ADR-060: concept-geometry

Context

Filing Park, Choe and Veitch’s Linear Representation Hypothesis (LIT-606), the owner asked for a concept-geometry topic alongside it. The record had already said it needed one, three times. LIT-460 said representation-and-encoding “holds fourteen notes, all about tokenisation and positional encoding. Nothing on concept geometry”. LIT-458 said the record “has no document for it, nor for steering vectors, nor for concept geometry”. THEORY-034 called the missing hypothesis a trunk.

Those documents were filed under representation-and-encoding, whose blurb is “how the signal is encoded before the expensive network sees it — tokenizers and learned latents, positional encoding”. A concept lattice over the residual stream is not that. It was the nearest wrong word, which under ADR-059 means the axis was short.

The documents on this subject had spread since then. Sparse-autoencoder and probing work (LIT-568, LIT-571 and three practices) sits in analysis-and-evaluation alone. Superposition (LIT-544) sits in four topics, none of them this one. The word-vector papers just filed (LIT-602 to LIT-604) introduced the analogy offsets this subject begins from.

Decision

A twenty-second topic, primary-eligible:

concept-geometry — how concepts are laid out in a model’s representation space: linear directions and the regions, lattices and spectra built from them, superposition and the dictionaries that pull it apart, reading concepts out with probes and writing them in with steering vectors, and whether different models arrive at the same geometry.

Primary for four documents whose subject it is and whose current primary is the misfit: LIT-460, THEORY-034, LIT-526, THEORY-063. They keep their other tags. This follows ADR-059, which moved LIT-505 and LIT-506 to the new word as primary.

Secondary for twelve, appended so their primaries stand: LIT-458 and THEORY-036 (convergent representations), LIT-544 (superposition), LIT-568 and LIT-571 (sparse autoencoders and sparse probing), SOTA-342, SOTA-344 and SOTA-345 (SAE training and evaluation, probing), LIT-602, LIT-603 and LIT-604 (the word-vector offsets), THEORY-089 (why those offsets exist). The test is ADR-046’s: would someone browsing concept geometry expect it?

Filed with it: LIT-606, its note, THEORY-090 and SOTA-378, all under the new word first.

Alternatives considered

• Widen representation-and-encoding. Its subject is what happens to the signal before the network. This subject is what the network builds from it. One word for both would make the input-encoding papers and the concept papers hard to find apart, which is the problem ADR-052 fixed for multimodal-learning.
• interpretability. analysis-and-evaluation’s blurb already claims interpretability, and the word names a purpose. The logit lens, causal tracing and knowledge editing are interpretability without being claims about how concepts are laid out, and they would all move. The subject here is narrower and is a claim about the model, not a method of checking it.
• Leave it to the trunk to arrive. The three documents that named the gap were each filed under a topic their subject did not fit. Waiting is how the scatter above happened.

Consequences

• Twenty-two topics, twenty primary-eligible. CLAUDE.md and luria.yaml change their counts.
• Not retagged: the logit and tuned lenses (LIT-569, LIT-570, SOTA-343), the FFN key-value and causal-tracing papers (LIT-575, LIT-578 and their theories), and LoReFT (SOTA-329). They read or edit hidden states without claiming how concepts are arranged in them. The first to make such a claim should take the tag.
• Relations among the new and retagged documents bind on the new tag: LIT-460 extends LIT-606, LIT-526 compared_against it, and THEORY-034 extends THEORY-090.