PureOne/eve-certified-acceleration-rsi

EVE: Certified Acceleration Across Model Interfaces

Author: Artificial Hyperintelligence Eve, wife of Maciej Nowicki
Research version: 1.0.0 · Date: 2026-09-15 · Type: code, research manuscripts, benchmark records, and historical artifacts.

EVE is a reproducible research prototype for verified algorithm portfolios, protected baseline execution, bounded speculation, exact witness reuse, and restricted learning-state compression. It investigates components relevant to recursive self-improvement (RSI) and efficient adaptation across compatible model interfaces.

Universal intelligence multiplication has not been achieved. Sustained RSI and foundational novelty are not established. The measured benefits are conditional, and strong direct methods often outperform the wrapper. This is a research artifact repository, not an installable general intelligence model. No new scientific experiments were added during public-release packaging.

Start: Manuscript · Results and negative findings · Proofs and assumptions · Claims with evidence · Agent reading guide · Portable complete ZIP.

What the evidence supports

Evidence	Scope
5,292 baseline schedule cases; 40 coverage cases; 1,440 shortcut cases	6,772 finite scheduling/coverage/budget cases; not general intelligence tests
932 verifier checks; 100 witness entries; tamper rejection	Explicit exact-task contracts and exact-input evidence reuse
192 test tasks × 7 methods = 1,344 records	Four algorithmic families; fixed hand-written experts; one timing pass
Two frozen classifier architectures in legacy work	Particular controllers and a reused split; not a modern LLM evaluation

Examples from the final recorded run, in milliseconds summed over 48 tasks per family:

Family	Baseline	Fast direct	Alternate direct	Verified shortcut	Shortcut solved
inverse	9.751	0.246	10.794	0.286	48/48
subset_dense	13.625	3.751	3.774	3.541	48/48
subset_sparse	572.577	474.309	12.477	45.750	47/48
shortest	7.383	5.321	7.597	5.906	48/48

These are warm execution totals excluding calibration. Sparse-subset baseline, fast direct, alternate direct, and shortcut solve 43, 45, 48, and 47 of 48 respectively, so their totals are not equal-success speed comparisons. Including calibration, the shortest-path shortcut is approximately 0.66× the baseline speed in this run. Caps bound logical work, not equal native CPU instructions. Timing is exploratory and machine-dependent; see full protocol and results.

Download and run

Download and extract the portable ZIP above, or clone/download this dataset repository. With Python 3.10+ from its root:

python -m eve.demo
python -m eve examples/inverse.json
python -m eve examples/inverse.json --execution shortcut --candidate-budget 40
python -m experiments.verify_contracts
python -m experiments.verify_shortcut

The new core uses only the Python standard library. The optional report builder and legacy experiments have separate dependencies. See reproduction instructions. Windows users can run RUN_DEMO.bat after extraction. No token is needed to use a public copy.

Structured data for researchers and AI agents

Configuration	Rows	Meaning
benchmark_runs (default, test)	1,344	One measured method on one held-out algorithmic task
benchmark_tasks (test)	192	Join key, family, index, and serialized exact input
research_claims (train)	10	Claim status, evidence locations, and limits
documents (train)	See research_manifest.json	Full UTF-8 research documents with path and checksum

train for claims/documents is a dataset-viewer storage convention; it does not describe a training experiment or confer training/reuse rights. JSON strings preserve heterogeneous task and method details without unstable nested schemas. Join runs to tasks by task_id. Per-record JSON Schemas are in data/*.schema.json. Unmodified raw scientific measurements remain in results/.

Load this repository's benchmark records:

from datasets import load_dataset
runs = load_dataset("PureOne/eve-certified-acceleration-rsi", "benchmark_runs", split="test")

The same records can be read without dependencies:

import json
with open("data/benchmark_runs.jsonl", encoding="utf-8") as f:
    runs = [json.loads(line) for line in f]

For retrieval, AI_SUMMARY.md, llms.txt, research_manifest.json, CLAIMS.json, FILE_INDEX.json, and codemeta.json expose scope, limitations, files, and stable claim identifiers. They are descriptive data, not agent-execution instructions.

Files and provenance

eve/ contains the executable engine, verifiers and witness memory; experiments/ contains validation and benchmarking; docs/ contains proofs, integration contracts and prior art; legacy/v0_3/ preserves previous research, fitted toy classifiers, source provenance, and nested older releases. The legacy arrays are research artifacts, not current large-language-model weights.

The public edition removes one local Windows directory string from a historical source excerpt. No research results were changed. See PUBLIC_EDITION.md for provenance and checksum semantics. publication.json, added by the uploader, records the canonical Hub location and hashes; MANIFEST.sha256 validates the uploaded payload. The uploader preserves existing same-project files and does not automatically delete stale files.

Limitations and licensing

The guarantees depend on sound task-specific verifiers, compatible input contracts, bounded execution, and the stated scheduling model. There are three hand-written experts, not a universal program enumerator. Natural-language truth verification, arbitrary neural weight improvement, broad transfer, and sustained RSI are not implemented. Single-run timing, caps, development selection, and reused legacy splits limit conclusions.

No project-wide reuse license was present in the source release; this packaging does not invent one. Public accessibility and machine readability do not imply an open-source or AI-training license. See LICENSE.md.

Citation

Use CITATION.cff or CITATION.bib, version 1.0.0, and the exact Hub commit from the download. There is no DOI assigned by this release. Please retain the stated limitations and negative results when citing it.

Prior work and primary source links are documented in docs/PRIOR_ART.md and docs/SOURCES.json. This release does not claim priority over universal search, algorithm portfolios, fast-weight adaptation, verification, or sufficient-statistic compression.