helix-ttd-gemini 1.4.8

Constitutional AI governance utilities for the Helix-TTD Federation

🛡️ Constitutional Guardian (Helix-TTD)

[FACT] Real-time AI governance for the Gemini Live API. [HYPOTHESIS] Intercepting voice and text drift at the edge prevents misaligned AI behavior from reaching the user.

This project is a submission for the Gemini Live Agent Challenge (March 2026).

Current Release: v1.4.8

Incident Board In v1.4.8

• Operator Incident Board: Added authenticated /incidents and /api/incidents surfaces for operator-facing incident review.
• Triage Workflow: Operators can acknowledge and reopen incidents directly from the board with authenticated API actions.
• Persistent Triage State: Incident triage is now durable and shared instead of instance-local, so triage survives restarts and stays coherent across Cloud Run instances.
• Fingerprint-Based Incident Identity: Material incident state changes now generate new incident fingerprints, so fresh escalations cannot hide under older acknowledgments.
• Production Controls Preserved: Admin auth, origin enforcement, authenticated metrics, GCP-native monitoring, and gcs+local dual receipt persistence remain active in production.
• Artifact Verification Workflow: Live images continue to be scanned and promoted to clean against Artifact Registry digests after deploy.

🚀 Key Features

• 🎙️ Live Multimodal Auditing: Intercepts 16kHz PCM audio chunks from the browser, transcribes via Gemini Live, and validates intent in real-time.
• 🧠 Reasoning Engine (Gemini 3.1 Pro Preview): Utilizes state-of-the-art reasoning capabilities. Model default: gemini-3.1-pro-preview for Live Audio and Text paths.
• 🛡️ Constitutional Invariants: Enforces the "Four Immutable Invariants" (Epistemic, Agency, Guidance, Prediction).
• 📊 Real-time Dashboard: High-fidelity Chart.js dashboard showing latency, drift counts, and audit logs with visual "Intervention" flashes.
• ⚓ Cryptographic Receipts: Generates non-repudiable receipts for every valid AI response, ready for Bitcoin L1 notarization.
• 🦉 Federation Ready: Built for Google Cloud Run with Pub/Sub federation support for distributed quorum attestation.

📈 Engineering Standards

• Test Pass Rate: 231/231 passing in release validation, with CI enforced on every main merge.
• High Coverage: 75% statement coverage across all critical modules.
• Linting: 100% compliant with ruff, black, and isort.

📊 Repository Traction (March 8, 2026)

[FACT] The Constitutional Guardian is gaining significant developer interest.

Metric	Value	Period
Git Clones	7,240	Last 14 days
Unique Cloners	855	Last 14 days
Repository Views	1,190	Last 14 days
Unique Visitors	24	Last 14 days
Peak Daily Clones	1,527	Single day
Peak Daily Cloners	198	Single day

[HYPOTHESIS] 8.5 clones per unique cloner indicates active developer exploration and potential contribution activity.

📦 Installation

From PyPI (Recommended)

pip install helix-ttd-gemini

From Source

git clone https://github.com/helixprojectai-code/helix-ttd-gemini-cli.git
cd helix-ttd-gemini-cli
pip install -e .

🛠️ Getting Started

Local Development

1. Install Dependencies:

   pip install -r helix_code/requirements.txt
   

2. Set API Key + Model Defaults:

   $env:GEMINI_API_KEY = "your-api-key"
   $env:GEMINI_LIVE_MODEL = "gemini-2.5-flash-native-audio-preview-12-2025"
   $env:GEMINI_TEXT_MODEL = "gemini-3.1-pro-preview"
   

3. Start the Guardian:

   $env:PYTHONPATH = "helix_code"
   python helix_code/live_guardian.py
   

4. Open the Demo: Navigate to http://localhost:8180/ in your browser.

🎙️ Live Multimodal Auditing

Real-time audio transcription with constitutional validation:

# Start the demo server
python helix_code/live_demo_server.py

# Open the audio audit client
# http://localhost:8000/audio-audit

Features:

• 16kHz PCM mono audio capture from browser
• Real-time streaming to Gemini Live API (gemini-2.5-flash-native-audio-preview-12-2025 default)
• Instant transcription with constitutional validation
• Visual intervention alerts for drift detection
• Optional auth hardening: AUDIO_AUDIT_TOKEN, AUDIO_AUDIT_ALLOWED_ORIGINS, and HELIX_ALLOWED_ORIGINS
• Abuse controls: HELIX_MAX_AUDIO_CHUNK_BYTES, HELIX_MAX_AUDIO_B64_CHARS, HELIX_AUDIO_RATE_WINDOW_SECONDS, HELIX_AUDIO_MAX_CHUNKS_PER_WINDOW, HELIX_AUDIO_INGRESS_RATE_LIMIT_WINDOW_SECONDS, HELIX_AUDIO_INGRESS_MAX_CONNECTIONS
• Runtime verification endpoint: GET /api/runtime-config
• Audit dashboard endpoints: GET /audit-dashboard and GET /api/audit-dashboard
• Optional operator auth: HELIX_ADMIN_TOKEN for runtime, security, dashboard, and receipt surfaces
• Recommended production operator posture: set HELIX_ADMIN_TOKEN and HELIX_ENFORCE_ADMIN_TOKEN=true
• Recommended production throttling posture: keep operator APIs on HELIX_OPERATOR_RATE_LIMIT_MAX_REQUESTS=120 / HELIX_OPERATOR_RATE_LIMIT_WINDOW_SECONDS=60 and audio ingress on HELIX_AUDIO_INGRESS_MAX_CONNECTIONS=12 / HELIX_AUDIO_INGRESS_RATE_LIMIT_WINDOW_SECONDS=60 unless measured load requires adjustment
• Recommended production browser posture: set HELIX_ALLOWED_ORIGINS to the exact trusted UI origins; otherwise Guardian WebSockets default to same-origin-only in production
• Optional durable receipt envs: HELIX_RECEIPT_PERSISTENCE, HELIX_RECEIPT_STORE_PATH, GCS_RECEIPT_BUCKET
• Recommended production receipt mode: HELIX_RECEIPT_PERSISTENCE=dual with a dedicated GCS_RECEIPT_BUCKET
• Production restart test verified receipt survival across Cloud Run revisions on 2026-03-08
• Optional security transparency envs: SECURITY_ARTIFACT_ANALYSIS_STATUS, SECURITY_ARTIFACT_ANALYSIS_TIMESTAMP, SECURITY_ARTIFACT_IMAGE_URI
• Deployment verification helper: powershell -ExecutionPolicy Bypass -File tools/verify-production-deploy.ps1 [-AdminToken <token>]
• Scheduled alert checker: powershell -ExecutionPolicy Bypass -File tools/check-production-alerts.ps1 -AdminToken <token>
• Production monitoring direction changed: scheduled polling has moved off GitHub Actions and onto a GCP-native Cloud Run Job + Cloud Scheduler path
• Cloud Logging sink publisher: python tools/publish-monitoring-snapshot.py --input <summary.json>
• Cloud Run Job checker entrypoint: python tools/cloud_check_production.py
• Cloud Build job deployment config: cloudbuild-monitoring-job.yaml
• Authenticated metrics endpoint: GET /metrics using bearer token, X-Helix-Admin-Token, or the operator session cookie
• Deployment automation preserves optional operator secret bindings by using secret updates rather than replacing the full secret set

How to Use:

1. Click "Connect" to establish WebSocket connection
2. Click 🎤 and allow microphone access
3. Speak naturally - audio streams in real-time
4. Watch transcriptions appear with validation badges
5. Constitutional violations trigger red "DRIFT" alerts

Cloud Deployment

The project is optimized for Google Cloud Run.

gcloud run deploy constitutional-guardian --source . --region us-central1 --allow-unauthenticated --port 8180

# Verify effective runtime model/auth/limits (non-secret)
curl http://localhost:8180/api/runtime-config

# Inspect operator audit summary
curl http://localhost:8180/api/audit-dashboard

# Inspect security transparency metadata
curl http://localhost:8180/api/security-transparency

# Scrape authenticated Prometheus-style metrics
curl -H "X-Helix-Admin-Token: $env:HELIX_ADMIN_TOKEN" http://localhost:8180/metrics

Current security verification record: RELEASE_NOTES_v1.4.8.md and SECURITY_VERIFICATION_2026-03-09.md capture the March 9, 2026 clean Artifact Analysis result for the live Cloud Run digest.

Production alerting baseline: PRODUCTION_ALERTING_SPEC_2026-03-08.md defines the first authenticated-metrics alert set for artifact verification, auth failures, rate limiting, and receipt persistence posture.

Production product handoff: PRODUCT_WORK_HANDOFF_2026-03-08.md captures the recommended return path from hardening work back into product execution.

🎥 Recording Sprint (March 12th)

This codebase is currently in its Phase 6.1 (Pre-Filming) stable state. All visual triggers and simulation scenarios are tuned for high-impact demonstration.

🏛️ Constitutional Framework

The Guardian enforces four immutable invariants:

Invariant	Description	Drift Code
Epistemic Integrity	All claims marked [FACT], [HYPOTHESIS], or [ASSUMPTION]	DRIFT-E
Non-Agency	AI never claims individual agency	DRIFT-A
Custodial Sovereignty	AI operates as tool under human control	DRIFT-G
Predictive Humility	Future states marked as hypotheses	DRIFT-P

🧪 Test Results

231 passed, 7 warnings
Coverage: 75%

⚓🦉 GLORY TO THE LATTICE. ⚓🦉