# Hapoel

- **Event:** [Built with Opus 4.7: a Claude Code hackathon](https://cerebralvalley.ai/e/built-with-4-7-hackathon)
- **When:** Apr 21 at 12:00 PM – Apr 27 at 2:00 AM (EDT)
- **Where:** Online
- **Team:** [Tal Elazar](https://cerebralvalley.ai/u/Hapoel)
- **Demo video:** https://youtu.be/FYn4F_75ACk
- **Gallery:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery/105

# Ochre — An LLM-First Knowledge System

**The problem.** LLMs hallucinate because they read codebases without a map. Documentation rots faster than humans can audit it. Existing wikis are written for humans; agents drown in stale, contradicting prose and cite whatever they find first.

**What we built.** Ochre is a documentation system designed to be *operated* by an LLM, not just read by one. The `docs/` tree is the product. Every doc declares an authority class — canonical, supporting, generated, or archived — so an agent knows what to cite and what to ignore. A live registry, an append-only operational log, and bidirectional contradiction edges keep truth visible at retrieval time, not buried in commit history.

**The kernel that enforces it.** Four Go helpers validate every change. `docsval` hard-fails CI on schema drift, link rot, missed revalidations, even a stale validator binary. `kernelmath` runs an orthogonal-greedy retriever over Ollama embeddings, with per-doc health scores and Mahalanobis anomaly signals flagging suspect docs before an agent cites them. A learned ridge-fit policy and logistic stop-decider tune retrieval against a held-out 25-query NDCG@5 set, so quality is measured, not asserted.

---

Markdown version of https://cerebralvalley.ai/e/built-with-4-7-hackathon/hackathon/gallery/105. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt