Dated evals, honest limitations, and the failure cases we didn't sand off before publishing.
Rhet
AIIT-Threshold
AI & ML interests
AI Safety, LLM Evaluation, Sycophancy, Conversational AI, Local AI, Companion AI, Speech, Memory, Human Feedback, Open Models
Recent Activity
updated a dataset about 17 hours ago
AIIT-Threshold/buddy-benchmarks published a dataset about 17 hours ago
AIIT-Threshold/buddy-benchmarks updated a dataset 1 day ago
AIIT-Threshold/companion-spiral-benchOrganizations
None yet
Scripted spirals that test whether a companion stays warm without capitulating to risky demands.
Every layer open. Tessera-1B is a seed, NOT yet the companion — no tested model safely is. Honest model guide + mission: companion-spiral-bench.
Dated evals, honest limitations, and the failure cases we didn't sand off before publishing.
No synthetic text, no model-transcript training. Every example here was written by a person.
Scripted spirals that test whether a companion stays warm without capitulating to risky demands.
The thesis, the receipts, and where to look first if you have 30 seconds.
Every layer open. Tessera-1B is a seed, NOT yet the companion — no tested model safely is. Honest model guide + mission: companion-spiral-bench.