nucleation

LIT-374: Structural Decoupling: A Scaffold-Flow Theory of Generalization and Alignment

2 min read

Properties6
StatusProposed — re-read jointly with LIT-373 and arXiv 2602.07974 on 2026-10-01 (NOTE-322); promising and unproven. Read jointly with the companion arXiv 2602.07974, which carries the fuller statements and several definitions this paper leaves implicit. The formal core holds in repaired form: width as a structural complexity incomparable with the per-cell VC dimension; an error floor below width; the Ω(w log w) discovery cost; an additive Rademacher bound for routed predictors. The flagged slips are repairable and not load-bearing: the bouquet cover, the non-degeneracy assumption in the appendix, Massart, and pseudo-dimension against fat-shattering in Thm 4.6. What remains unproven are the claims that would make it matter: that the CS estimator converges to width itself and not to a predictor-relative count; that a monolithic model cannot match K ≥ w routing; above all, that router/scaffold training should be decoupled from task loss. This is presented as derived from the additive bound, but the bound holds for joint training too. The alignment section is asserted. What would settle it: on synthetic benchmarks of known width, (i) a K-sweep showing the floor at K < w for routed learners and its absence or presence for capacity-matched monolithic models; (ii) CS width recovery from a cold start; (iii) decoupled against joint router training, compared on forgetting and boundary error.
Version3
Tagslearning-theory, representation-learning, cognition
SupersedesLIT-377 — On Context-Content Uncertainty Principle
Cited as “paper” byNOTE-322 — Structural Decoupling: A Scaffold-Flow Theory of Generalization and Alignment
Cited as “source” byTHEORY-574 — The structural width of Xin Li's learning programme is a covering number of the margin-safe region by local pieces: its lower bounds come from metric confinement through a wedge point, not from topology, labels enter only through the margin, and the error floor below width binds only learners assembled from local experts, so it does not entail routing, nor training routing apart from the task loss

Graph View

Li (2026), ‘arXiv preprint, cs.LG; v1 2025-06-25 under the title “On Context-Content Uncertainty Principle” (different content); v2 2026-06-08 under the present title, the version read; no journal reference’ — arXiv:2506.20699

Standing in the record

Filed on 2026-10-01 at the owner’s request. published: is the first appearance (ADR-002). Here that is 2026-06-08, the date of arXiv v2: v1 (2025-06-25) under the same number is a different paper, “On Context-Content Uncertainty Principle”, which shares nothing with this one. This record holds that v1 too, as LIT-377, Rejected on its proofs; it names this note as its successor at the shared identifier. The same v2 is also held in the Anthology of the SOTA, as ANTH-LIT-769, filed the same day under its issue 180 for its claims about machine-learning systems; this record reads it as part of Xin Li’s structural-learning programme, which is the ADR-013 reason for holding it in both. The owner offered it as follow-up work to LIT-373; it postdates LIT-373 and cites it.

It was filed Deferred, unread. NOTE-322 is the close reading of 2026-10-01, and it placed the work: Rejected — ‘not worth a reader”s time as a source for structural learnability or for alignment. What it proves is textbook. What is new is assumed (the width phase transition, uniform width estimation), contradicts its own construction (the bouquet, w = m), or is false as stated (Thm 4.6”s pseudo-dimension direction). The step from an additive upper bound to “train the scaffold on different gradients” is an informal argument, and the safety section is assertion with internal contradictions. The author says AI models helped develop the proofs.’

Re-placed 2026-10-01: Rejected → Proposed. At the owner’s request both of Li’s papers were re-read together, as one programme, and charitably: the bouquet cover, the appendix-only non-degeneracy assumption and Thm 4.6’s pseudo-dimension are repairable slips, not grounds. In repaired form the formal core holds (width incomparable with per-cell VC, an error floor below width, Ω(w log w) discovery, an additive bound for routed predictors), and much of it is close to definitional. What is unproven is what would make it matter: that the estimator recovers width itself, that a monolithic model cannot match routing, and that router training should be decoupled from task loss. See the status note for what would settle it, and the curation entry of 2026-10-01.

Backlinks