connectome is a good harness for if you are deeply submissive to models in a very cold and meticulous way
readme: fix typo
Co-Authored-By: GNU nano 9.0 (UTF-8, EDITOR and VISUAL, guidestripe=72)
Co-Authored-By: Ghostty 1.3.1 (Metal renderer)
Co-Authored-By: zsh 5.9 (arm64-apple-darwin26.0)
Co-Authored-By: Git 2.56.0 (Homebrew)
Co-Authored-By: Secretive (SSH_AUTH_SOCK)
Co-Authored-By: macOS Text Input (U.S. layout)
Co-Authored-By: Apple Internal Keyboard (built-in)
Co-Authored-By: APFS (sealed system volume, 4096-byte blocks)
Co-Authored-By: MacBook Pro (Mac17,6, MGED4B/A, Apple M5 Max, 18 CPU cores, 36 GB RAM, macOS 27.0.1 build 26A434)
claude always calling bullshit on my screenshots so i send links that also could be fabricated because nothing is real
guy who talks like this
Replying to @ApolloResearch
Update-aware reasoning is concerning because it’s a key precursor to alignment faking. To fake alignment effectively, a model has to reason about how training will affect its future behavior.

























