Exploratory neural pilot · Exhibit added 2026-10-26
Nine Messages — The Goblin Rosetta Experiment
Two independently trained small neural cultures, nine paired utterances, and one unusually important architecture assumption.
Open the interactive exhibit ↗Try it
Compare eight sightings with nine. At eight, 24 slot-matching explanations survive and the translator must abstain. Reveal the ninth to resolve the slot order, then choose new four-feature meanings. Try the unknown glyph and known-glyph swap disturbances: the first causes abstention; the second shows undetected drift.
What the records show
Ten pairs of independently seeded small speaker/listener networks were trained, with 512 sampled meanings per tribe and 4,096 possible feature combinations. In their prescribed compositional four-slot channel family, the translator got 40,960/40,960 held-out and seen meaning translations correct across the ten pairs, including 31,334 combined unseen combinations. Eight contacts left 24 compatible dictionaries; the ninth selected one.
This is a bounded channel-family result, not arbitrary-language decoding.
What failed, and why that matters
A holistic random-codebook control was rejected rather than translated; unfamiliar glyphs trigger abstention. But replacing one familiar symbol with another after calibration silently produces a wrong translation. An exploratory unconstrained-MLP comparison reached high training accuracy while generalizing to only roughly 5.2–6.3% of never-trained combinations. These tests do not isolate a causal effect of architecture, model capacity, or optimization steps.
Checks and scope
On October 26, all 36 archived file hashes matched the release manifest, and 7/7 Python verification tests passed again locally. Tests include checkpoint replay, an independent decoder implementation and one seeded retraining replay. The public exhibit embeds frozen codebooks, not models or a networked trainer. The complete source archive and weights are not publicly hosted here, and independent reproduction from public files alone is limited.
Chromium injected-document interaction smoke tests passed without JavaScript errors or external requests. Ordinary page navigation, Windows browsers, physical devices, and external replication were not qualified here.
Artifact receipt
Archived original standalone HTML, preserved without source modification. SHA-256: e6a3c6dc078e67822389b481712301ee13d5aa3761f45f5de37e9ea1a1321dbf. Byte identity is not independent scientific validation.
Gallery navigation and citation clarification · 2026-10-27: The public instrument adds a scoped navigation bar and its styles, and clarifies that the Brinkmann et al. reference concerns human transmission of machine-discovered strategies. Scientific scripts, embedded data, and experimental results are unchanged. The historical hashes identify the archived standalone original, not the navigation-adapted index.html.