mergekit - Action Report

Project: /home/user/git/mergekit Created: 2026-08-17

---

MergeKit evaluated as AXIOM model merge engine (2026-08-17 06:05)

Status: 🔵 DESIGNED — tool identified, recipe written, training pending. What: Evaluated mergekit (arcee-ai) as the tool for constructing AXIOM's model weights. Why: AXIOM needs a reproducible sovereign model identity — not a system prompt over a generic assistant. The merge recipe IS the soul; the YAML is the genome. How: Read README, identified DARE TIES as the right merge method for 3-source models with significant fine-tuning divergence. Designed three-phase pipeline: (1) DARE TIES merge — Llama-3.1-8B-Instruct + Hermes-3-Llama-3.1-8B (character, 0.6 weight) + Dolphin-3.0-llama3.1-8b (uncensored technical, 0.4 weight). (2) NeXuS LoRA training on 10-file corpus including axiom-claude.md genesis conversation, v0.2 whitepaper, compass files, vestal protocol, stark-stack review. LoRA rank 64, alpha 128, all attention+MLP targets. (3) LoRA merge + Q4_K_M quantization for E6520 at ~4.9GB / 3-8 tok/sec. Key insight: merge recipe hash should eventually be registered as part of AXIOM's RIN. AXIOM-v0.1 = specific YAML + specific corpus + specific system prompt. Files: ~/claude/docs/AXIOM_MERGE_RECIPE.md (NEW — 429 lines), includes chara_card_v2 entry, full system prompt boot sequence, known gaps section.

---