korea-ai-foundation-model-verification

Use when planning, submitting, or debugging LLM-DNA fingerprint extraction jobs on SageMaker spot training. Trigger on keywords like "LLM-DNA jobs", "calc-dna", "submit_dna.py", "SageMaker spot", "Solar / K-EXAONE / A.X-K1 / EXAONE 추출", "Phase 5/6 재시도", "model fingerprint extraction", or follow-up verification rounds. Codifies 9 cost/operational lessons from Phase 4–6 of korea-ai-foundation-model-verification (2026-04-30, $64 / 77 jobs, 52.8% waste rate) so future rounds avoid the same learning curve.

Use when working on LLM model lineage, provenance, phylogenetic analysis, from-scratch verification, the project's verification targets (Solar-Open-100B, K-EXAONE, A.X-K1, HyperCLOVAX), reference families (Llama 3, Qwen 2.5, GLM, Mixtral, DeepSeek), or anything under experiments/llm-dna/. Routes through the project wiki/ before external search.

Use when writing, reviewing, or refactoring code in this repo to reduce common LLM coding mistakes — overcomplication, drive-by edits to unrelated lines, hidden assumptions, vague success criteria. Four principles: Think Before Coding, Simplicity First, Surgical Changes, Goal-Driven Execution.

Use when the user invokes /update-changelog, or when a session has produced concrete repository changes (file edits, verification results, methodology updates, structural changes) that should be recorded in CHANGELOG.md by date and category.

Use when the user invokes /update-tutorial, or when a verification-related Q&A exchange has just completed in the session and should be persisted into docs/tutorial/ as narrative-style content. Classifies by topic, finds the right file, inserts above the SECTION_MARKER, and updates the README index.