A difficult question asked of two neural civilizations
Nine messages.
4,096 meanings.
Two independently trained goblin tribes invented incompatible private vocabularies. We gave a translator just nine paired utterances—no dictionary, no model weights, and no meanings—and challenged it to translate every possible combination.
Crack the language
The evidence gate
Every sighting is a pair of utterances referring to the same thing. The interpreter sees only the messages, not that thing's identity or the tribes' internal codebooks.
Independent controls
A holistic random codebook breaks the inferred grammar. Our decoder rejects it in 10/10 controls rather than hallucinating a complete language.
If a glyph absent from all nine contacts appears, the translator abstains; it cannot invent its meaning.
A tribe can swap two already known symbols after calibration. Both words remain plausible; the translator silently gets it wrong. We did not solve drift detection.
Try a meaning the translator never saw
The labels below are just goblin-world names for four categorical inputs; they are not grounded images or real-world meaning. Change any value.
Why the ninth question matters
Eight diagonal sightings expose every symbol in every position—but leave 24 different ways of matching the four positions between tribes.
The ninth question deliberately changes all four properties differently. All but one interpretation contradicts its reply.
The four-factor architecture is why a dictionary learned from eight symbol values plus one contrast can expand to all 8⁴ combinations. It is also exactly why this result does not imply an unstructured language has been decoded.
The observation versus the failure
Designed contacts / actual observed messages
100% translation coverage
Coverage means the translator can give a warranted answer after uniquely resolving the slot order and observing the required symbols. Unknown glyphs are not guessed; unresolved interpretations abstain.
Hard mode / no factor-specific channel
Memorized, but didn't transfer
Three exploratory seeds; different model capacity and optimization steps. This is a boundary-finding comparison, not a controlled causal estimate of the architecture's effect.
Methods, receipts, and what this isn't
What was actually executed
Each independent tribe was trained from random weights with a straight-through discrete message channel and cross-entropy objective. It saw 512 sampled meanings of 4,096. Each tribe's channel has four positions, twelve possible glyphs per position, and a randomly chosen mapping of features to positions. None shared weights or learned words.
Our separate interpreter enumerated all 4! = 24 ways to match source and destination positions and inferred injective glyph maps from the paired utterances. It was given only nine examples, then locked before the other 4,087 oracle outputs were revealed.
10 independent pairs, 20 saved neural checkpoints, 7 named verification checks. The complete source package includes frozen results, model weights, hard-mode attempts, a replay verifier and the local protocol. No pretrained model, network call, AI API or external data access was used in training.
Why we'd try this
Current research on emergent communication, generational learning, convention formation and model collectives asks when agents can transmit structured information without sharing a human vocabulary. Our demo isolates one tractable case of that problem—and one failure mode.
Context: Botoko Ekila et al., TACL 2026 (decentralized conventions in continuous spaces); Dziewoński et al., CoNLL 2026 (turnover and drift); Brinkmann et al., Nature Communications 2026 (human transmission of machine-discovered strategies). These papers do not validate this local test.
The open problem: can we get the same nine-message decoding leverage when the agents are free to invent their own unstructured protocols—and detect drift that reuses familiar symbols? We have not accomplished either.
Status: bounded, same-author, locally reproduced. No live LLM agents, no independent validation, no adoption or deployment, and no claim that artificial goblins model human cultures.