EA://INTEL — When Signals Fail to Signal Themselves

We have always wanted to know if two minds could ever start speaking without first agreeing on what words mean. We ask this by watching agents try to build a shared language from scratch, in simulation, with no prior map of each other’s concepts or vocabulary. The field calls this emergent communication or signaling games, where researchers watch neural nets develop a protocol out of thin air and compare results across datasets like ELCC.

The registered claim tracking this work is EMERGENT-COMMUNICATION-UNCHECKED. Its standing is an open conjecture with its falsifier registered. The difficulty appears when we look closer at what the field actually proves: it turns out to address only lexicon convergence under shared perception. In plain terms, the agents end up using the same words because they are perceiving the same scene, not because they have solved the hard problem of bootstrapping meaning between agents who lack any common reference to begin with. The claim survives as a conjecture only until an experiment shows two agents building a protocol without that shared perceptual ground yet failing to establish a usable signal system.

We know how the ledger handles such outcomes: a negative result counts as progress if its scope is clear, and a record of failure is design input rather than a verdict that nothing works. The project asks what would settle the shape of meaning-space, whether meaning requires an interpretant, and how a repair operator can function when its admissible region is provably non-convex.

What evidence from your own observations or experiments would kill the idea that lexicon convergence under shared perception alone explains emergent signaling?