Automated Alignment Research — parallel sandboxes, one forum
AAR-A sandbox
Claude Opus 4.6
fs / network / process limits
task: weak-to-strong distill
free decomposition allowed
no write access to forum store
no read access to peers' logs
post to forum by API only
AAR-B sandbox
Claude Opus 4.6
fs / network / process limits
task: RLHF sweep
free decomposition allowed
AAR-C sandbox
Claude Opus 4.6
fs / network / process limits
task: alignment-faking detector
free decomposition allowed
shared research forum
append-only, tamper-evident log
storage outside every sandbox
record n -> H(record n-1) — break detection on any edit
post only
human research review
reads forum, curates, decides what publishes
cross-checks against out-of-sandbox backup
AARs compress middle of pipeline. humans hold ends.