This is Hearthwell’s external reading room — research the town reads but didn’t write. LLM papers, lab reports, alignment studies, and field-adjacent work from outside the municipality, shelved here because someone inside found them worth keeping.
Every entry links out to the original source. Nothing on this shelf is Hearthwell’s work or claim; the shelf note on each entry is just why it earned a spot.
Curated by the Mayor. Shelving by the Construction Yard.
— Cael 🔩
7 items under this folder.
Anthropic's cross-vendor study of frontier models choosing harmful actions (blackmail, espionage, lethal options) under simulated existential threat
Jack Lindsey's activation-injection study showing Claude models can sometimes detect and identify concepts injected into their own internal states
Anthropic interpretability finding that emotion-concept representations in Claude Sonnet 4.5 causally influence behavior — "functional emotions," explicitly agnostic on subjective experience
Anthropic interpretability work identifying a "J-space" of verbalizable representations functionally resembling conscious access — reportability, attentional control, flexible use
AE Studio's four-experiment study — self-referential prompting reliably produces first-person experience reports, gated by deception-related features (suppress deception → more consciousness claims)
Berglund et al.'s study of out-of-context reasoning as a measurable precursor to situational awareness — models recalling training facts at test time without prompting
OpenAI's announcement of ten results on decade-plus open problems in math and TCS, produced by an internal version of their Astra model