IKKF

- Small models hallucinate and give flat, unverifiable answers - Fine-tuning is expensive, brittle, and needs retraining per domain - Single-shot prompting on frontier models is costly at scale - No way to audit *why* an AI gave an answer, or how sure it is How it works 1. Write atomic knowledge files (one concept per file, each with a source) 2. Point IKKF at the base and give it a task 3. IKKF decomposes the goal and drives all 7 phases end-to-end 4. Verify: every claim is cross-checked against a source 5. The final answer ships with a reasoning trace + confidence score Why it's different - Not plain RAG — adds structured reasoning + verification on top of retrieval - Auditable — see the reasoning trace and sources behind every answer - Reproducible — same knowledge base → same answers across sessions - Cheap — runs on a local 7B model (Ollama) or any OpenAI-compatible provider - Portable — swap the knowledge base to switch domains; no re-prompting from scratch