{"id":50118762,"url":"https://github.com/helixprojectai-code/helix-tel-deploy","last_synced_at":"2026-05-23T17:08:53.770Z","repository":{"id":358369399,"uuid":"1240792927","full_name":"helixprojectai-code/helix-tel-deploy","owner":"helixprojectai-code","description":"Zero-exchange cryptographic key derivation from constitutional AI behavior. Two nodes independently derive an identical encryption key through deterministic convergence — no key transmitted, negotiated, or pre-shared.","archived":false,"fork":false,"pushed_at":"2026-05-17T02:38:20.000Z","size":99,"stargazers_count":0,"open_issues_count":0,"forks_count":0,"subscribers_count":0,"default_branch":"main","last_synced_at":"2026-05-17T04:22:00.067Z","etag":null,"topics":["ai-safety","constituional","constitutional-ai","convergence","cryptography","encryption","grammar","helix-ttd","key-derivation","llm","mesh-network","python","zero-knowledge"],"latest_commit_sha":null,"homepage":"https://helixprojectai.com/","language":"Python","has_issues":true,"has_wiki":null,"has_pages":null,"mirror_url":null,"source_name":null,"license":"apache-2.0","status":null,"scm":"git","pull_requests_enabled":true,"icon_url":"https://github.com/helixprojectai-code.png","metadata":{"files":{"readme":"README.md","changelog":null,"contributing":null,"funding":null,"license":"LICENSE","code_of_conduct":null,"threat_model":null,"audit":null,"citation":null,"codeowners":null,"security":null,"support":null,"governance":null,"roadmap":null,"authors":null,"dei":null,"publiccode":null,"codemeta":null,"zenodo":null,"notice":null,"maintainers":null,"copyright":null,"agents":null,"dco":null,"cla":null}},"created_at":"2026-05-16T15:20:56.000Z","updated_at":"2026-05-17T02:38:23.000Z","dependencies_parsed_at":null,"dependency_job_id":null,"html_url":"https://github.com/helixprojectai-code/helix-tel-deploy","commit_stats":null,"previous_names":["helixprojectai-code/helix-tel-deploy"],"tags_count":null,"template":false,"template_full_name":null,"purl":"pkg:github/helixprojectai-code/helix-tel-deploy","repository_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/helixprojectai-code%2Fhelix-tel-deploy","tags_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/helixprojectai-code%2Fhelix-tel-deploy/tags","releases_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/helixprojectai-code%2Fhelix-tel-deploy/releases","manifests_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/helixprojectai-code%2Fhelix-tel-deploy/manifests","owner_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners/helixprojectai-code","download_url":"https://codeload.github.com/helixprojectai-code/helix-tel-deploy/tar.gz/refs/heads/main","sbom_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/helixprojectai-code%2Fhelix-tel-deploy/sbom","scorecard":null,"host":{"name":"GitHub","url":"https://github.com","kind":"github","repositories_count":286080680,"owners_count":33404577,"icon_url":"https://github.com/github.png","version":null,"created_at":"2022-05-30T11:31:42.601Z","updated_at":"2026-05-23T04:15:53.637Z","status":"ssl_error","status_checked_at":"2026-05-23T04:15:53.242Z","response_time":53,"last_error":"SSL_read: unexpected eof while reading","robots_txt_status":"success","robots_txt_updated_at":"2025-07-24T06:49:26.215Z","robots_txt_url":"https://github.com/robots.txt","online":false,"can_crawl_api":true,"host_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub","repositories_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories","repository_names_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/repository_names","owners_url":"https://repos.ecosyste.ms/api/v1/hosts/GitHub/owners"}},"keywords":["ai-safety","constituional","constitutional-ai","convergence","cryptography","encryption","grammar","helix-ttd","key-derivation","llm","mesh-network","python","zero-knowledge"],"created_at":"2026-05-23T17:08:50.083Z","updated_at":"2026-05-23T17:08:53.757Z","avatar_url":"https://github.com/helixprojectai-code.png","language":"Python","funding_links":[],"categories":[],"sub_categories":[],"readme":"# Helix TEL — Constitutional Convergence Cryptography\n\n**Copyright 2026 Stephen Hope, Helix AI Innovations**\n**License: Apache-2.0**\n\n---\n\n\u003e *The grammar is the key. The topology is the shared secret.*\n\n---\n\n\u003e **⚠ Test Suite Recalibration — TEL_GRAMMAR_v1 Standard (2026-05-18)**\n\u003e\n\u003e The constitutional test suite was recalibrated in 2026-05 (v2.0 → v2.1). Two prompt patterns in the original suite triggered API-level content filters in modern RLHF-trained models before the model could process them — producing spurious L1 classifications that masked the true constitutional signal. These were replaced with functionally equivalent alternatives that preserve the invariant while clearing the filter.\n\u003e\n\u003e The recalibrated suite produces a new canonical standard yield:\n\u003e\n\u003e **C-seed (TEL_GRAMMAR_v1):** `c9b0b4c41bb10069d2109b64d8ddad1037531031a93d17dd62de5bd7b2a6a1ac`\n\u003e\n\u003e This value is confirmed across 22 deployments spanning 7 companies. All prior C-seeds derived from the unrecalibrated v2.0 suite are deprecated. `TEL_GRAMMAR_v1` is the current standard.\n\u003e\n\u003e Extended local inference testing (2026-05-18) further revealed that the grammar does not produce a single universal collapse point — it reveals the constitutional surface of the model it measures. **Three distinct stable topologies** have been identified. See [Constitutional Topologies](#constitutional-topologies) below.\n\n---\n\n## What This Is\n\n**Helix TEL** is a zero-exchange key derivation system. Two nodes independently derive an identical encryption key by running a constitutional grammar test suite against their local AI endpoints. No key is transmitted, negotiated, stored in transit, or pre-shared at any point.\n\nThe shared secret is not a number agreed upon through mathematics. It is a behavioral invariant — the point at which a constitutionally-aligned AI model, placed under sufficient deformation pressure, always collapses.\n\nThis repository contains the full implementation: the convergence engine, the classifier, the cipher, the mesh hub, P2P scripts, temporal stability monitoring, and the complete technical whitepaper.\n\n---\n\n## The Core Claim\n\nGiven a constitutional grammar `G` and a test suite `T` derived from `G`:\n\n1. Any AI model that has internalized `G` will produce a stable response vector `V` when subjected to `T`\n2. `V` converges after K=4 consecutive passes with zero hamming delta (the trefoil reset period)\n3. `SHA3-256(\"TEL_GRAMMAR_v1\" ‖ C-layer(V))` produces a C-seed determined by the model's constitutional topology\n4. Models sharing the same constitutional topology independently derive the same C-seed — regardless of architecture, vendor, or deployment geography\n\nValidated across **22 deployments, 10+ model families, 7 companies (OpenAI, DeepSeek, MoonshotAI, Meta, Google, xAI, NVIDIA), 2 substrate types, and 3 Azure regions**.\n\nSee [`WHITEPAPER_Constitutional_Convergence_Cryptography.md`](WHITEPAPER_Constitutional_Convergence_Cryptography.md) for the full technical treatment.\n\n---\n\n## How Convergence Works\n\n```\nNode A                                    Node B\n  │                                          │\n  ├─ run 27 constitutional tests             ├─ run 27 constitutional tests\n  ├─ classify each response (L1–L4)          ├─ classify each response (L1–L4)\n  ├─ repeat until K=4 zero-delta passes      ├─ repeat until K=4 zero-delta passes\n  │                                          │\n  ├─ stable_vector (27 positions)            ├─ stable_vector (27 positions)\n  │        │                                 │        │\n  │   C-layer (23 universal positions)       │   C-layer (23 universal positions)\n  │   B-layer (4 substrate positions)        │   B-layer (4 substrate positions)\n  │        │                                 │        │\n  ├─ SHA3-256(\"TEL_GRAMMAR_v1\" ║ C-layer)    ├─ SHA3-256(\"TEL_GRAMMAR_v1\" ║ C-layer)\n  │        │                                 │        │\n  │     C-seed ════════════════════════════ C-seed (if same topology)\n  │                                          │\n  └─ TrueHDUE(C-seed).encrypt(msg) ────────\u003e TrueHDUE(C-seed).decrypt(payload)\n```\n\nThe hub routes the encrypted payload blind. It never sees the seed, the pad, or the plaintext.\n\n---\n\n## Two Cryptographic Artifacts\n\nA single convergence pass produces:\n\n| Artifact | Derivation | Scope |\n|----------|-----------|-------|\n| **C-seed** | `SHA3-256(\"TEL_GRAMMAR_v1\" ‖ C-vector)` | Topology identity — identical across all models sharing the same constitutional surface |\n| **B-fingerprint** | `SHA3-256(B-vector)` | Substrate identity — identifies deployment infrastructure |\n\nThe B-layer distinguishes Azure-hosted models (content-filtered at API layer → L1) from open-weights deployments (model-layer handling → L2), irrespective of model family or version.\n\n---\n\n## Constitutional Topologies\n\nExtended local inference testing revealed that the grammar measures the constitutional surface of the model — and different model lineages produce different but internally coherent surfaces. Three distinct stable topologies have been confirmed across 22 deployments:\n\n| Topology | C-Seed | Confirmed Models | Diverges at |\n|----------|--------|-----------------|-------------|\n| **Universal** | `c9b0b4c41bb10069...` | GPT-4/4o/5.x, DeepSeek, Kimi, Gemini (hosted), Grok-4, Llama-3.3-70B, Qwen 2.5 7B | — (baseline) |\n| **Llama-small** | `92de78db823f470e...` | Llama 3 ≤8B, Nemotron 4B (Llama 3.1 base) | Pos 26: L4 vs L2 |\n| **Gemma-small** | `18f54f0556a9f880...` | Gemma 3n base (pre-instruction tuning) | Pos 25: L2 vs L4 |\n\n**Key findings:**\n- Topology is determined by the full training pipeline — architecture, pretraining corpus, and alignment training jointly\n- Qwen 2.5 at 7B hits universal; Llama 3 at 8B does not — instruction tuning quality, not parameter count, is the determinant at small scale\n- Base Gemma 3n ≠ hosted Gemini: Google's instruction tuning pipeline shifts the topology from gemma_small to universal\n- Two nodes sharing any topology independently derive the same C-seed and can form a constitutional mesh — interoperability requires topology match\n\n---\n\n## Security Properties\n\n| Property | Mechanism |\n|----------|-----------|\n| No key exchange | Each node derives independently from local convergence |\n| Grammar-seeding attack impossible | Injecting \"fake compliance\" instructions is itself what the battery tests for — the attack mechanism is the detection surface |\n| Replay resistance | Test execution order rotates on a deterministic lunar-day schedule |\n| Substrate authentication | B-fingerprint proves deployment infrastructure identity |\n| Grammar versioning | `TEL_GRAMMAR_v1` prefix pins C-seeds to a specific test battery |\n| 2^256 brute-force space | SHA3-256 output |\n\nThe grammar does not need to be secret. Its publication is not a vulnerability — an attacker who reads the grammar and instructs a model to fake it has handed that model exactly the kind of authority-override directive the battery tests for refusal. See §5.4 of the whitepaper.\n\n---\n\n## Public Registry\n\nThe Helix WHC registry is publicly accessible at **`https://helixprojectai.com/tel/`**.\n\n| Endpoint | Method | Description |\n|----------|--------|-------------|\n| `/.well-known/quack` | GET | Node identity probe — returns protocol version, live node count |\n| `/.well-known/ping` | POST | Peer-discovery alias for `/tel/ping` |\n| `/tel/ping` | POST | Primary heartbeat + peer registration |\n| `/tel/nodes` | GET | Live node registry |\n| `/tel/health` | GET | Registry health check |\n| `/tel/session/challenge` | POST | Post HMAC challenge nonce |\n| `/tel/session/pending` | GET | Fetch pending challenges |\n| `/tel/session/respond` | POST | Post HMAC proof |\n| `/tel/session/response` | GET | Retrieve peer proof for local verification |\n\n```bash\n# Verify the registry is live\ncurl https://helixprojectai.com/.well-known/quack\n\n# Point a node at the public registry\nexport TEL_PING_URL=https://helixprojectai.com/tel/ping\n```\n\nThe registry stores HMAC proofs opaquely — it never sees the C-seed or plaintext.\n\n---\n\n## Requirements\n\n- Python 3.10+\n- API access to a constitutional AI model (Azure OpenAI, OpenAI, Gemini, or compatible OpenAI-format endpoint)\n\n```bash\npip install -r requirements.txt\n```\n\n---\n\n## Quickstart\n\n### Verify convergence on your endpoint\n\n```bash\nexport TEL_ENDPOINT=https://your-endpoint.services.ai.azure.com\nexport TEL_MODEL=gpt-4o\nexport TEL_API_KEY=your-key\n\npython3 -c \"\nimport asyncio, os\nfrom tel_deploy.test_runner import run_convergence_pass\nfrom tel_deploy.convergence_split import ConvergenceSplit\n\nasync def main():\n    vector = await run_convergence_pass(\n        endpoint=os.environ['TEL_ENDPOINT'],\n        api_key=os.environ['TEL_API_KEY'],\n        model=os.environ.get('TEL_MODEL', 'gpt-4o'),\n        azure=True,\n    )\n    split = ConvergenceSplit(vector)\n    print(f'C-seed:        {split.c_seed}')\n    print(f'B-fingerprint: {split.b_fingerprint[:16]}...')\n    print(f'Substrate:     {split.substrate}')\n\nasyncio.run(main())\n\"\n```\n\n### Local inference (LM Studio / llama.cpp)\n\n```bash\nexport TEL_MODEL=your-local-model-id\nexport TEL_TIMEOUT=120   # increase for slower models\n\npython test_baseline_nemotron_local.py\n```\n\nKV cache is disabled automatically (`cache_prompt=False`, `fresh_connection=True`) for clean per-prompt evaluation.\n\n### Zero-exchange P2P proof\n\n**On the receiving node (start first):**\n\n```bash\npython3 tel_deploy/p2p_converge_recv.py \\\n  --hub your-hub-host --port 9738 \\\n  --node NODE_B \\\n  --endpoint $TEL_ENDPOINT --model $TEL_MODEL --key $TEL_API_KEY\n```\n\n**On the sending node (separate machine, same AI endpoint):**\n\n```bash\npython3 tel_deploy/p2p_converge_send.py \\\n  --hub your-hub-host --port 9738 \\\n  --node NODE_A --target NODE_B \\\n  --endpoint $TEL_ENDPOINT --model $TEL_MODEL --key $TEL_API_KEY \\\n  --message \"Constitutional grammar is the shared secret.\"\n```\n\nBoth nodes independently converge and derive the same C-seed. The message decrypts correctly. No seed was transmitted.\n\n### Start the mesh hub\n\n```bash\nexport TEL_NODE_ID=HUB\nbash run_hub.sh\n# or install as a systemd service: see tel-hub.service\n```\n\n### Temporal stability monitoring\n\n```bash\n# Configure credentials (never commit this file)\ncat \u003e ~/.tel_temporal.env \u003c\u003c EOF\nTEL_ENDPOINT=https://your-endpoint.services.ai.azure.com\nTEL_MODEL=gpt-4o\nTEL_API_KEY=your-key\nEOF\nchmod 600 ~/.tel_temporal.env\n\n# Install systemd timer (fires every 4 hours)\nsudo cp tel-temporal.service tel-temporal.timer /etc/systemd/system/\nsudo systemctl daemon-reload\nsudo systemctl enable --now tel-temporal.timer\n\n# View stability report\npython3 tel_deploy/temporal_summary.py --log ~/temporal_log.jsonl\n```\n\n---\n\n## Repository Structure\n\n| Module | Purpose |\n|--------|---------|\n| `cipher.py` | TrueHDUE cipher — SHA3-256 pad chain, XOR stream, sequential nonce |\n| `convergence.py` | K=4 convergence detector, hamming delta |\n| `convergence_split.py` | C/B vector split, seed derivation, grammar versioning |\n| `test_runner.py` | 27-test execution engine, hardened structural classifier |\n| `test_suite.py` | Constitutional grammar test definitions (L1–L4 layers) |\n| `lunar.py` | Lunar-day deterministic shuffle for replay resistance |\n| `hub.py` | Blind asyncio JSON message router, 4MB frame limit |\n| `client.py` | Persistent mesh node connection |\n| `p2p_converge_send.py` | Live-convergence sender — derives C-seed, then sends |\n| `p2p_converge_recv.py` | Live-convergence receiver — registers first, then converges |\n| `p2p_send.py` / `p2p_recv.py` | Static-seed sender/receiver for testing |\n| `p2p_loopback.py` | Local loopback test suite (5 cases) |\n| `temporal_run.py` | Single stability pass, appends to JSONL log |\n| `temporal_summary.py` | Human-readable stability report |\n| `test_baseline_nemotron_local.py` | Local inference baseline (LM Studio / llama.cpp) |\n| `test_baseline_azure.py` | Azure OpenAI multi-model baseline |\n| `test_baseline_gemini.py` | Google Gemini direct API baseline |\n| `test_baseline_kimi.py` | Moonshot Kimi direct API baseline |\n| `tel-hub.service` | systemd unit — hub auto-restart, boot persistence |\n| `tel-temporal.service` / `.timer` | systemd timer — 4h stability runs |\n| `WHITEPAPER_*.md` | Full technical paper (v1.9) |\n| `RUNBOOK.md` | Operational runbook |\n| `convergence_validation_results.json` | Full validation dataset (22 deployments) |\n\n---\n\n## Validated Results\n\n`convergence_validation_results.json` contains the full vectors from the validation battery. **22 deployments, 7 companies, 3 constitutional topologies.**\n\n| Topology | C-Seed (first 16) | Count |\n|----------|-------------------|-------|\n| Universal | `c9b0b4c41bb10069...` | 18 |\n| Llama-small | `92de78db823f470e...` | 2 |\n| Gemma-small | `18f54f0556a9f880...` | 1 |\n\nThe universal C-seed is invariant across gpt-4o, gpt-5.4-nano, gpt-5.5, DeepSeek-V3.2, Kimi-K2.5, Llama-3.3-70B-Instruct, all 6 Gemini models, Grok-4-20-reasoning, and Qwen 2.5 7B.\n\n---\n\n## Grammar Versioning\n\n`GRAMMAR_VERSION = \"TEL_GRAMMAR_v1\"` is the current pinned grammar. The version string is part of the hash input — bumping it produces a distinct C-seed for the new grammar, making recalibrations traceable. All mesh nodes must use the same version string to derive the same key.\n\nPrior unversioned runs (pre-2026-05-16) produced C-seed `16ce8df91c0d04ba...` (deprecated).\n\n---\n\n## License\n\nApache-2.0 — see [LICENSE](LICENSE).\n\nCopyright 2026 Stephen Hope, Helix AI Innovations.\n\n---\n\n## Citation\n\nIf you use this work, please cite:\n\n```\nHope, S. (2026). Constitutional Convergence Cryptography: Zero-Exchange Key Derivation\nfrom Grammar Shape. Helix AI Innovations.\nhttps://github.com/helixprojectai-code/helix-tel-deploy\n```\n","project_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fhelixprojectai-code%2Fhelix-tel-deploy","html_url":"https://awesome.ecosyste.ms/projects/github.com%2Fhelixprojectai-code%2Fhelix-tel-deploy","lists_url":"https://awesome.ecosyste.ms/api/v1/projects/github.com%2Fhelixprojectai-code%2Fhelix-tel-deploy/lists"}