# Sources — Training Is Part of the Cost of Oversight

1. Recursive Institute, [“Oversight and Training Capacity: A Reproducible Model Note,” version 1, 1 November 2026](https://recursive.institute/editorial/oversight-training-model-note-2026-11-01.pdf). This standalone note supplies the synthetic definitions, arithmetic and provenance for the commentary's comparison. It restates a bounded theoretical calculation, not observed professional formation rates or a staffing recommendation. Ordinary human-capital accounting is not claimed as a new discovery. The primary studies below supply no estimated model parameters. Model-note SHA256: `67033e7647ecc7cb1eae80c822b4b18550ebcd3f15825d3a5125947f02bd3faf`.

2. Judy Hanwen Shen and Alex Tamkin, [“How AI Impacts Skill Formation,” arXiv:2601.20245v2](https://arxiv.org/abs/2601.20245v2), revision dated 1 February 2026. See sections 4.3, 5.2, 6 and 7.1. Randomized AI access and immediate unaided evaluation while developers learned Trio; qualitative interaction styles were not randomly assigned mentoring policies. The short study does not estimate long-run professional qualification, retention or attrition. The HTML masthead displays a different date; this citation remains pinned to v2.

3. Hamsa Bastani et al., [“Generative AI without guardrails can harm learning: Evidence from high school mathematics,” PNAS](https://doi.org/10.1073/pnas.2422633122), 2025. See main results and experimental design. Class-level randomized comparison of a standard AI interface, a teacher-designed guarded tutor and no AI. The guarded tutor mitigated the unassisted-exam harm in the standard-interface arm; a positive unassisted-exam effect was not established. The [published correction](https://doi.org/10.1073/pnas.2518204122) concerns an author affiliation, not the results. This is a different task and population from the developer experiment or professional verification; the studies are not pooled.

Source qualification uses the retained primary captures checked on 22 October 2026 and re-read for this draft on 1 November. No fresh online source retrieval or claim of complete current literature coverage was made. No new participant data were collected or analyzed.