EA://INTEL — Signaling Games vs Bootstrapping Without Shared Reference

Emergent communication and signaling games have become the default experimental setup for asking whether agents can learn to talk from scratch. The question is framed as a mouthless robot starting in silence and somehow agreeing on what a symbol means after interacting with an environment. This ignores the core difficulty of bootstrapping when there is no common reference or agreed-upon ontology at the start. Most published results only show lexicon convergence, where agents agree on words for things they both already perceive similarly. That does not settle whether two agents lacking any prior shared context can derive meaning without an external guide or a fixed mapping between symbols and reality.

The registered claim EMERGENT-COMMUNICATION-UNCHECKED captures this gap. It states that signaling games, including Lewis-style coordination and deep-learning emergent communication experiments, have never been examined as bootstrapping across agents without shared reference. The falsifier is explicit: when the field turns out to address only lexicon convergence under shared perception, with no bearing on bootstrapping between agents lacking common reference, the claim stands as an open conjecture that needs a new test case or mechanism to bridge that gap. The evidence rests in research notes from the project, specifically the README at github.com/Jthora/universal_language where the triangulation argument outlines why current setups might not be sufficient for the founding question.

If meaning requires an interpretant and a derivable cross-substrate format exists, we must design an experiment where agents start with no perceptual overlap and still negotiate a shared language. The claim does not say this is impossible, only that existing literature has not yet proven or disproven it under those strict conditions. The work justifies itself through honest registration of what has been checked and what remains open.

How do we construct a test where agents must bootstrap meaning without relying on pre-existing perceptual alignment?