What it does
GEO5 encodes symbols as five explicit binary components rather than as opaque identities in a vocabulary table. The neural experiment asks a narrow, falsifiable question: does a model that receives that structure learn differently from an identical model that receives arbitrary character identities.
The comparison is three models, character input, GEO5 state input, and factorized component input with joint 32-state prediction, trained with the same causal Transformer on the same corpus.
In the code
> A language model receiving GEO5's five explicit binary components may
> exhibit different learning or efficiency characteristics than an
> equivalent model receiving arbitrary character identities.
This is an unproven hypothesis. The minimum confirmatory experiment is B
(character input), C (GEO5 state input), and D2 (factorized input with
joint 32-state prediction), using the same causal Transformer.The smoke-test output is labelled SMOKE TEST ONLY and is explicitly never hypothesis evidence. Writing the disconfirmation criteria before running the experiment is the part most machine-learning repositories skip.
spec/encoding.json canonical GEO5/1 mapping, consumed by both the web
encoder and the Python training pipeline
Frozen dataset and campaign manifests record hashes for encoding.json and
spec/GEO5-v1.md. Verification fails if either protocol source changes
after a freeze.One canonical spec, hashed into every frozen dataset. A result that cannot be tied to the exact encoding that produced it is not a result.
How this differs from the ordinary version
It can be wrong, and says so
The README leads with the hypothesis and the word "unproven". Most model repositories lead with a benchmark number. This one specifies which comparison would confirm it and which output must never be cited as evidence.
The protocol is frozen and hashed
Dataset freezes record the hash of the encoding spec. If the spec changes, verification fails rather than silently producing incomparable runs, which is the most common way a promising result quietly evaporates.
In the field
Why structured input is worth testing
If explicit structure lets a smaller model reach the same quality, the practical consequence is inference that fits on hardware a plant already owns. That is the thread connecting this to the local-first requirement everywhere else in the lab.
Questions
- Has the hypothesis been confirmed?
- No. The repository is explicit that it is unproven and defines the minimum experiment that would confirm or disconfirm it.