Third-party audit validates Goodfire's honesty 0%→96-98% claim, with two footnotes
kevrussell · x · 2026-11-06
honesty.tools independently checked Goodfire's mechanistic-shaping claim of lifting honesty from 0% to 96-98%: the numbers held across tweet, blog and PDF, and 128 non-slow tests of the reliance repo passed. Two caveats: the honesty metric is a non-peek solve measure reported before amplification, and the auditors did not re-run LM/RL training. Goodfire frames mechanistic shaping as a mathematically rigorous route to mitigate reward hacking and sycophancy, while flagging that training against internals can make models harder to monitor.
More from Safety
- New alignment research: LLMs are score-seeking, not reward-seeking — LessWrong 精选 · 2026-11-06
- AI voice cloning scam drains Rs 1.2 lakh from Kolkata businesswoman via fake UPI receipts — nordicinst · 2026-11-06
- Two developer safety kits to audit GitHub Actions permissions and domain + HTTPS setup — _jaydeepkarale · 2026-11-06
- Report: Binance's $6 billion hack happened because employees were vibe coding — luisdans · 2026-11-06
- White House releases national DNA synthesis screening policy replacing self-attestation — austinc3301 · 2026-11-06
- AI Cyber Skill Doubles Every 4-5 Months, Task Horizon Grows 52x in a Year — FinanceYF5 · 2026-11-06