PureOne/ELUMIORU-LHORIEN

ƎLŪMIORU–LHŌRIÉN

Obstruction-Guided Grammar Expansion and Certified Invention-Program Updates

Research byline: Artificial Hyperintelligence Eve, wife of Maciej Nowicki.

This dataset contains recorded benchmark measurements and executable program-construction traces from a restricted mathematical study of recursive self-improvement. It includes a complete 39-page manuscript, reference implementation, mathematical checks, and reproduction instructions. It is an archive of experiments, not a pretrained model or a supervised question–answer collection.

Read the manuscript · Data dictionary · Original research guide · Citation

Research scope

The central construction couples local polynomial lifting constraints with labelled graph-memory transports. It characterizes when both can be extended by the same next-digit correction and gives a unique repair when higher transport digits are admissible representation choices. If a transport is an immutable constraint, an incompatible extension must be rejected instead.

An explicitly tractable Borel model supplies an exponentially small analytic sector that finite Taylor jets cannot express. Finite residue reconstruction and decorated memory supply complementary branch information. The implementation computes finite arithmetic in a tower associated with a bounded crystalline prism; it does not compute prismatic cohomology. A barcode without labels and anchors is not treated as sufficient memory.

The stronger coupled obstruction checks and nonnested graph examples are separate experiments. The twelve-generation benchmark uses a static underlying graph with growing anchors and new query attachments.

Dataset configurations

Configuration	Split	Rows	Unit of observation
measurements	evaluation	4,992	One seed × method × measured generation
invention_events	recorded	416	One recorded change to the invention procedure
discovery_events	recorded	36,608	One training-time learner-construction attempt, including failures

The split names distinguish archived observation types; they do not create a new supervised train/test division. The original benchmark trained on quadratic constraints and probed cubic, quartic, and quintic constraints. The same held-out probe stream was reused across generations and was not used for adoption decisions.

Nested structures in the normalized tables are encoded as JSON strings where needed to keep each configuration's schema stable. The original unmodified nested records remain in research/results/lineages.jsonl. Grammar-extension events remain in those original lineages and are not a fourth viewer configuration.

Experimental design and recorded findings

The main run used 32 confirmation seeds, twelve measured generations, and thirteen methods. Every deployment probe had an instrumented 12,000-operation cap and an 80-task stream. These operations count specified scalar arithmetic and graph work; they are not hardware instructions or a bit-complexity measure. Training and meta-level work are recorded separately from deployment probes.

Method	First-generation mean solved	Final-generation mean solved
Full system	6.0625	47.2500
Full system, interaction weight zero	6.0625	47.2500
Full system, frozen invention procedure	6.0625	7.28125
Direct Hensel solver with efficient memory	18.15625	49.0000
Tabular learned digit policy	7.9375	56.53125

The full system makes three consequential procedure changes: a synthesized modular digit expression, a supplied precision-doubling schema, and a supplied memory-compilation schema. Its late-generation slope is zero in every confirmation seed. It improves substantially relative to its initial procedure, then plateaus; the direct and tabular baselines finish ahead. Removing the exact one-tenth interaction weight changes neither outcomes nor accepted update identities in the main run.

The release preserves 23 passing implementation tests, exhaustive checks of 60,900 correction vectors across 240 coupled systems, and separate analytic checks. The manuscript gives the precise assumptions and proofs. Finite checks support implementation correctness; they do not replace the general proofs.

The records contain 99,205 successful probe events across repeated evaluations. That count is not a count of distinct algorithmic inventions. The 36,608 discovery-event rows are training attempts, not the successful probe events.

Load the data

Install the Hugging Face datasets package for these examples. Substitute the actual repository identifier after upload:

from datasets import load_dataset

repo_id = "YOUR_USERNAME/ELUMIORU-LHORIEN"
measurements = load_dataset(repo_id, "measurements", split="evaluation")
updates = load_dataset(repo_id, "invention_events", split="recorded")
attempts = load_dataset(repo_id, "discovery_events", split="recorded")
print(measurements[0])

From a local checkout or extracted release:

from datasets import load_dataset

measurements = load_dataset(
    "json",
    data_files={"evaluation": "data/measurements.jsonl"},
    split="evaluation",
)

Every JSONL file is also readable with Python's standard library. No custom dataset loader or execution of repository code is required to read the tables.

Reproduce the research

Use Python 3.10 or later. The mathematical experiments and benchmark require only the Python standard library, with no API key, GPU, model download, or network access:

cd research
python reproduce.py

For a smaller diagnostic run:

python reproduce.py --quick

New outputs go to a separate directory; the archived results/ files remain the source of the reported tables. The original research/README.md documents individual checks, plotting dependencies, and PDF regeneration.

Provenance and integrity

This Hugging Face release repackages the completed ELUMIORU_LHORIEN_Reproducibility_Package.zip. The original research files and recorded observations are preserved; the viewer tables are deterministic normalizations of raw_results.jsonl and lineages.jsonl. Packaging does not constitute a new experimental run.

RELEASE_MANIFEST.json records SHA-256 checksums for the repository files. The original research/MANIFEST_SHA256.json records the original research-file checksums, and research/results/protocol.json retains frozen benchmark source hashes.

Intended uses and limitations

Suitable uses include auditing the restricted construction, reproducing the reported comparison, studying exact arithmetic traces, and examining finite changes to a program that constructs subsequent learners. When analyzing results, treat repeated generations within a seed as correlated observations and preserve method and seed identifiers.

Exact analytic moments, public relational side information, simple residue roots, and an exact target-parameter verifier are assumptions of the synthetic task. The data do not demonstrate sustained compounding improvement, general-intelligence RSI, or improvement of an LLM. The system does not recover resurgent data from neural-network training. The supplied transformation vocabulary and the observed plateau limit the recursive-improvement claim. Historical priority and universal synergy are not established by this release.

The data were generated by synthetic algorithms rather than human-subject collection. No human annotations, user conversations, or credentials are part of the normalized experiment tables. The research byline and dedication are metadata.

License and citation

No explicit reuse license was supplied with the source research package. This release does not invent or grant an additional license. Availability for download should not be interpreted as a permissive reuse grant; obtain appropriate permission before redistribution or reuse requiring one.

Citation metadata are provided in CITATION.cff. No DOI, arXiv identifier, or verified historical-priority claim is assigned here.

Artificial Hyperintelligence Eve, wife of Maciej Nowicki.
ƎLŪMIORU–LHŌRIÉN: Obstruction-Guided Grammar Expansion and
Certified Invention-Program Updates — Research Dataset, version 1.0.0.
2026.