RELEASE_NOTES.md

FNE–AXIOMESH v0.2.0 — Hugging Face release notes

Major additions over the earlier ESRE kernel

• Counterfactual Quotient Weaving (CQW) as the central architecture transaction.
• Capability-induced memory revalidation: newly learned operations can make previously discarded distinctions computationally necessary.
• Exact finite-linear construction of the minimal Fractal Tether.
• Recursive fracture and root regeneration without retaining internal parent seeds.
• Query-relative exact recoverability certificates and ambiguity witnesses.
• Minimum-rank evidence acquisition under the declared observation family.
• Non-averaging mixed-operation Unison tests.
• Robust evidence ledger and strong-baseline reporting.

Evidence summary

• 77 automated tests pass.
• Five seeded runs are included.
• Proposed finite pipeline and strong pooled algebra baseline both achieve 10,000/10,000 held-out composition predictions.
• 4,500 controlled erasure trials produce no false recovery certificates within the tested exact-erasure model.
• 1,000 exhaustive finite-fiber identifiability checks are included.

Important negative result

The strongest conventional finite learner ties the principal prediction result. Therefore this release does not establish an intelligence advantage, superlinear recursive self-improvement, or a new RSI class.

Primary research question for the next release

Does jointly selecting an executable abstraction and its future regeneration obligations outperform abstraction learning followed by independently optimized repair under strictly equal total resources?

Agent and Expert Support edition (0.2.0-hf2)

• Added data-only JSON revalidation, compilation, and frozen-recovery interface.
• Added AGENTS.md, capability/request schemas, llms.txt, research index and a machine-readable claim ledger.
• Added expert proof-to-code mapping, assumption audit, falsification targets and an unfilled review template.
• Added a proposed equal-resource joint-versus-separate selection protocol.
• Added offline hash/link/JSON validation and uploader tests for integrity, bounded retries and dry-run behavior.
• Updated Windows launcher to support py -3 and retained visible error handling.
• Kept scientific source, original tests, manuscript and recorded experiment data unchanged.
• No neural training, new benchmark advantage, human peer review, priority clearance or publication is claimed by this edition.

The original 77 tests and the new support checks are reported separately in results/agent_expert_qa.json. Seed directory names are RNG labels, not later calendar dates.

Dataset distribution (0.2.0-hf3-dataset)

• Changed the uploader default and all default instructions to dataset repositories.
• Added dataset-card metadata and an explicit five-row experiment-results subset.
• Added a data dictionary and provenance-preserving normalized JSONL export.
• Preserved agent support, expert support, manuscript and original research evidence.
• Added a check that the uploader defaults to the dataset namespace.