The case
Two established practices recover latent organization from a matrix of relations, with no feature of any row or column given in advance.
- Collaborative filtering. A user–item matrix is fitted as R ≈ PQᵀ. A207 §1, with its displayed formulas transcribed: “For any invertible matrix A, define p’_u = Aᵀp_u, q’_i = A⁻¹q_i. Then (p’_u)ᵀq’_i = p_uᵀq_i. Thus an entire family of distinct latent-coordinate systems predicts exactly the same relationship matrix.” A218 §4.2: collaborative filtering “recovers predictive relational profiles of users and items without assigning unique ontological meaning to particular latent coordinates.” The standard reference is Koren, Bell and Volinsky 2009 (doi:10.1109/MC.2009.263), which A207 names.
- Topic models. A document–term matrix is fitted as X ≈ WH with nonnegative factors (NMF), or by latent Dirichlet allocation (ANTH-LIT-592). A218 §4.1: “The reconstruction is not uniquely determined in general.” And: “Under appropriate separability, support, or generative assumptions, topic components can become identifiable.” The works behind this are Lee and Seung 1999 (doi:10.1038/44565); Donoho and Stodden 2003 (url:https://proceedings.neurips.cc/paper/2003/hash/1843e35d41ccf6e63273495ba42df3c1-Abstract.html), whose uniqueness result needs separability and complete factorial sampling; and Arora, Ge and Moitra 2012 (arXiv:1204.1956), whose provable recovery needs an anchor word for every topic. None of the three is held in either record. Each is a methods work an anthology topic could hold.
The record already held readings that bear on the case, and the exchange cited none of them. Zhang and Blei (LIT-677) claim, without reported results, that LDA’s optima are joined by flat paths, a practical form of topic-model non-identifiability. Word embeddings learned contrastively factorize a co-occurrence matrix’s deviation from independence (THEORY-182, LIT-855; also LIT-863), and QUESTION-025 asks what such factorizations do with correlated attributes. The nearest precedent for a human test of recovered structure is Chang et al. 2009, “Reading Tea Leaves” (ANTH-LIT-597), in which human judges detect intruding words and topics, and better held-out likelihood goes with less interpretable topics.
The laboratory proposed
The owner proposed the case as a method at U57: “I think collaborative filtering via matrix factorization might give us a useful laboratory here”. A207 built it out. It is a design, not a study, and every number in it is invented: its 4×4 matrix of jokes and rebukes is labelled “Hypothetical ratings, not empirical data”.
A207 §4 replaces preferences by communicative judgements in a tensor Y_uiac, where u is an interpreter, i a communicative realization, a a judgement task “such as teasing, condemnation, solidarity, authority”, and c a presentation or framing condition. It proposes three experiments.
- Recognition without surface similarity. Realizations that perform one act in different wording, and realizations in near-identical wording that perform different acts. The hypothesis is that similarity of response profiles, Similarity(R_·i, R_·j), predicts shared-kind judgements “even after controlling for lexical or embedding similarity”. It is to be scored on “new realizations and held-out interpreters”, as CLAIM-011 requires.
- Gauge transformation. Transform the fitted factors by invertible changes that preserve the predicted matrix, and “Ask which proposed signatures remain stable.”
- Intervention. Estimate P(Y_uia | do(c)), keeping apart “actual situation changes from information supplied to the interpreter”. In a vignette design the second is all that is available (CLAIM-127, CLAIM-130).
A207 §5 adds the design rule that matters most. Observational ratings are missing for non-random reasons, and “Missing-not-at-random exposure can make learned factors reflect selection patterns rather than preferences or communicative distinctions.” The primary experiment is therefore to assign realizations to interpreters at random and record nonresponse separately.
A207 §6 also calls a held-out comparison of decision risk between two encodings “a direct Blackwell experiment”. By the record’s reading it is not one. It compares a chosen family of losses, which THEORY-176 shows is not Blackwell’s order (THEORY-156). A207 hedges (“The exact relationship to Blackwell deficiency depends on which decision classes and statistical experiments are included”) but keeps the heading.
What it can show
That restricted relational data support the recovery of latent structure, and that the recovery has stated limits. A218 §4.3 gives both halves: “Matrix factorization already demonstrates that restricted relational data can support latent structure discovery.” And: “Neither result alone demonstrates that recovered components are constitutive kinds, that their identities survive translation, or that they retain information for independent communicative decisions.”
So the case is evidence for what a factorization identifies and up to what (CLAIM-148), for the prior-art boundary (CLAIM-146), and for the gap between identification and constitution (CLAIM-135). It is evidence for A218’s Recovery step, not for its Ontology step.
What it does not show
- It does not show that a relational profile is the act. A207 §2: “If all interpreters share the same misconception, matrix factorization may recover a highly stable representation of that misconception.”
- It does not show that any topic is a real kind. A210’s own table answers “Every recovered topic is an objectively real kind” with “No”.
- It does not bear on contextuality. A207 §7: “A low-rank factorization succeeding or failing does not establish contextuality” (CLAIM-041).
- Experiment 2 is not an experiment. Which quantities survive a change of gauge is settled by algebra before any data arrive: exactly the functions of the predicted matrix (CLAIM-148). Only its second half, whether different sparse completions support conflicting classifications, is empirical.
- No part of the laboratory has been run.