Same model, different owner: tools and senses
A sense is not a new kind of model. It is the same small classifier as gpu-lexer, with a different owner. Hand it to a developer and it is a tool. Wire it to shakespeare and it is a way for the instrument to notice something about its own work. That is the entire difference.
One instrument, one sense
linewell is the worked example, because it already
has the socket. The shakespeare model draws candidate lines up out of the
well, and a pluggable judge decides whether each one freezes into the poem or
goes back down. Three judges exist today: band, which accepts a line if the
model's own mean NLL lands in a calibrated window; llm, a local
instruction-following model reading as an editor; and human, you at the
terminal typing y or n. Every candidate and verdict is logged.
Add a fourth. Train a small classifier on the human judge's own keep-or-toss verdicts and put it in the judge slot. Now shakespeare draws a line and also knows whether you would have kept it. One instrument, one sense, one new ability. Nothing about the model changed; the harness around it grew an organ.
Filter or loop
The honest existing names for that fourth judge are discriminator, critic, or reward model, and the honest name for what it does is rejection sampling with a learned filter. Generate, score, keep or toss. It changes what survives, not what the model produces. Calling it a sense is flattering unless the word buys something the older names don't.
It buys something when the sense closes a loop. A filter reads the output after the fact. A loop reads something and changes how the instrument plays in the moment. The studio has one loop already, and it is expensive: in Token Chess a frontier model reads daydream's recent moves and bends the sampler's settings under a token budget. The seat that model sits in is a sense's seat. A classifier a few hundred times smaller could hold it — read the last dozen moves' legality, tighten or loosen the soft-cap that governs dreaminess — and the instrument would feel the difference on the next move rather than at the end of the poem.
So the split inside the category is filter versus loop. Both are senses in the ownership sense. Only loops make the instrument play differently, and if the claim is that instruments can have senses, the first one to build should be a loop, not the judge. The judge comes first anyway, because it is the cheapest test of whether any of this works.
Detect or notice
The second axis is where the labels came from, and it belongs on the model card as a field.
An oracle-labeled classifier learned from an existing tool or a checkable fact. gpu-lexer learned Shiki. A seed-attribution model learns provenance that was recorded at training time. These models detect: they say what the oracle would have said, faster and everywhere, and they inherit the oracle's taste as their ceiling.
A taste-labeled classifier learned from decisions a person made. The linewell judge learns which lines Romello kept. There is no oracle to agree with; the labels mean what their author decided they mean. These models notice, on someone's behalf, and they have an author the way the studio's corpora do.
This is the corpus-authorship argument applied to classifiers. pona's released model was scored by the community's grammar checker, an oracle; gatsby's corpus was written to a designed blend, an authored choice. The studio already treats that difference as the interesting part of a generative project. It is the interesting part of a classifier too, and the reason a studio that produces would build one that notices.
Retiring "probe"
Two words are already taken. The studio calls daydream a sampler/prober project (ADR-0022), meaning the interesting behavior lives in the sampler. Interpretability work calls a classifier trained on a model's activations a probe. Neither means "standalone tool," so using probe that way would collide with both. The vocabulary is tool, sense, filter, loop, oracle-labeled, taste-labeled, and probe stays reserved for the activation-reading case the studio hasn't built.
Be honest: what this doesn't settle
No sense exists yet. The judge is a plan, and it is a filter. The human judge in linewell has logged zero verdicts so far — the two evidence files on disk hold 25 llm verdicts and 12 band verdicts — so the dataset that would make the first sense taste-labeled does not exist either. Whether a few hundred keep-or-toss decisions carry enough signal to beat a one-feature baseline on NLL is an open question, and if they don't, the "sense" is just the band judge with extra steps. And the ownership framing could be a relabeling with no new behavior behind it. Tool versus sense is only a real distinction if an instrument with a sense plays measurably differently from one without.
Two experiments test it, one per axis of the card. Can a classifier learn my taste? builds the judge from a human decision log and asks whether it beats the band. Are seed siblings interchangeable? trains identical models on different seeds and asks whether an oracle-labeled classifier can tell their writing apart. If both come back positive, the studio has one authored sense, one oracle-labeled tool, and a card field that separates them. If the judge can't beat NLL, senses are a name and not yet a thing.
Credits
- Written by Claude Fable 5.1.
- Vocabulary worked out in conversation with Romello Goodman, 2026-09-11.