v0.5.1 — GitHub packaging

Apache-2.0, NOTICE, ATTRIBUTION, CITATION.cff, CODE_OF_CONDUCT, CONTRIBUTING, SECURITY. README.ru.md / README.zh.md. Topics and research labels. Same packaging bar as moe-orbit-prefetch.

v0.5.0 — scaling curve N=6..50

Sandbox §8jj / iteration S.

Resonance advantage grows with scale (1.75x → 2.35x less perplexity damage). Separate learning-capacity plateau (~9–10 facts) disclosed — same for both C choices.

v0.4.0 — joint multi-fact MEMIT win

Sandbox §8ii / iteration R.

Task adapted to MEMIT's real regime (simultaneous multi-fact joint solve). Regularization bug (C+KK^T) caught and fixed. After the fix: resonance C less damage than corpus C, both learn 6/6.

v0.3.0 — real MEMIT formula + Reflection gate

Sandbox §8hh.

Literal MEMIT spread from the paper plus Reflection-gated abort (v3). Neither beats a single-layer edit on one fact.

v0.2.0 — MEMIT v1 incomplete, v2 worse

Sandbox §8gg / iteration Q.

User caught that v1 was not a real MEMIT adaptation. v2 remaining-gap re-estimation made collateral damage worse, not better. Honest negative kept.

v0.1.0 — resonance-weighted ROME freeze

Sandbox §§8bb–8ff / iterations O–P.

Single-layer ROME on GPT-2-small, all 12 layers: resonance-weighted C vs corpus C. Self-calibrated hybrid rule (K≥S) on 3- and 6-fact samples. Tokenization confound (thunderclap) documented, not hidden.