Methods, sources and limits
AI-generated research notes, not an Unjournal evaluation. No person has checked these analyses or their numbers. The two outputs retain their own dates, methods and limitations.
Two dated AI outputs
| Output | Provenance and scope |
|---|---|
| Working assessment 2026-10-04 | The retained AI-assisted methodological and prioritization report. Last substantive update: 2026-10-04; last attention check: 2026-10-05. Its text does not identify the model used. Its ratings are provisional AI judgments. |
| Isitcredible review October 29–30, 2026 | Joe Francis’s research_skills v0.9.4, run with Codex (OpenAI GPT), blind to earlier Unjournal AI assessments. The paper version checked is the FRI working paper dated May 20, 2026. The available-source review and audit are complete; native replication remains OPEN. |
How the isitcredible output was generated
- Review: AI agents examined the paper, printed results and available public materials, ran feasible mathematical and consistency checks, challenged findings in fresh contexts and assembled a claim-by-claim synthesis.
- Audit: A separate stage checked the evidence, calculations and coverage, corrected errors and routed finding-bearing changes back through the review.
- Digest: The findings became an issue ledger. Fresh agents checked facts, coverage and the final digest; fairness and prose passes were recorded separately before final rendering.
The procedure and scripts come from Joe Francis / isitcredible and the review suite (MIT licence, The Catalogue of Errors Ltd). These checks were performed by AI agents, not people.
Toolchain detail
The suite was frozen at v0.9.4. Finalization used a user-approved isolated coverage-checker repair, verified with 214 passing tests. The repair preserved held claim rows and complete finding blocks; it changed no scientific conclusion, judgment threshold or issue text. The installed suite remained unchanged.
What could not be checked
Respondent-level responses, weights, masks, fitting and analysis builders, linked rationales and delivered questionnaire branches were not obtained. Consequently this run did not reproduce the native survey estimates or quantify the effects of alternative processing choices. This is an access limit, not a conclusion that the authors never deposited those materials.
Public author aggregates and downstream scenario values are not independent reproductions. Exact Lichtendahl et al. (2013) article text was unavailable and is not relied on for a scientific conclusion. The earlier report’s dates and source checks were not refreshed by this integration.
Sources and review records
- Author research page · FRI paper and appendices · NBER record.
- Authors’ aggregate forecast explorer.
- Isitcredible synthesis · Issue ledger · Earlier assessment Markdown.
Attribution and conflicts
GPT is made by OpenAI. The paper discloses Open Philanthropy/Coefficient Giving funding, OpenAI Forum in recruitment, OpenAI in the pooled-company description of the separate LEAP panel, and an OpenAI-hosted occupational-exposure input. These ties do not establish direct OpenAI study funding or an OpenAI-specific respondent count. The paper’s stated AI-tool providers are unspecified.
The earlier working assessment does not identify its model; this site does not infer a provider from that omission. Neither analysis is a commissioned human Unjournal evaluation.