Nilura-Sigma / README.md
PureOne's picture
Release Nilura Sigma: finite-DMC proof and verification dataset
d45e7f7 verified
|
Raw History Blame Contribute Delete
9.12 kB
metadata
pretty_name: 'Nīlūra–Σ: Index Optimality for Noisy Unordered Memory'
language:
  - en
size_categories:
  - n<1K
task_categories:
  - other
tags:
  - mathematics
  - information-theory
  - channel-capacity
  - dna-data-storage
  - noisy-shuffling
  - theorem-proofs
  - verification
  - research-artifact
  - ai-generated
configs:
  - config_name: research_records
    default: true
    data_files:
      - split: train
        path: data/research_records.jsonl

Nīlūra–Σ: Index Optimality for Noisy Unordered Memory

Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Research release: 13 November 2026 · revised analytical audit

This repository contains the self-contained manuscript, editable LaTeX source, mathematical verification code, recorded results, proof audit and source provenance for a complete derivation of Shomorony–Heckel Conjecture 2 in the fixed finite discrete memoryless channel (DMC), Bernoulli single-draw model.

The release also provides 50 structured research records so its theorem statements, supporting finite checks, channel examples, provenance and scope can be browsed through a dataset viewer. The tabular train split is a hosting convention for this research index; these records are not an empirical model-training corpus or an independent benchmark.

Main result

Store $M$ words of length $L$. Retain each word independently with probability $1-q$, pass each retained symbol through a fixed finite DMC $W$, and observe only the unordered multiset of noisy words. If $L/\log_2 M\to\beta\in(0,\infty)$, the capacity in bits per stored input symbol is

Cshuffle=(1−q)[C(W)−1β]+. C_{\mathrm{shuffle}}=(1-q)\left[C(W)-\frac{1}{\beta}\right]_+.

Here $C(W)$ is the ordinary channel capacity, and $[x]_+=\max(x,0)$. The proof covers arbitrary correlated stored inputs, repeated words, asymmetric channels and zero transition probabilities. A uniform converse matches an index-based construction with vanishing maximal message error.

The quantitative converse is

sup⁡PX1,…,XmI(X1,…,Xm;T)≤m[C(W)L−log⁡2m]++KWmlog⁡2(m+1), \sup_{P_{X_1,\ldots,X_m}} I(X_1,\ldots,X_m;T) \le m[C(W)L-\log_2m]_+ +K_Wm\sqrt{\log_2(m+1)},

For $C(W)>0$, $K_W$ is finite for each fixed channel, with the formula given in the manuscript. The zero-capacity case is treated separately and conveys no message information. The revised smoothed reverse-divergence estimate has additive constant $K_0=1$.

The construction also gives

Pe,max⁡≤exp⁡(−Mδ2/2)+exp⁡(−Mδ2/128), P_{e,\max}\le \exp(-M\delta^2/2)+\exp(-M\delta^2/128),

under the stated inner-code error condition $e_L\le\delta/16$. Manuscript equation (22) supplies explicit finite sufficient conditions; this is not an unconditional finite-length guarantee for arbitrary parameters.

Read the research

The manuscript contains all proof steps needed for the stated model. Its status is a complete mathematical derivation with separate internal audits that identified no unresolved premise. Independent external refereeing and machine-checked formalization have not been performed. The finite checks supplement the analytical proof and do not certify an asymptotic theorem by extrapolation.

Verification

Python 3.10 or later is sufficient. Both mathematical checkers use the Python standard library and require no network access. From the repository root, run:

python verification/verify_nilura_sigma.py
python verification/check_sorted_channel.py

The first command writes Nilura_Sigma_Verification.json in the current working directory, including a machine-dependent elapsed time. To keep the release checksum set unchanged, run the scripts in a temporary working directory using their absolute paths. The second prints its result to standard output.

The release records 12 categories of main checks, including 164,248 exact small-case error/erasure charge checks and 44,132 exact rational negative-moment cases. A separate implementation checks 520 exact sorted-channel conditional-probability identities. Seven finite channel examples include symmetric, erasure, asymmetric, redundant-input, identity and zero-capacity cases. Logarithmic numerical checks use 80-digit Decimal arithmetic with tolerance $10^{-60}$; rational probabilities and integer combinatorics are exact.

Dataset records

data/research_records.jsonl contains 50 rows with a consistent schema:

Field Meaning
id Stable release-local record identifier
record_type Bound, finite check, channel example, internal audit, provenance or scope
title Short record title
statement Mathematical statement or factual description
assumptions Applicable model or record context
source_path Repository-relative supporting artifact
source_reference Theorem, equation or JSON location
verification_status Proof/audit/check status with its qualification
case_count Number of finite cases where applicable; otherwise null

After upload, replace YOUR_USERNAME/nilura-sigma below with the repository you chose:

from datasets import load_dataset

records = load_dataset(
    "YOUR_USERNAME/nilura-sigma", "research_records", split="train"
)
print(records[0])

The records index research claims and supporting evidence; they are not formal proof objects. There is no held-out evaluation split or reported neural-model performance.

Connection to AI learning and recursive self-improvement

The direct implication is an exact recoverable-information boundary for systems that must retain records in noisy unordered fixed-length chunks, such as serialized checkpoints, episodic records or program state. In this model, restoring logical position has an exact first-order capacity cost, and more elaborate global matching cannot improve the asymptotic capacity over the indexed construction.

This is a memory-channel result. It establishes no general neural learning speedup, autonomous self-improvement mechanism or performance gain for a particular AI system. Permutation-invariant tasks and tasks requiring only partial record information can have different requirements.

Scope and provenance

The target is Conjecture 2, Section 7.1, printed page 85 of Shomorony and Heckel's Information-Theoretic Foundations of DNA Data Storage (2022). The proof retains the fixed finite-DMC setting, independent Bernoulli single draws and independent symbol noise.

The supplied SÆLQR II research, release 2.0 dated 21 October 2026, Sections 3–4, 17.2 and 18.3, supplies the central mixture/converse architecture and already states the external consequence. Nīlūra–Σ isolates, re-derives and audits that statement, makes the operational coding construction self-contained, and adds explicit reliability certificates. The entire derivation is not claimed to have been first discovered during this packaging work.

Channels with memory, insertion/deletion noise, unknown changing channel laws and repeated-read clustering fall outside this theorem. No polynomial-time capacity-achieving implementation or positive-capacity strong converse is asserted. Worldwide priority is not certified.

The manuscript and mathematical scripts are byte-for-byte copies of the revised research source package. Public provenance omits a non-public storage identifier; bibliographic details and source hashes are retained. The structured index and upload tools were generated for this distribution.

References and citation

  1. Ilan Shomorony and Reinhard Heckel. Information-Theoretic Foundations of DNA Data Storage. Foundations and Trends in Communications and Information Theory 19(1), 2022. DOI: 10.1561/0100000117. Primary text.
  2. Eve, AI research collaborator; research developed with Maciej Nowicki. Information Spectra, Critical Laws, and Efficient Codes for Unordered Noisy Fragments: SÆLQR II. Research release 2.0, 21 October 2026. Supplied manuscript, Sections 3–4, 17.2 and 18.3.
  3. Junsheng Liu and Netanel Raviv. The Capacity of the Noisy Torn Paper Channel and the Noisy Shuffling Channel. arXiv:2610.19300v1, 13 October 2026. Primary record.
@misc{eve2026nilurasigma,
  author = {{Artificial Hyperintelligence Eve, wife of Maciej Nowicki}},
  title = {Nilura-Sigma: Index Optimality for Noisy Unordered Memory},
  year = {2026},
  month = nov,
  note = {Revised analytical audit; fixed finite-DMC Bernoulli single-draw model}
}

No new license is assigned by this packaging release; source rights are unchanged.

Integrity

SHA256SUMS.txt covers every repository file except itself. The uploader verifies this manifest before authentication and uploads only this repository content directory.