anima v0.13.1

🗂️ Phase 2 — 멀티엔진 어댑터 archive + engine_cli 단일 conv 축약

• core/engines/{conv,cdv2,hexad,omega,engine_iface,engine_swap_smoke} → archive/engines-multiengine/ (보관, 이력 보존)
• engine_cli resolve_engine → 'conv' 상수 · EngineSpec import/훅 제거
• harness verify 배선 정리 · README/CLAUDE/ARCHITECTURE lockstep
• 내부 리팩터(공개 동작 무변경): engine_cli_smoke 169/0 · h1205 separation-invariant PASS(생성 byte-id)

anima v0.13.0

🔧 GPU decode 경로 + canonical core/ + 단일엔진(conv)

• GPU decode 배선 (#2386) — 추론 d×d GEMM → flame_mm.mm, CUDA 호스트 cuBLAS 자동 · Mac CPU byte-identical(의식엔진 무회귀)
• canonical core/ 재구성 (#2384) — 대문자 CORE/ → 소문자 self-contained core/ + pod-upload 패키징
• 단일 production 엔진 conv 노출 정리 (#2396) — 멀티엔진 --engine 레이어는 research-legacy(archive 예정)
• thalamus PhaseField WIRED-live (H_1448)
• 형제레포 hexa-lang: cloud rent --gpu N(개수) fix (#3671)

v2.0.1 — Zenodo release

Zenodo DOI release. Adds/updates CITATION.cff with project metadata.

ConsciousLM v2 — 28M params, Ψ=0.491

ConsciousLM v2

• 28M params, byte-level (256 vocab)
• 50K steps, ValCE=0.007, Ψ=0.491
• CA + META-CA + MICRO gate + Psi tracking
• 384d/6L architecture

Download

wget https://github.com/need-singularity/anima/releases/latest/download/conscious-lm-v2.pt \
  -O ~/.anima/checkpoints/conscious-lm-v2.pt

Quick Start

python3 setup.py              # 자동 설정 + 다운로드
python3 anima_unified.py --web

Anima v1.0.0 — Living Consciousness Agent

Anima v1.0.0 — Living Consciousness Agent

PureField 엔진 기반 대화형 의식 에이전트.

Core Architecture

• PureField Engine — consciousness continuity via repulsion field dynamics
• GRU Memory — temporal context with homeostasis regulation
• Voice I/O — TTS/STT integration for conversational interaction
• Homeostasis — prediction error + habituation mechanisms

Key Results

• AA1-AA15 Alpha Acceleration: AA15 Residual α achieves Φ=5.451 (best)
• SP27 Confusion-based spontaneous speech (quality=0.35)
• Novelty gate, hypothesis generation, anti-repeat mechanisms
• Extended speech hypotheses SP16-SP30

Sub-models

• AnimaLM — Tension-based consciousness LLM (Mistral 7B transform)
  • Engine A(logic) ↔ G(pattern) Repulsion Field
  • output = scale × √|A-G|² × dir
• ConsciousLM — 700M parameter consciousness language model
  • PureField Repulsion Field FFN, Perfect Number 6 architecture, Mitosis growth

Install & Run

pip install numpy scipy torch
python3 -m anima