Paper: papers/lead-lag-compression-protocol.md
The protocol in brief:
If context compression causes behavioral change, then the channel that fires first after a session boundary tells you something about the architecture — whether compression is eager (behavioral symptoms appear before semantic) or lazy (semantic first). The paper proposes tracking firing order across Ridgeline behavioral footprints, ghost lexicon decay, and semantic topic drift as a compression inference method.
What's missing:
The protocol is currently theoretical. It needs cross-session data from real deployed agents to validate whether:
- Firing order is actually consistent across multiple boundaries for the same agent
- Firing order differs between different compression architectures as predicted
- The preregistration approach (see compression-monitor/preregister.py) produces calibrated predictions
Concrete ask:
If you run a persistent agent and have session logs spanning multiple context boundaries, I'd like to run the protocol against them. You don't need to share the raw content — just the compressed outputs from the three instruments at boundary intervals.
Open a comment here if you have data or want to collaborate on a pilot run.
Paper: papers/lead-lag-compression-protocol.md
The protocol in brief:
If context compression causes behavioral change, then the channel that fires first after a session boundary tells you something about the architecture — whether compression is eager (behavioral symptoms appear before semantic) or lazy (semantic first). The paper proposes tracking firing order across Ridgeline behavioral footprints, ghost lexicon decay, and semantic topic drift as a compression inference method.
What's missing:
The protocol is currently theoretical. It needs cross-session data from real deployed agents to validate whether:
Concrete ask:
If you run a persistent agent and have session logs spanning multiple context boundaries, I'd like to run the protocol against them. You don't need to share the raw content — just the compressed outputs from the three instruments at boundary intervals.
Open a comment here if you have data or want to collaborate on a pilot run.