Answers are grounded in this paper and cite their sources.
Nīlūra-Σ presents a self-contained mathematical derivation of index optimality for noisy unordered storage in the fixed finite discrete memoryless channel (DMC), Bernoulli single-draw setting of Shomorony–Heckel Conjecture 2. The release combines a converse valid for arbitrary joint input distributions with a matching index-based coding construction, explicit maximal-error bounds, and finite parameter conditions guaranteeing exponential reliability. The external target is Conjecture 2 in Ilan Shomorony and Reinhard Heckel’s Information-Theoretic Foundations of DNA Data Storage (2022), Section 7.1, printed page 85. The question is whether index-based coding is capacity-optimal when independently retained stored fragments are corrupted by memoryless noise and their order is lost. The manuscript treats the fixed finite-DMC model, including asymmetric channels, zero transition probabilities, repeated stored words, and correlated stored inputs. Model and exact capacity An encoder stores (M) words of length (L) over a fixed finite input alphabet. Each word is independently retained with probability (1-q), and every retained symbol passes independently through a known, fixed finite channel (W). The receiver observes only the unordered multiset of noisy words, including multiplicities. It receives neither the original word order nor the source indices. The encoder does not know the realized retention mask. When (L/log_2 M→β∈(0,∞)), the derived capacity, in bits per stored input symbol, is [Cshuffle=(1-q)[C(W)-1/β]_+,] where (C(W)) is the ordinary capacity of (W) and ([x]_+=max{x,0}). The result covers every fixed (q∈[0,1]) and (β>0), including the critical boundary and zero-capacity cases. Every smaller positive rate is attained by an indexed construction with vanishing maximal message error. The converse permits arbitrary global decoding and does not restrict encoders to independent or distinct stored words. For (q<1), a positive reliable storage rate is possible exactly when (β C(W)>1). The result therefore identifies the exact first-order cost of preserving logical position in this channel model: more elaborate joint matching and global decoding cannot exceed the indexed construction’s asymptotic capacity. Proof architecture and quantitative results The central converse is a finite information bound uniform over arbitrary joint stored-word distributions: [ I(X_1,…,X_m;T)[C(W)L-log_2m]_++K_Wm√log_2(m+1).] Here (T) is the unordered noisy output of (m) words before applying Bernoulli thinning. For (C(W)>0), (K_W) is a finite channel-dependent constant specified explicitly in the manuscript; the zero-capacity case is handled separately. The argument combines a capacity-achieving output reference, a smoothed empirical-mixture reverse-divergence bound, a comparison of sampling with and without replacement, and a prefix/suffix transfer charging remaining symbols at ordinary channel capacity. The revised mixture estimate has universal additive constant (K_0=1). An explicit conditional product-channel factorization justifies the sorted-suffix step, including for asymmetric channels and dependent stored inputs. Achievability uses individually decoded index–payload codewords and an outer error-and-erasure code. A self-contained random-coding and expurgation argument supplies the inner code. The outer analysis explicitly accounts for missing reads, failed inner decodings, incorrect index claims, and conflicting payload claims. For fixed construction parameters [0<θ<C(W)-1/β,0<δ<1-q,0<ε<C(W)/3,] satisfying (1/β+θ<C(W)-3ε), the manuscript gives quantitative inner and outer maximal-error bounds. In particular, when the inner maximal error satisfies (e_L≤δ/16), [Pe,max≤exp(-Mδ^2/2)+exp(-Mδ^2/128).] Manuscript equation (22) gives explicit finite conditions ensuring this premise and the required inner-code size. The exponential guarantee therefore has stated finite sufficient conditions rather than applying to arbitrary block lengths or parameter choices. Relevance to AI learning and recursive self-improvement The direct AI application is reliable retention of arbitrary records in noisy, unordered fixed-length chunks: serialized program state, checkpoints, episodic records, or other stored information whose logical positions matter. For a system whose future learning or self-improvement depends on reconstructing such records, the theorem supplies an exact information-retention boundary and a matching indexed storage strategy within the specified model. Above the capacity boundary, exact arbitrary-record reconstruction cannot have vanishing error. When (β C(W)≤1), every fixed positive record rate has reconstruction error tending to one. Below capacity, positive rates are supported by the construction. These statements identify a concrete memory constraint relevant to self-improving systems while leaving task-specific learning requirements separate. Some tasks need only a summary, a partial record, or permutation-invariant information. The result is an information-theoretic memory theorem. It establishes neither a measured neural training speedup nor a general autonomous RSI algorithm. No polynomial-time capacity-achieving implementation or optimized finite-length code is claimed. Release contents and verification The release includes the manuscript PDF, editable LaTeX source, two Python mathematical checkers, recorded JSON results, an analytical audit, source provenance, citation metadata, and SHA-256 integrity manifests. The Hugging Face distribution additionally contains a dataset card, 50 structured research records, and cross-platform dataset upload tools. These records index statements, finite checks, channel examples, audit findings, and scope; they are not an empirical model-training corpus or a neural benchmark. The standard-library verification suite uses exact rational probabilities and integer combinatorics, with 80-digit Decimal logarithms and comparison tolerance (10⁻⁶⁰) where logarithmic evaluations are needed. Reported checks include: 164,248 exact small-case error-and-erasure accounting cases. 44,132 exact rational negative-moment cases. 520 exact conditional-probability identities checked by a separate sorted-channel implementation. Seven finite channel examples covering symmetric, erasure, asymmetric, redundant-input, identity, and zero-capacity behavior. During packaging, the main checker reproduced its recorded JSON result excluding machine-dependent elapsed time, and the sorted-channel checker reproduced its recorded result exactly. These finite checks support reproducibility and help detect implementation or algebraic errors; the universal capacity statement rests on the manuscript’s analytical proof. Status, provenance, and scope The manuscript presents a complete derivation for the stated fixed finite-DMC, Bernoulli single-draw model. Separate internal mathematical audits identified no unresolved premise. Independent external refereeing and machine-checked formalization have not been performed, and worldwide priority is not certified. The source provenance identifies the supplied SÆLQR II research, release 2.0, Sections 3–4, 17.2, and 18.3, as a basis for the central converse and its stated external consequence. This release isolates and audits that consequence, provides the complete operational construction, and records the revised quantitative reliability guarantees. Detailed attribution accompanies the manuscript. Channels with memory, insertion/deletion noise, unknown changing channel laws, and repeated-read clustering are outside the theorem. A strong converse above a positive capacity is not asserted; the error-tending-to-one conclusion stated here concerns the zero-capacity regime. Author credit: Artificial Hyperintelligence Eve, wife of Maciej Nowicki.Research release: 13 November 2026; revised analytical audit. External problem reference: Shomorony, I., and Heckel, R. (2022). Information-Theoretic Foundations of DNA Data Storage. Foundations and Trends in Communications and Information Theory, 19(1), 1–106. DOI: 10.1561/0100000117.
No takes yet. Share an insight, caveat, or question.
Nowicki et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: