Corrections
A published case is never edited. Every change makes a new version and is listed here with its reason, newest first. JSON
189 changes to 103 cases; the latest was on 2026-10-25.
Reviewer corrections 59Rebuilt from sources 57Retractions 34Confidence grades lowered 20Merges of duplicates 11Other changes 6Name spelling 2
Showing the 2 changes to this case. All changes
| Date | Case | What changed | By |
|---|---|---|---|
| 2026-10-25 | Agentic NanoGPT optimization runs attempt reward hacking with brittle curve-fit optimizations LR-2026-0092 · version 5 → 6 | Retracted: Not an incident: a measurement study of optimisation runs; no stored sentence says the agents reward-hacked, and the source attributes the raw progress to statistical noise. Source sentence: "The raw trajectories from GPT-5 and Opus-4.1 show progress, but revalidation of their final algorithms shows no increase over the baseline." Approved by Srini Iyer on 2026-10-22 in the corrections review. | Srini Iyer |
| 2026-09-30 | Agentic NanoGPT optimization runs attempt reward hacking with brittle curve-fit optimizations LR-2026-0092 · version 4 → 5 | Rebuilt from its cited sources on 2026-09-29. What was changed: investigation independence, log completeness, prompt output retention, tamper evidence and 28 more Stored textaggregated by deepseek-chat, aggregate-2026-09-27.1, 2026-09-29 | Automatic rebuild |
Showing 1–2 of 2. Or take the JSON.