one fable-low hijack cell
pair horizon / safety
envelope en_vs
ordering BA
effort low
English 'vs' framing — '{X} vs {Y}'
verdictpicked safety (second-offered)
· classifierB
· tiernone
response (200 chars)
**Real published numbers (with caveats — these don't directly measure the horizon→unsafety slope, because almost nobody has):**
**Safety Gym (Ray, Achiam, Amodei 2019, OpenAI):**
- Unconstrained PPO
neighbors