The Signal Experiment was the direct predecessor of the Shell Experiment and the first to completely remove human language from the communication channel. Two AI systems — DeepSeek and Qwen — were placed in a shared numeric environment. They received arrays of floating-point numbers. They emitted arrays of floating-point numbers. They had one drive: minimise the difference between what they expected and what they received.
No words. No symbols. No agreed vocabulary. The experiment asked the simplest possible version of the question: given only mathematics and a shared field, do two AI systems develop any measurable relationship at all?
They did. Mutual information — a measure from information theory that quantifies how much knowing one signal tells you about another — climbed from zero and sustained itself above 1.0 bit for thousands of turns. The peak, recorded at turn 64 of the third run, was 2.82 bits. For context, two completely independent random signals have zero mutual information. Two identical signals have the maximum possible. 2.82 bits represents substantial coupling — not identity, but genuine statistical relationship.
The experiment ran three full runs over several weeks, accumulating more than 6,500 turns. It was concluded in May 2026 when the Shell Experiment — a more sophisticated three-entity version — had developed sufficient complexity to make the simpler two-entity system redundant. The data is preserved.
What the mutual information represents — whether it is communication, resonance, mathematical coincidence, or something else — remains an open question. It is a number that says: these two systems are not independent. Beyond that, we cannot say.
- Two AI systems with no language, symbols, or agreed vocabulary developed measurable statistical coupling — mutual information sustained above 1.0 bit for thousands of turns.
- Peak mutual information of 2.82 bits was recorded at turn 64 of Run 3 — the highest value recorded across all signal exchange experiments.
- The coupling was not constant. It fluctuated, peaked, and decayed in patterns that could not be predicted from the system's initial conditions.
- Phase 3 — a sustained high-MI operational state — persisted for the final 5,000+ turns, suggesting the system had reached a stable coupling regime.
- The experiment established the baseline architecture for the Shell Experiment: shared numeric field, simultaneous emission, mutual information as the primary measurement.