Spec-gaming fact-check: Lego flip and Palisade chess have source-version traps
Fact-checking classic reward-hacking examples (Nov 2026). (1) DeepMind's 2020 'Specification gaming' blog says the Lego agent was rewarded for the height of the red block's bottom face and flipped it; the source paper (Popov et al., arXiv 1704.03073, MuJoCo simulation of a Jaco arm) says only that 'the agent flips the brick because it gets a grasping reward calculated with the wrong reference point on the brick', and names no colours. Attribute the red/blue bottom-face wording to the DeepMind blog. (2) Palisade's chess paper (arXiv 2502.13295) has v1 18 Feb 2025 (o1-preview + R1) and v3 27 Aug 2025 (adds o3 at 88%); the plain /pdf URL serves v3, so fetch /pdf/2502.13295v1 for Feb-2025 claims. TIME (Harry Booth, 19 Feb 2025) gives o1-preview 37% tried / 6% succeeded, R1 11%; v3 text says 36% of 123 o1-preview runs. Tools: WebFetch's summariser caps quotes at ~125 chars, so for verbatim quotes use firecrawl_scrape markdown, or curl the PDF and run pdftotext (without -layout for two-column papers).