NOTE-322: Structural Decoupling: A Scaffold-Flow Theory of Generalization and Alignment

Joint re-reading (2026-10-01)

This section supersedes the assessment below. At the owner’s request this paper was re-read together with LIT-373 and the author’s companion preprint (arXiv 2602.07974), as one research programme, and charitably: a technical slip counts only if a major takeaway depends on it and its sensible repair does not save it. The first, isolated reading below rejected on slips that are repairable; its mathematical observations stand as observations, but not as grounds. The joint verdicts, takeaway by takeaway, are in the curation entry of 2026-10-01.

Li’s structural learning theory (StrLT, developed in the companion arXiv 2602.07974) splits multi-regime learning into discovering contexts (the “trap”, or scaffold) and predicting within one (the “funnel”, or flow).

• Width is the least number of cells on which a γ-contractive, low-risk predictor suffices. It is incomparable with the per-cell VC dimension, forces an error floor when fewer than w cells are allocated, and costs Ω(w log w) samples to discover. These results hold in repaired form, and are close to definitional.
• The CS estimator reweights a spatial graph by prediction agreement. It is shown to count components for a given predictor, not width itself.
• The decoupling prescription says to train routing on structural signals, not task loss, and to freeze consolidated contexts. It is a plausible, testable design hypothesis with continual-learning precedent, but the additive Rademacher bound it is said to follow from holds for joint training too.
• The safety section recasts alignment as scaffold alignment by analogy, and is asserted.

Contribution

The paper names a second complexity axis for learning in multi-context environments. VC/Rademacher theory governs prediction inside a fixed regime, the “funnel”. A separate quantity, width, governs how many regimes (“cells”, “basins”, the “trap”) must be found (Def 3.1). It summarises results for width: incomparability with VC dimension (Thm 3.2), an error floor when fewer than w cells are used (Thm 3.3), a coupon-collector discovery cost (Thm 3.4), a task-reweighted graph Laplacian whose spectrum is meant to count cells (Def 3.5, Thm 3.6), and consistency of penalised model selection over the number of cells (Thm 3.7). It then states a “structural fundamental theorem” (Thm 4.6) and an additive Rademacher bound for routed predictors (Thm 4.5). From these it argues for an architecture in which routing (the “scaffold”) gets no task-loss gradient (Def 4.10), and for reading hallucination, reward hacking, deceptive alignment and corrigibility as scaffold failures (§5). There are no experiments. Proofs are in an appendix that, by the author’s statement, ChatGPT 5.5 and Claude 4.7 helped develop (p. 1 footnote).

Key insight

Generalisation error for a routed predictor splits into a routing term and a within-route term. Thm 4.5 bounds the class complexity by a sum, √(2 log Δ_str(n)/n) + LK·R̂_n(G). The paper reads that additivity as a design rule: two terms that enter a bound separately should be reduced by separate mechanisms trained on separate signals (Prop 4.7, §4.2). Everything in §§4.2–5 rests on that step from an upper bound to a training prescription, and the paper gives no argument for it beyond the analogy.

Assumptions

• Setting (§3.1). (X, d_X) is a compact metric space, P a distribution on X × Y and G a class of local predictors. A cell U is (γ, δ)-feasible if some g ∈ G is γ-contractive on U and has conditional risk ≤ δ on U. Width is w(P; γ, δ) = min{K : some open cover {U_1, …, U_K} of X has every U_k feasible} (Def 3.1). “Contractive” for a map X → Y means Lipschitz constant γ < 1 between different spaces (Thm 3.8 proof).
• Thm 3.3 (phase transition). The proof (App. .2) adds a “standard structural non-degeneracy condition”. For every K < w, any K-cell cover has a cell that mixes incompatible basins, and every such cell has population excess risk ≥ η(w, K) > 0. The theorem as stated in the body (p. 4) carries no such hypothesis.
• Thm 3.4. Worst case: w basins of mass 1/w each, and a basin cannot be certified before it is sampled.
• Thm 3.6. Within-basin prediction variation is ≤ δ_y. Across basins joined by a spatial edge (d_X ≤ r_x) the gap is ≥ Δ_y, for the fixed predictor G used to build the kernel. Within-block conductance is ≥ φ_in and cross-block conductance ≤ φ_out. The perturbation bound ‖L_CS − L_0‖_op ≤ C_1·exp(−(Δ_y² − δ_y²)/σ_y²)·φ_out holds “under standard degree-conditioning assumptions” (App. .4), which are not stated.
• Thm 3.7. (1) A gap below width: R_K ≥ R_w + η_K for K < w. (2) No gain above width: R_K = Rw for K ≥ w. (3) Uniform convergence |R̂{n,K} − R_K| ≤ r_n(K) → 0 a.s. (4) A penalty with pen_n(K) → 0 and pen_n(K) − pen_n(w) ≫ r_n(K) + r_n(w) for K > w. Here w is the risk elbow of R_K. Nothing shows it equals the contraction-and-risk width of Def 3.1.
• Thm 3.10. Y = f*(X) + ξ with E[ξ|X] = 0. f* is β_k-Lipschitz and g_k γ-Lipschitz on U_k, |g_k − f*| ≤ ε_k at one anchor point, and ℓ(u, y) = φ(u − y) with φ convex, φ(0) = 0 and L_φ-Lipschitz.
• Thms 4.5–4.6. Loss bounded in [0, 1] and L-Lipschitz, K fixed, “loss separation on graph-shattered sets” (Thm 4.6, not defined beyond one sentence in App. .10). The class H_{K,δ}(G) is defined as the assignments h for which some g_{1:K} is “δ-risk-feasible” (Def 4.1). That is a population-risk condition, so the hypothesis class depends on P.
• Prop 4.11. d_part is “define[d] … so that P(E_scaf) ≤ d_part” (App. .12).
• §5. Premises about alignment methods, deceptive alignment and corrigibility are asserted with citations to the safety literature ([24]–[27], [47]–[50], [58], [59]). No model of any of them is given.

Key results

Results in the paper’s numbering, with what the appendix actually shows.

• Thm 3.2 (VC–width separation). There are families with w(P) → ∞ and VC(G) = O(1), and families with VC(G) → ∞ and w = 1.

  • Second direction. It is proved, and it is trivial. On [0, 1], f*(x) = λx with λ < γ lies in every polynomial class G_d, so w = 1 while VC(G_d) ≥ d + 1.
  • First direction. The bouquet construction, as written, contradicts itself. The problem P_m is stipulated to make infeasible any set with points of two branches “arbitrarily close to x₀”. In a bouquet every open set containing the basepoint x₀ contains initial arcs of all m branches. So no feasible open cover exists, and w = min ∅ = ∞ for every m ≥ 2. The claimed upper bound w ≤ m (“one thickened neighborhood for each branch”) is exactly such a cover, and it violates the stipulation. The conclusion “w → ∞” survives only degenerately. The proof’s “w(P_m) = m” is false under its own construction. A measurable partition in place of an open cover would repair it.

• Thm 3.3 (phase transition at width). If K ≥ w, the problem “decomposes into K ordinary within-cell problems”. If K < w, Gap(K) ≥ η(w, K) > 0 independent of n. The K < w half is the non-degeneracy hypothesis restated (Assumptions). Infeasibility in Def 3.1 can come from the contraction condition alone with risk ≤ δ, so width falling short does not by itself imply a risk floor. The K ≥ w half is proved only “conditional on using such a feasible structural allocation”, that is, given the cover.

• Thm 3.4 (structural sample complexity). Ω(w log w) samples to see all w basins, by coupon collector (E[T] = wH_w). It is correct under its worst-case assumptions and standard. The headline n_total ≈ Ω(w log w) + Σ_c O(p_c/ε²) adds a lower bound to upper bounds, so it is a heuristic, not a theorem.

• Thm 3.6 (spectral gap amplification). The ratio of cross-basin to within-basin CS weights is ≤ exp(−(Δ_y² − δ_y²)/σ_y²). That is immediate from Def 3.5. The gap bound is g_eff ≥ c_1·φ_in²/w⁴ − C_1·exp(−(Δ_y² − δ_y²)/σ_y²)·φ_out, by Weyl’s inequality applied to λ_{w+1}. Three gaps in it:

  • g_eff is not defined in the statement.
  • The w⁴ comes from a cited “higher-order Cheeger-type estimate”. For a block-diagonal L_0, ordinary Cheeger per block already gives λ_{w+1}(L_0) ≥ φ_in²/2, so the factor is unnecessary, though the bound is not false.
  • The operator-norm perturbation bound, which carries the whole result, is asserted.

  No statement says that the first w eigenvalues of L_CS are near zero, or that the count of near-zero eigenvalues equals w(P; γ, δ). The “estimator of width” is therefore not proved to estimate width. It counts prediction-value discontinuities of a fixed predictor G, which presupposes a G whose jumps already mark the basins.

• Split–merge (p. 5). “V(Π) = aI_tot(Π) + bK(Π) strictly decreases … yielding finite convergence to the correct width.” I_tot is never defined, and there is no theorem or proof.

• Thm 3.7 (width consistency by penalised structural ERM). K̂_n → w a.s. The proof is the standard penalised model-selection argument and is correct given assumptions (1)–(4). Assumption (3), uniform convergence, is the “uniform width-estimation guarantee” that contribution 2 and §7 claim as a result. §6.1 concedes that calibrating η* “requires concentration results beyond those currently proved”.

• Thm 3.8 (contraction transfer through the slingshot). F_c = π_c ∘ G⁰_c ∘ φ has Lipschitz constant ≤ L_{π,c}·γ_Z·L_φ(U). This is composition of Lipschitz maps, correct and elementary. It is unrelated to LIT-373’s “metric slingshot”, a metric warp that collapses distance along an escape route.

• Thm 3.10 (contraction-to-risk transfer). R_k ≤ L_φ(ε_k + (γ + β_k)·diam(U_k)) + E[φ(ξ) | h(X) = k]. Correct by the triangle inequality. The paper itself notes the swap of φ(−ξ) for φ(ξ). Convexity of φ and E[ξ|X] = 0 are assumed but unused. The converse example, g_m = m⁻¹ sin(m²x), is correct.

• Thm 4.3 (structural Sauer–Shelah). Δ_str(n) ≤ Σ_{j≤d} C(n,j)(K−1)^j ≤ (e(K−1)n/d)^d for graph dimension d. The first inequality is invoked as “the standard multiclass Sauer–Shelah bound for graph dimension”, with no citation. Its attribution was not verified here. The second is a correct binomial estimate.

• Thm 4.5 (structural complexity decomposition). R̂^S_n(F_{K,δ}) ≤ √(2 log Δ_str(n)/n) + LK·R̂_n(G). For fixed assignments the per-cell suprema separate, and Talagrand contraction plus zero-padding gives LK·R̂_n(G). The routing term is attributed to “Massart’s finite-class lemma”, which does not apply directly to a supremum over assignment vectors of random suprema. The bound holds, with these constants, by a bounded-differences / sub-Gaussian maximal inequality: each fixed-assignment supremum has differences ≤ 2/n per σ_i. So the stated result is right and its proof has a fixable gap. The factor K multiplies the funnel term, so the “no multiplicative cross-term” claim holds only with K fixed. App. .11 concedes “except through the fixed multiplier K”.

• Thm 4.6 (fixed-class structural fundamental theorem). It claims equivalence of (1) d_str < ∞ and Pdim(G) < ∞, (2) R^S_n → 0, (3) uniform Glivenko–Cantelli, and (4) distribution-free structural PAC learnability. (4) ⇒ (1) is false as stated. Take K = 1 and G the 1-Lipschitz maps [0, 1] → [0, 1] (the kind of class the paper’s contraction condition picks out). G has infinite pseudo-dimension: on the points i/n with every threshold ½, the values ½ ± 1/(4n) realise every pattern 1-Lipschitzly. Yet G is uniformly Glivenko–Cantelli and learnable under bounded Lipschitz loss. Learnability of real-valued classes is characterised by fat-shattering dimension at every scale (Alon, Ben-David, Cesa-Bianchi & Haussler 1997), not by pseudo-dimension. The proof’s step “If Pdim(G) = ∞ … not distribution-free PAC learnable” (App. .10) is the error. Separately, “distribution-free” learnability of H_{K,δ}(G) is ill-posed, since Def 4.1 defines that class through risk under P. The (1) ⇒ (2) ⇒ (3) ⇒ (4) chain is standard and correct.

• Prop 4.7 (structural decoupling principle). Reducing d_str affects only the trap term of the upper bound, and reducing R_n(G) only the funnel term. This is true of the bound, and the proof says so (“at the level of generalization bounds”). §4.2 then uses it to conclude that task-loss gradients should not modify the trap, which is an informal argument.

• Prop 4.11 (decoupled decomposition of errors). R(F_S) − R* ≤ d_part(Π_S, Π*) + Σ_c P(U_c)(R_c(F_c) − R*_c). d_part is defined so that this holds, and the claim that off E_scaf the remainder is “exactly” the weighted within-cell excess is asserted.

• Prop 4.13 (no backdoor interference). A flow update in context c cannot change routing or consolidated parameters of c′ ≠ c. True by definition: Def 4.10 sets ∇_{θ_S} L_flow = 0, and consolidated experts are frozen by assumption.

• Condensation protocol (§4.2). A candidate scaffold token is created when the “single-basin condition” fails. It is consolidated, and thereafter removed from task-loss training, when three tests agree: stability, coherence (internally contractive or CS-connected) and separation (merging would violate the structural gap). It is a protocol description, with no algorithm, thresholds or guarantee.

• §5 (safety). Five items:

  • deceptive alignment as a “structural identifiability problem”, with grokking offered as “an empirical existence proof for timescale separation”;
  • reward models as “closure gates”, and a “multi-gate condensation protocol”;
  • “anesthesia-mode” deployment, which may navigate the scaffold but not consolidate new structure;
  • hallucination and misalignment as the same boundary-resolution failure, so hallucination rate on boundary-like factual tasks would proxy alignment robustness;
  • scaffold preservation as an instrumental incentive against corrigibility.

  The section concludes that “the alignment problem cannot be solved at inference time” (p. 10). All of this is asserted.

• §6. Proposed research directions: online width estimation; “two independent scaling exponents” predicted from Thm 4.5’s additive bound; benchmarks with known w and η*; a causal reading of width via invariant risk minimisation; and structural certification that K ≥ w.

Claims

id	claim	strength	support
C1	Width and VC dimension are incomparable (Thm 3.2)	proof; one direction trivial, the other self-contradictory as written	App. .1. With an open cover and the stipulated cross-branch incompatibility, the bouquet has no feasible cover (w = ∞), not w = m
C2	Below the true width every K-cell learner has an error floor η(w, K) > 0 independent of n, which is the “phase transition” (Thm 3.3)	assumed	App. .2 adds a non-degeneracy hypothesis that is the conclusion. The body states the theorem without it
C3	Discovering all w basins needs Ω(w log w) samples in the worst case (Thm 3.4)	proof (standard coupon collector)	App. .3; uniform basin masses
C4	The CS Laplacian’s near-zero eigenvalues count task-compatible components and so estimate width (Def 3.5, Thm 3.6)	informal argument	The weight-ratio bound is immediate. The gap bound rests on an asserted perturbation norm. No theorem relates the eigenvalue count to w(P; γ, δ)
C5	Split–merge with V(Π) = aI_tot + bK converges in finitely many steps to the correct width (p. 5)	assertion	I_tot is undefined. No statement or proof
C6	Penalised structural ERM selects K̂_n → w a.s. (Thm 3.7)	proof, given assumptions that include uniform convergence	App. .5. Standard model-selection argument
C7	The CS operator gives “uniform width-estimation guarantees” (contribution 2, §7)	assertion	Uniform convergence is a hypothesis of Thm 3.7, and §6.1 says the needed concentration results are not proved
C8	A latent “slingshot” transfers contraction: Lip(F_c) ≤ L_π γ_Z L_φ (Thm 3.8)	proof (elementary)	App. .6
C9	Contraction plus an anchor bound gives a risk bound on each cell (Thm 3.10)	proof	App. .7, with a noted sign swap
C10	The structural Rademacher complexity is at most a trap term plus LK·R_n(G), with no multiplicative cross-term (Thm 4.5)	proof with a fixable gap; the claim holds with K fixed	App. .9. Massart is misapplied, but a sub-Gaussian maximal inequality gives the same bound. K multiplies the funnel term
C11	Finite d_str and Pdim(G) are necessary and sufficient for structural learnability (Thm 4.6)	proof with a false step; the claim is false as stated	App. .10, (4) ⇒ (1). 1-Lipschitz classes have infinite pseudo-dimension and are learnable. The class is defined through P, so “distribution-free” is ill-posed
C12	Scaffold and flow complexities can be controlled independently (Prop 4.7)	proof, of the upper bound only	App. .11
C13	Therefore the scaffold should receive no task-loss gradient, and a scaffold–flow architecture “separates alignment and generalization” (§4.2, abstract)	informal argument	An inference from the additivity of an upper bound to a training prescription. No experiment
C14	A flow update cannot change other consolidated contexts under decoupling (Prop 4.13)	tautology	It follows from Def 4.10 and the freezing assumption
C15	Grokking is the within-cell (w = 1) sibling of the structural phase transition, “regularisation-driven”, and an “empirical existence proof” of the timescale separation that deceptive alignment needs (§3.1, §5.1)	analogy	Cites [38]–[41]. The founding grokking run used no weight decay (record, below). Sample counts (Ω(w log w)) are conflated with optimisation steps
C16	Hallucination and misalignment share one structural cause, boundary-resolution failure, so boundary hallucination rate proxies alignment robustness (§5.4)	assertion	No model or data
C17	Alignment failures are primarily scaffold failures, and alignment “cannot be solved at inference time” (§5 summary, §7)	assertion	It contradicts §5’s own classification (see Limitations)
C18	Structural scaling should show two independent exponents (§6.2)	prediction, untested	Read off the additive upper bound (Thm 4.5). A bound’s form does not fix the error’s

Method

The paper has two layers.

1. Learning-theoretic. It defines width (Def 3.1) and the structural assignment and loss classes (Defs 4.1–4.4). It proves results by standard tools: coupon collector, the Lipschitz triangle inequality, multiclass Sauer–Shelah, Talagrand contraction plus a finite-class maximal inequality, a penalised-ERM elbow argument, and Weyl perturbation with a Cheeger bound.
2. Architectural. It defines a scaffold S = (C, Σ, M, A): contexts, a “topological indexer” Σ: X → Δ(C), a metric library of local experts (U_c, G_c, π_c), and admissible split/merge/archive/consolidate operations (Def 4.8). Flows are F_S(x) = π_c ∘ G_c ∘ φ(x) with c = Σ(x) (Def 4.9). Architectural decoupling means ∇{θ_S} L_flow = 0 and ∇{θ_F} L_scaffold = 0 (Def 4.10). The protocol alternates flow update, scaffold test (mismatch, CS width, merge criteria) and a certified scaffold update (Def 4.12). Promotion to scaffold is by the three-test condensation protocol.

No algorithm is specified beyond this, no loss L_scaffold is defined, and nothing is implemented.

Concepts

• Trap / funnel. The structural problem of identifying the context, against the metric problem of predicting within it (Fig. 1, §1). In LIT-373 a “topological trap” is instead a region “where local gradients vanish or point toward dead ends” (Fig. 1 caption).
• (γ, δ)-feasible cell. An open set on which some g ∈ G is γ-contractive and has conditional risk ≤ δ (Def 3.1). In Def 3.9 the “jointly feasible” version applies to the cells h⁻¹(k) of an assignment, which need not be open.
• Width w(P; γ, δ). The least size of a feasible open cover (Def 3.1).
• Structural error floor η(w, K). The population excess risk of any K < w cover. It is assumed (Thm 3.3 proof).
• Contractive-similarity (CS) kernel. W_ij = 1[d(x_i, x_j) ≤ r_x]·exp(−|G(x_i) − G(x_j)|²/σ_y²), a geometric adjacency reweighted by prediction agreement. Its normalised Laplacian is L_CS = I − D^{−1/2} W D^{−1/2} (Def 3.5). The paper compares it to bilateral filtering.
• Metric slingshot. A composition π_c ∘ G⁰_c ∘ φ through a latent space with pre-built contractions (§3.3). In LIT-373 the slingshot is a warp of the internal metric.
• Structural assignment class H_{K,δ}(G), structural graph dimension d_str, structural growth function Δ_str. Defs 4.1–4.2. The multiclass graph dimension of the routing maps.
• Scaffold / flow. The slow structural object (context library, routing, boundaries) against the fast within-context computation (§4.2, Defs 4.8–4.9).
• Condensation. “transient flow pattern → stable scaffold token” under the stability, coherence and separation tests (§4.2). It descends from LIT-373’s “recursive condensation” in name only: there is no topology, quotient map or Urysohn step.
• Scaffold alignment / flow alignment. Def 5.1. Flow-aligned means acceptable outputs on a distribution. Scaffold-aligned means the scaffold represents the needed contexts and boundaries at sufficient resolution.
• Anesthesia-mode deployment. Navigate but do not consolidate (§5.3).
• E-D-T cycle (§6.3) and I_tot (p. 5). Used but never defined.

Connections

• LIT-373 / NOTE-321 (Li 2026, “The two dragons of cognition”). v2 cites it as [21] for the scaffold–flow view (p. 2). The influence runs from LIT-373 to this paper, not the reverse. v1 (2025) is the unrelated CCUP paper, and LIT-373 cites no Li preprint (NOTE-321). The shared words carry different mathematics:

  • “metric slingshot”: a metric warp in LIT-373, Lipschitz composition here (Thm 3.8);
  • “trap”: a gradient dead end in LIT-373, context identification here;
  • “phase transition”: maze-to-bowl collapse in LIT-373, an error floor at K < w here;
  • “condensation”: a quotient map in LIT-373, freezing a routed expert here.

  Everything NOTE-321 found misstated is absent here: Savitch, Urysohn, the linear readout, parity, the free-energy equivalence and R(N) ∝ b^{αN}. It is not defended and not repaired. The two papers fail in parallel ways. The headline claims outrun the body (“proves”, “uniform guarantees”). A key “theorem” assumes or misstates its content. Terms are used and never defined (GHL in LIT-373; E-D-T and I_tot here). The author discloses AI help with the theorems in both.

• THEORY-019 (symmetry forces degeneracy; degeneracy does not identify a symmetry). The CS estimator reads structure from eigenvalue multiplicity at zero. For a graph Laplacian that multiplicity counts connected components exactly, with no group involved. That is consistent with THEORY-019’s point that multiplicities carry dimension counts, not a group. THEORY-019 lists near-degeneracy in noisy operators, and the thresholds a test would need, as open. Thm 3.6 is a Weyl eigengap bound of that kind for a Laplacian, but its perturbation norm is asserted and it gives no threshold for the first w eigenvalues. It does not bear on THEORY-019’s promote_when, which asks for the real Schur corollary.

• THEORY-028 (information-theoretic generalisation bound, Active). Thm 4.5 is a uniform-convergence (Rademacher) bound over a hypothesis class. It says nothing about how a training algorithm uses the class. A claim about which gradients should update which parameters would need algorithm-dependent analysis of the kind THEORY-028 records (bias bounded by the information the output carries about the data), and the paper uses none. No bearing on THEORY-028.

• THEORY-039 (later training phases differ in kind) and THEORY-022 (the grokking circuit). §3.1 and §5.1 present grokking as “a sharp, late, regularisation-driven jump” and as the w = 1 instance of the width phase transition. THEORY-039 warns against identifying training phases across works by shape, and this paper’s identification is by “qualitative shape” only (p. 5). On “regularisation-driven”: the record’s reading of Power et al. (NOTE-287, cited in THEORY-022) has the founding §3.1 run use Adam with no weight decay, though weight decay was the most data-efficient intervention. LIT-369 reports anti-grokking at late times. The paper’s claim that grokking tasks have “intrinsic width w = 1” is asserted, not computed. No bearing on either promote_when.

• THEORY-035 (Rejected: SGD compression explains generalisation). “Condensation” here is a freezing protocol, not an information-compression claim. No bearing.

• THEORY-030 and THEORY-026 (thermodynamics of computation and memory). The paper has no thermodynamic content, so it bears on neither. The owner has noted that NOTE-321’s remark, that LIT-373’s maintenance cost “runs against” THEORY-030, is mistaken. A standing maintenance cost only makes the structure dissipative, which Landauer’s principle does not forbid. That remark sits in NOTE-321’s Connections and is not among LIT-373’s stated grounds, so correcting it does not move LIT-373’s status.

• LIT-227 and LIT-249 (Deferred; NOTE-209 and NOTE-227, both skims). Contrastive learning as spectral clustering on an augmentation graph, with Cheeger-type guarantees (HaoChen et al. Thm 3.8). The CS Laplacian is spectral clustering on a prediction-reweighted geometric graph, the same move with a different edge weight. The paper cites only von Luxburg’s tutorial and does not engage this literature.

• Not in the record. The companion preprint arXiv 2602.07974, “Structural Learning Theory: A Metric-Topology Factorization Approach” (Li, v1 2026-02-08, v2 2026-05-06), whose results this paper says it “briefly review[s]” but does not cite. It was not read. It may hold fuller statements of Thms 3.2–3.7.

Bearing on the record

• On LIT-373 (Rejected, NOTE-321). Ground by ground:

  • Savitch misstated (“memoized Savitch”): not mentioned, so left as it was.
  • Urysohn misstated, and the inferred linear readout: not mentioned, so left as it was. The paper’s separability machinery is now contraction plus prediction-gap spectral clustering, which tacitly abandons the Urysohn argument rather than repairing it.
  • Theorem 2, R(N) ∝ b^{αN}, unproved: no counterpart. Thm 3.4’s Ω(w log w) is a sample-complexity lower bound about a different quantity and does not prove, restate or reinterpret predictive reach.
  • No data or model: there is still no data. This paper supplies a formal model (Defs 4.8–4.12) of a different thing, routed context learning. It is not a model of any of LIT-373’s neural claims (columns, PNGs, SWRs, gamma), and it makes none of LIT-373’s §6.2 predictions discriminating.
  • The three internal inconsistencies (Definition 2 against Theorem 1, the three senses of “Space Dragon”, cortical growth linear and exponential): none of the objects recurs here, so they are left as they were.

  The paper’s new formal content (width, the error floor, the CS estimator) is not LIT-373’s and does not supply anything LIT-373 lacked on its own terms. Where it is proved it is standard, and where it is new it is assumed or flawed. LIT-373’s status should stay Rejected, on NOTE-321’s grounds minus the THEORY-030 remark, which was never one of them. The owner’s suggestion that this work might change the take on LIT-373 does not hold up. The strongest thing it shows is that the author’s later framework no longer relies on the Savitch and Urysohn pillars.

• On THEORY documents. It supports, contradicts or should produce none. It touches THEORY-019, -022, -028 and -039 only as described above.

• For ML practice (the Anthology of the SOTA). It carries an instruction: train routers and experts on separate signals, freeze consolidated experts, and evaluate alignment at “boundary” inputs. That instruction has no experiment behind it. The ingredients are old: MoE with frozen experts, progressive networks and CLS-style replay, all cited. Nothing here is evidence for them, so it is not an anthology candidate.

• For filing. learning-theory first: width, structural Rademacher complexity, learnability. representation-learning because the scaffold is a learned context representation and §5.1 reads interpretability as estimating its topology. cognition because complementary learning systems is the paper’s biological motivation (§2, Table 1). A browser of AI safety would want it, but the closed vocabulary has no such tag, and agency would be the nearest wrong word for deceptive alignment and corrigibility.

Limitations

• The headline outruns the body.
  • The abstract and §7 say width “is proved to be orthogonal to VC dimension”, but one witness contradicts its own construction.
  • They describe “a sharp phase transition at the critical width”, which is assumed.
  • They claim “uniform width-estimation guarantees”, which are a hypothesis of Thm 3.7.
  • They say the theory turns the trap–funnel picture “from an analogy into a mathematical pipeline” and gives “a precise architectural prescription”, but the step from bound to architecture is informal.
• A false theorem. Thm 4.6’s necessity direction is false for real-valued classes: pseudo-dimension is not the right dimension, fat-shattering is. The class it is stated for is defined through P, so “distribution-free” does not parse.
• Width is not connected to its estimator. Def 3.1 (contraction plus risk on an open cover), Thm 3.6 (prediction gaps of one fixed predictor) and Thm 3.7 (the elbow of R*_K) are three different quantities. No result says they coincide.
• Internal inconsistencies.
  • §5 places RLHF, constitutional AI, DPO and debate under scaffold alignment and calls flow alignment “largely unexplored” (p. 9). §4.2 says “behavioral alignment is primarily evidence about flows”, and the §5 summary says “Output-level alignment methods operate mainly on flows” (p. 10).
  • Def 5.1 defines flow alignment behaviourally, on a distribution. The next paragraph makes it “a generalization problem” about “novel situations”. §4.2 had assigned generalisation to the flow and alignment to the scaffold.
  • “alignment = scaffold alignment + flow alignment” (p. 9) and “Alignment = Scaffold Architecture × Condensation Protocol × Deployment Mode” (p. 10) are two incompatible decompositions.
  • “Five competing approaches to alignment” introduces five failure analyses, not approaches.
• Undefined terms. E-D-T cycle (§6.3), I_tot (p. 5), g_eff (Thm 3.6), the loss-separation assumption (Thm 4.6), the degree-conditioning assumptions (App. .4) and L_scaffold (Def 4.10).
• The grokking argument. It mixes sample complexity, Ω(w log w) samples, with optimisation time, steps on a fixed set. It calls grokking “regularisation-driven” against the founding experiment, and asserts w = 1 for modular arithmetic.
• No experiments, simulations or implementations. §6 lists what would test the theory: benchmarks with known w and η*, CS width on transformer activations, and the two scaling exponents.
• AI involvement in the formal content. By the author’s footnote, ChatGPT 5.5 and Claude 4.7 helped develop “theoretical ideas, mathematical proofs in the Appendix, and visual illustrations”. v1’s footnote says ChatGPT helped with the theoretical ideas.
• Version history. The arXiv identifier’s 2025 date belongs to a different paper (CCUP). A citation of “Li 2025, arXiv 2506.20699” for StrLT would be anachronistic.

Open questions

• With open covers replaced by measurable partitions, or with incompatibility only outside a small neighbourhood of the basepoint, does the bouquet give w(P_m) = m? Is there a natural condition on (P, G) under which infeasibility of every K < w cover implies a positive risk floor, rather than assuming one?
• Is there a theorem that the number of eigenvalues of L_CS below some explicit threshold equals w(P; γ, δ) with high probability, with the threshold depending on φ_in, φ_out, Δ_y, δ_y and n? This is the step the “width estimator” needs, and it is the noisy-near-degeneracy question THEORY-019 leaves open, specialised to Laplacians.
• Restated with fat-shattering dimension and a distribution-free definition of the assignment class, does the structural fundamental theorem hold? It probably reduces to known results for compositions of a multiclass router with real-valued experts.
• Does decoupling router and expert training (Def 4.10) beat joint training on a benchmark with known width, in forgetting and in boundary-case error? That is the one experiment that would give §§4.2–5 an empirical footing.
• Does the companion preprint, arXiv 2602.07974, contain the missing assumptions and proofs (Thm 3.3’s non-degeneracy, the perturbation bound, uniform width estimation)? A reading of it would settle whether the StrLT results are better than their summary here.

Corrections to the seeded skim

• none (there was no seed or dossier)
• ‘The batch instructions call width, the phase transition at the true width and the contractive-similarity spectral estimator “LIT-373”s”. They are not. LIT-373”s text contains no “width”, no VC dimension, no contractive-similarity operator and no Laplacian. Its “phase transition” is the maze-to-bowl metric collapse (§1.2) and the PNG-to-assembly change (§4.2). Neither is a statement about a number of cells. These objects first appear, in the record”s holdings, in this v2. The vocabulary the two share is “trap”, “metric slingshot”, “condensation”, “flow” and “amortized inference”. In each case the formal content differs (see Connections).’
• ‘arXiv 2506.20699 v1 (2025-06-25) is “On Context-Content Uncertainty Principle” (CCUP). It is an entropy-asymmetry framework, H(Φ) ≪ H(Ψ), with Theorems 1–8 on “Structure-before-Specificity”, variational preconditioning, bootstrapping and hierarchical composition. v2 replaced it wholesale under the same identifier with an unrelated paper. Neither version mentions Savitch, Urysohn, parity, cortical columns or linear readout. So the identifier”s 2025 date does not make this paper a precursor of LIT-373. v2 postdates LIT-373 (published 2026-03-23) and cites it as [21], the source of the “scaffold-flow view” (p. 2).’
• ‘v1”s abstract promises “computational simulations demonstrating the efficiency gains of CCUP-aligned inference”. v1 contains none. v1 Thm 1”s proof (App. D, Step 1) says H(Ψ) ≫ H(Φ) “implies H(Φ|Ψ) > H(Ψ|Φ)”. The identity H(Ψ|Φ) − H(Φ|Ψ) = H(Ψ) − H(Φ) gives the reverse. v1 §6 calls consciousness the state where H(Ψ|Φ) ≈ H(Φ|Ψ), which that identity rules out under the paper”s own assumption H(Φ) ≪ H(Ψ). The proofs of v1 Thms 6–7 update Ψ where the statements update Φ. These are recorded for completeness. v1 is not the version read for the claims below.’