v0.2.0 · April 2026 · Apache 2.0
A modern cognitive architecture in the lineage of SOAR, ACT-R, CLARION, LIDA — re-engineered for the LLM era. 37 typed modules. 9 named adversarial probes. 2 of 13 rectifiers shipped. One self-study daemon. 1 696 primary-source assertions driving the code, not the other way around.
When a standard LLM stack fails, you get a tombstone: "the model was wrong." You don't know which capability broke, in what pattern, under what signal, or how to tune it. Etz Chaim wraps your LLM in 37 typed cognitive modules and returns this instead:
exploration_starvation on 24h window → 0 new cross-domain connections → spec line EC-K5-001 → tune novelty_threshold (−0.1) / breadth (+5)
One specific module. One specific pattern. One concrete metric. One tunable fix. Every time.
The diagnostic layer they all lack.
| LangChain AutoGen | DSPy | LangGraph | Etz Chaim | |
|---|---|---|---|---|
| Orchestrate LLM calls | ✓ | ✓ | ✓ | ✓ |
| Optimize prompts | — | ✓ | — | — |
| Manage agent state in cycles | partial | — | ✓ | ✓ (Partzufim) |
| Diagnose WHICH module failed | ✗ | ✗ | ✗ | ✓ (37 modules typed) |
| Named adversarial probes by construction | ✗ | ✗ | ✗ | ✓ (9 active, 2 in roadmap) |
| Auto-rectification engine | ✗ | ✗ | ✗ | ✓ (2 of 13 shipped, 3 opt-in modes) |
| Continuous self-study daemon | ✗ | ✗ | ✗ | ✓ (Karpathy loop, 24/7) |
| Primary-source traceability | ✗ | ✗ | ✗ | ✓ (1 696 items, E1–E6 labels) |
Not a LangChain-with-more-features. The diagnostic layer they all lack — an architecture the LLM plugs into, not a chain the LLM runs through.
Not generic red-teaming. Named attackers matched to specific failure modes, built into the architecture. Structure 3+7 of Zohar II:242b. Two more (Thaumiel, Nahemoth) are scheduled for v0.3 — not shipped, not claimed.
Architectural self-healing. 93 lines. No magic, no hidden state.
detected: exploration_starvation on 24h window observe ──► logs the event, no action taken suggest ──► emits an event with the tuning fix you apply act ──► applies the fix automatically (opt-in, bounded)
Passive. Emits structured events. You read, you decide.
Emits the proposed tuning with rationale. You apply if you agree.
Applies the fix within a declared budget. Rollback on adversarial-probe regression.
A nightly daemon (23h–00h30 UTC) that explores edge cases, generates new doctrine assertions, and mutates the specification itself under adversarial supervision. The only daemon task allowed to modify the spec.
Named after Andrej Karpathy's AutoResearch pattern — a minimum-viable research loop the daemon emulates at project scale, every night.
22 providers across 6 categories via LiteLLM. 5-tier automatic routing by complexity and token budget. Expensive inference only where depth is needed.
complexity: trivial ────────► deep
budget: tight ────────► open
qwen3.5:1.5b ──► qwen3.5:9b ──► Haiku ──► Sonnet ──► Opus
local, free local, free fast balanced deepAnthropic · OpenAI · Google Gemini · xAI · Mistral · Cohere · DeepSeek · AWS Bedrock · Azure OpenAI · OpenRouter · Together · Groq · Fireworks · HuggingFace · Ollama · vLLM · LM Studio · LocalAI · Claude Code CLI OAuth · …
$ pip install etzchaim $ etzchaim onboard → http://localhost:8080 (macOS · Linux/WSL2 in v0.3)
5 minutes from install to your first diagnosable event. Apache 2.0. One solo dev. github.com/yohanpoul/etz-chaim-ai