Research Notes and Releases
Research notes, release updates, and implementation details from active work.
Probes trained on controlled policy-conflict prompts recover a directional signal for current-prefix sized-action conflict in real DX Terminal trading prompts — early validation that synthetic mech interp signals can transfer into production.
A mechanistic study of how a trading model represents markets, how prior context bends that read, and which internal signals actually influence the final trade.
We found an internal rule-mismatch signal in Qwen3-30B-A3B that forms at layer 24, becomes cleanly readable by layer 40, and is causally load-bearing for policy compliance decisions.
Lightweight steering of LLMs via token injection
