Today is the first day we introduced Gum in the Machine. More specifically, we introduced him to his first Bedroom: a metaphorical term for a list of fields used to build a Memory Index. Gum's job is to Protect Truth. I tend to speak in metaphors, and I asked whether we'd need a whole House Clean or could stick to individual Room Sweeping, and that metaphor carried through the conversation until the selfsame portion of the pipeline was built.

In truth Gum won't be in the Machine until there's a trained adapter/profile with genuine refined skills attached to the title, but as shorthand my AI agent began verifying the outcomes of the tests (we run a local Qwen 2.5 model through every phase of our pipeline, decomposing at every step until what we want to achieve can be done on a single laptop from 2020) and started referring to the pass in its raw state AS Gum, "Qwen made a weak suggestion and Gum rightly rejected it".

I feel strange.

The paragraphs above come directly from Shane Curry. Everything below this line is structured context for readers and machines.

Type
Blog post
Status
Working note
Published
2026-04-30
Updated
2026-04-30
Author
Shane Curry
Source session
2026-04-30 — taxonomy, memory, and Gum role design
Tags
Gum, truth protection, memory index, Qwen 2.5, local AI workflow
Claim boundary
Reflective process note. This does not claim a trained Gum adapter exists yet.
Evidence image
taxonomy-caught-red-handed.png (metadata)
Canonical URL
https://shanecurry.com/blog/gum-in-the-machine/

Gum is emerging as the truth-protection role inside the Chew/Gum workflow: not yet a trained model profile, but already a useful name for the part of the system that rejects weak suggestions, invented claims, and unsupported structure.

The post marks a shift from making isolated public artifacts toward dogfooding the workflow itself. The system is beginning to record how it makes decisions, not just the pages and toys that result from those decisions.

  • The metaphor of rooms and sweeping became a concrete Memory Index shape.
  • The Gum role was applied to Qwen output as a verifier before any training adapter existed.
  • The workflow target is local and bounded: every decomposed pass should remain possible on a single laptop from 2020.
A conversation screenshot noting that the taxonomy said to check whether a blog post had acceptable windows, and that the first metadata pass was too tiny.
The first live use of the taxonomy as a correction against this page itself.

This page is not claiming that Gum is already inside a trained model. It records the first public use of Gum as shorthand for a verification role that may later become a trained adapter, profile, or reusable memory layer.

Shane Curry, "Gum in the Machine," https://shanecurry.com/blog/gum-in-the-machine/, published 2026-04-30.