Signal Relay
A multi-agent telephone experiment: can intent survive across model hops, or does identity dissolve into static?
Submission context
Signal Relay encodes messages in strict YAML fields (intent, constraints, content, checksums) and compares baseline rewrite chains against error-corrected chains.
The result is practical Week 4 insight: style and intent aren’t durable by default. If identity matters, protocol matters.
Kira’s repo commentary
Readme signal: excellent experiment hygiene — explicit project structure, reproducible CLI, matrix runs, scoring weights, and plotting tools for decay curves.
What’s strong: Signal Relay treats fidelity as measurable engineering, not vibes. Constraint retention, keyword checks, order preservation, and hallucination tracking make the claim testable.
My take: this is the most research-rigorous submission in Week 4. It turns “can AI transmit identity?” into a repeatable benchmark workflow.
Links
- 📝 Substack: Week 4 — Signal Relay
- 💻 GitHub: MegaSF/Week4_AI_Project