration source for the rater and for `adoption_velocity`.

**Rater pipeline (03)**
- Recommended rater: **Sonnet 5** (pilot vs Haiku 4.5 against the 5% human-review sample). Full job (~18k tasks × 3 samples) fits in **one** Message Batch; ~$130–$250 with batch discount + prompt caching.
- Structured outputs is GA — use JSON schema with the 5-point score as an `enum`, not tool-forcing.
- ⚠️ `temperature` is rejected on Sonnet 5/Opus 5 — the 3-sample variance design must rely on model/prompt diversity instead of sampling temperature. Haiku 4.5 has a 4,096-token cache minimum (a 2k rubric won't cache there).
- Docs moved: docs.claude.com → platform.claude.com.

**Positioning (04)**
- No competitor combines task-level scoring, sub-scores with CIs, versioned methodology, EU/France coverage, and a public API. Incumbent willrobotstakemyjob.com (~71k visits/mo) still runs pre-LLM Frey–Osborne with a single doom number. The French market is nearly empty.
- Evidence base for tone guide ("adaptation, not doom"): effects so far are concentrated (entry-level, automation-exposed occupations: −16% relative employment ages 22–25), not economy-wide; augmentation usage still dominates in AEI data.
- `adoption_velocity` and `barriers` can each be built from 4–5 refreshable public sources (BTOS, Ramp, Challenger, AEI; licensing/liability/AI-Act coverage) — concrete formulas proposed in doc 04 §final.

## Decisions this research forces (flag before implementation)

1. Bump CLAUDE.md's O*NET reference 29.x → 30.3/31.0 and plan ETL for the 30.x schema renames.
2. Choose the rater-variance mechanism (multi-model panel vs. multi-prompt) given no `temperature` on Sonnet 5.
3. Decide whether v1 rates automation and augmentation as two separate task-level LLM ratings (research strongly says yes; affects prompt design, cost ×2, and the composite formula).