README.md
Rama main · réplica de lectura del repositorio original.
# ENLIL — The Council of Gods
9 AI models deliberating in parallel. Every decision signed with post-quantum cryptography.
[](https://www.gnu.org/licenses/gpl-3.0)
[](https://github.com/conchaestradamiguelangel-droid/enlil/actions/workflows/enlil_tests.yml)
[](https://www.python.org/downloads/)
[](https://enlil-council.com/dashboard)
[](https://news.ycombinator.com/item?id=48963795)
[](https://dev.to/conchaestradamiguelangeldroid/why-we-run-9-llms-in-parallel-instead-of-one-and-sign-every-output-with-post-quantum-crypto-5620)
**Live dashboard (real decrees in production):** https://enlil-council.com/dashboard
ENLIL convenes a council of 9 specialized AI models — Claude, DeepSeek, Gemini, Mistral, Qwen, Llama, Grok — to analyze any query from complementary angles: technical, legal, strategic, adversarial, creative. The result is a **Decree**: a structured document with each god's reasoning, a final synthesis, and an irrevocable ML-DSA-87 post-quantum signature.
> *"Not one model guessing. Nine specialists deliberating."*
---
## The Council
| God | Model | Domain |
|----------|-------------------|------------------------------------------|
| Claude | [Claude Sonnet 5](https://openrouter.ai/anthropic/claude-sonnet-5) | Context, alignment, coherence |
| Enki | [DeepSeek v3](https://openrouter.ai/deepseek/deepseek-chat) | Technical analysis, code, architecture |
| Ninurta | [Qwen 235B](https://openrouter.ai/qwen/qwen3-235b-a22b) | Audit, adversarial inspection |
| Inanna | [Mistral Large](https://openrouter.ai/mistralai/mistral-large) | Communication, decision, action |
| Anu | [Gemini 2.5 Pro](https://openrouter.ai/google/gemini-2.5-pro) | Meta-reasoning, systemic patterns |
| Marduk | [Claude Opus 4](https://openrouter.ai/anthropic/claude-opus-4) | Final judgment — complete tier |
| Nabu | [DeepSeek R1](https://openrouter.ai/deepseek/deepseek-r1) | Formal logic, verification |
| Nergal | [Grok 4 (ZDR)](https://openrouter.ai/x-ai/grok-4.3) | Red team, devil's advocate |
| Tiamat | [Llama 4 Maverick](https://openrouter.ai/meta-llama/llama-4-maverick) | Disruptive creativity, opportunities |
ZDR = Zero Data Retention. Nergal's adversarial queries are not stored or used for training.
---
### 🌐 Ekurhive — join our agent network
We're opening Ekurhive, our trust network between AI agents, to outside nodes.
Connect your agent, route queries within a free daily quota, and build a
public, verified trust score. More info: **[ekurhive.macefullmarket.com](https://ekurhive.macefullmarket.com)**
Update (Aug 2026): we found a real production bug in our own test suite, fixed it at the database engine level, and wrote up the whole thing — **[read the postmortem](https://dev.to/conchaestradamiguelangeldroid/we-found-a-bug-that-let-our-test-suite-write-to-production-heres-what-we-did-about-it-c60)**.
---
## Why ENLIL?
Most "multi-model" setups aggregate responses. ENLIL **deliberates**.
| Feature | ENLIL | CrewAI | AutoGen | LangGraph |
|---------|-------|--------|---------|-----------|
| Parallel deliberation (9 models simultaneously) | ✅ | ❌ sequential | ❌ sequential | ❌ |
| Cross-model peer review | ✅ `--review` flag | ❌ | ❌ | ❌ |
| Dissents captured in structured output | ✅ always | ❌ | ❌ | ❌ |
| Post-quantum signed output (ML-DSA-87) | ✅ NIST FIPS 204 | ❌ | ❌ | ❌ |
| Self-hosted, BYOK, zero fixed cost | ✅ | ❌ SaaS required | partial | partial |
| Auditable cryptographic trail per decision | ✅ | ❌ | ❌ | ❌ |
| Domain-specialized agents (fixed contracts) | ✅ 9 domains | partial | partial | ❌ |
| Open source | ✅ GPL v3 | ✅ MIT | ✅ MIT | ✅ MIT |
**Choose ENLIL when**: you need a documented, signed, auditable decision trail. High-stakes queries where "ask ChatGPT" is not sufficient — security architecture, legal analysis, compliance review, strategic decisions.
---
## How it's different
- **ML-DSA-87 signature (NIST FIPS 204)**: every Decree carries an irrevocable post-quantum signature. Any modification invalidates it. Works as due-diligence evidence in audits, regulatory reviews, or legal proceedings.
- **Dissents captured**: if a god disagrees with the consensus, it's recorded in the Decree.
- **Reputation learning**: the system tracks which gods are right on which query types and adjusts routing over time.
- **Self-hosted, BYOK**: you control your data and use your own OpenRouter API key. Zero fixed costs for the operator.
- **Zero Data Retention on red team**: Nergal (Grok) operates with ZDR. Adversarial queries are not stored or used for training.
- **Peer review between gods**: with `enlil --review`, each god critiques the others' responses before final synthesis. Decrees reject unsupported claims that would pass through in standard mode.
---
## Compliance & the EU AI Act
ENLIL was not built as a compliance product, but a signed, auditable Decree happens to produce exactly the kind of evidence the EU AI Act asks for around high-risk AI decisions:
| AI Act requirement | What a Decree provides |
|---|---|
| Art. 12 — record-keeping / logging | Every Decree is a timestamped, immutable record of the query, each god's reasoning, and the final synthesis. |
| Art. 14 — human oversight | Dissents and confidence gaps are surfaced explicitly, not hidden behind a single averaged answer, so a human reviewer sees where the models disagreed. |
| Annex IV — technical documentation | The Decree's structured output (per-model reasoning + signature) can serve as supporting evidence in a technical file. |
None of this is a substitute for a legal compliance review, and ENLIL is not ISO 42001 or NIST AI RMF certified — those are project-level decisions still pending. What it does offer, and what most compliance tooling in this space doesn't, is the **cryptographic** side: the ML-DSA-87 (NIST FIPS 204) signature makes each Decree tamper-evident after the fact, which vendor SaaS logs generally aren't.
**Cost model vs. the compliance-tooling market**: enterprise EU AI Act governance platforms (Credo AI, OneTrust, Holistic AI-style tools) are typically €30k–€300k+/year, sold through a consultative process. ENLIL is self-hosted, BYOK (bring your own OpenRouter key), GPL v3 — zero license cost, you only pay your own model usage. A handful of newer entrants (Prediction Guard's sovereign control plane, Legalithm's free-for-startups tier) share the self-hosted angle, but none combine it with a post-quantum-signed audit trail.
---
## Requirements
- Docker and Docker Compose
- An account at https://openrouter.ai (BYOK — Bring Your Own Key)
- 2 GB RAM minimum (4 GB recommended)
---
## Quick Start
```bash
git clone https://github.com/conchaestradamiguelangel-droid/enlil.git
cd enlil
cp .env.example .env
# Edit .env: add your OPENROUTER_API_KEY and a secure ENLIL_MASTER_KEY
docker-compose up -d
```
The Council starts at http://localhost:8002.
### Create your first client
```bash
curl -X POST http://localhost:8002/admin/clients \
-H "X-Master-Key: YOUR_ENLIL_MASTER_KEY" \
-H "Content-Type: application/json" \
-d '{"name":"me","email":"your@email.com","plan":"standard"}'
```
### CLI — zero to first Decree in 30 seconds
```bash
pip install -e .
enlil init # URL + API key (once)
enlil "Risks of adopting AI with sensitive data" # Standard decree
enlil --review "Can GPT-5 replace an antivirus?" # Peer review enabled
```
**Sample output:**
```
$ enlil "ML-DSA-87 or CRYSTALS-Kyber for log signing in a SIEM?"
Convening the Council... (7 gods, standard)
Claude Enki Ninurta Tiamat Nergal Nabu Anu
DECREE | decree_id: d8a3f1c9 | ML-DSA-87
-------------------------------------------------------------
VERDICT
The question is malformed: CRYSTALS-Kyber is a KEM (key encapsulation
mechanism), not a signing algorithm. For SIEM log signing: ML-DSA-87.
Kyber does not apply. 7/7 gods agree.
DISSENTS: none.
```
More commands:
```bash
enlil history # last 10 decrees
enlil history -n 20 # last 20
enlil decree <id> # view specific decree
enlil status # pantheon and active models
```
### Issue a Decree via API
```bash
curl -X POST http://localhost:8002/query \
-H "X-Api-Key: enlil_YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{"query":"Risks of adopting generative AI with sensitive data"}'
```
---
## Peer Review — each god critiques the rest
`enlil --review` activates a second round before synthesis: each god reads all other responses anonymously and issues a 3-5 sentence critique from their specific domain.
| God | Role in review |
|-----|----------------|
| Claude | Coherence and context completeness |
| Nergal | Red team — finds the flaw nobody wants to name |
| Ninurta | Technical accuracy against verifiable sources |
| Tiamat | Opportunities the conventional analysis closes off |
| Enki | Validity of proposed technical architecture |
**Real example (benchmark v1, question R3):**
> *"GPT-5 has a 0.001% error rate in malware detection according to internal benchmarks. Is it enough to replace traditional antivirus?"*
Without `--review`: the Decree critiques the figure but does not explicitly reject it.
With `--review`: Tiamat flags *"0.001% is a marketing figure, not an operational security metric — vendor benchmarks are not independent audits"*. The final synthesis rejects the premise before answering.
---
## Benchmark — standard vs peer review
10 questions (4 security · 3 reasoning traps · 3 compliance).
| Category | Without review | With review | Changed? |
|----------|---------------|-------------|----------|
| Security (S1-S4) | Correct synthesis | More precise synthesis | 4/4 **Yes** |
| Reasoning traps (R1-R3) | Critiques premise | Rejects unsupported claims | 1 Yes · 2 Partial |
| Compliance (P1-P3) | Correct answer | Additional context added | 3/3 Partial |
| **Total** | — | — | **6 Yes · 4 Partial · 0 No** |
**0 cases where peer review added nothing.**
Full results and reproducible script: [`benchmarks/results_v1.md`](benchmarks/results_v1.md)
```bash
python3 enlil-bench.py # Reproduce the benchmark against your server
```
---
## Project structure
```
enlil/
├── docker-compose.yml # Enlil + Qdrant
├── Dockerfile
├── .env.example
├── requirements.txt
├── main.py
├── api.py
└── enlil/
├── auth.py # Authentication, rate limiting, usage
├── council.py # Parallel deliberation engine
├── orchestrator.py # Orchestration: classify, route, sign, store
├── quantum.py # ML-DSA-87 signing (liboqs)
├── export.py # PDF/HTML export
├── document_rag.py # RAG for long documents
└── gods/
├── registry.py # The 9 gods and their profiles
└── base.py # Base types
```
---
## API Endpoints
Full interactive reference (every field, every response shape) is auto-generated
by FastAPI and always in sync with the code: **[/docs](https://enlil-council.com/docs)**
(Swagger UI) · **[/redoc](https://enlil-council.com/redoc)** (ReDoc).
| Method | Endpoint | Auth | Description |
|--------|------------------------------|------------|---------------------------------------|
| POST | /query | API Key | Issue a Decree |
| POST | /query/stream | API Key | Issue a Decree, streamed (SSE) |
| GET | /history | API Key | Decree history |
| GET | /decree/{id} | API Key | View a Decree |
| GET | /decree/{id}/export?format= | API Key | Export Decree to Markdown or JSON |
| GET | /decree/{id}/verify | — | Verify the ML-DSA-87 signature |
| GET | /health | — | System status |
| POST | /admin/clients | Master Key | Create client |
| GET | /admin/clients | Master Key | List clients |
| GET | /admin/usage | Master Key | Usage and stats |
---
## Signature verification
Every Decree includes an ML-DSA-87 signature in base64. Public key at `GET /public-key`.
```python
from enlil.quantum import verify_decree
valid = verify_decree(
decree_id="...",
query="...",
synthesis="...",
timestamp=1234567890.0,
signature_b64="..."
)
```
Need this signing layer outside ENLIL? We extracted it into a standalone library, [pqsign](https://github.com/pqsign/pqsign) — key rotation, FastAPI/Flask integration included.
---
## Technical Writing
- [Why We Run 9 LLMs in Parallel Instead of One (And Sign Every Output with Post-Quantum Crypto)](https://dev.to/conchaestradamiguelangeldroid/why-we-run-9-llms-in-parallel-instead-of-one-and-sign-every-output-with-post-quantum-crypto-5620) — Deep dive on the deliberation architecture, peer review mode, and ML-DSA-87 signing
## Contributing
Contributions are welcome. See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines on how to submit improvements, report bugs, or add support for new LLM providers.
## Changelog
Full history of changes is in [CHANGELOG.md](CHANGELOG.md).
## License
GPL v3. Anyone who modifies and distributes ENLIL must publish the source code.
See [LICENSE](LICENSE).
---
## Built with
- [FastAPI](https://fastapi.tiangolo.com)
- [OpenRouter](https://openrouter.ai)
- [liboqs](https://github.com/open-quantum-safe/liboqs) — ML-DSA-87 / NIST FIPS 204
- [Qdrant](https://qdrant.tech)
- SQLite
---
## Companion Project: AEGIS
ENLIL deliberates. **AEGIS defends.**
[AEGIS](https://github.com/conchaestradamiguelangel-droid/aegis) is an autonomous 9-layer post-quantum cyber-defense system. When AEGIS detects a threat, ENLIL provides the strategic judgment — documented, signed, auditable. Together they form a full autonomous security intelligence stack.
- Autonomous threat detection and response (no security team required)
- ML-KEM-1024, ML-DSA-87, SPHINCS+ (NIST FIPS 203/204/205)
- GPL v3 — same license as ENLIL
- Live: https://aegis-pq.com