Deployment & Roll-Out Guide

This guide describes how RABS is built, shipped, observed, and iterated in production. It blends standard DevOps mechanics with the human-in-the-loop safety plan required during RABS’ first 24 months.

---

1. Hosting & Environments

Stage	URL / Location	Data Source	Primary Purpose
DEV	dev.rabs.local	Synthetic seed data	Rapid local builds, no PII
STAGE	staging.rabs.company.lan	Scrubbed prod snapshot	Feature-flag dry-runs, load testing
PROD	app.rabs.company.au	Live data	Real users, gated roll-outs

The core API and Brainframe are deployed to the on-prem LUCAS Kubernetes (k3s) cluster.

---

2. CI/CD Pipeline

git push
   │
   ▼
CI Build  →  Unit Tests & Lint  →  Container Scan
   │
   ▼
Integration Tests (STAGE DB snapshot)
   │
   ▼
Manual QA Gate + Feature-Flag Review
   │
   ▼
Security Scan
   │
   ▼
Canary Deploy  (10 % of PROD traffic)
   │
   ▼  (observe ≥ 48 h)
Full Rolling Deploy to PROD

---

3. Human-in-the-Loop Monitoring (first 24 months)

Phase (months)	Review Mode	Roll-out Target
0 – 3	Review-All – every decision requires manual approval before execution	0 – 10 % users
4 – 12	Pre-Exec Sampling – high-risk decisions block for review	25 – 50 % users
13 – 24	Post-Exec Audit – all actions execute immediately; high-risk ones are audited after	100 % users

Reviewers use the Cognitive Alignment UI. A misalignment rate above 5 % automatically rolls back the feature flag and sends an alert.

---

4. Observation Windows

No further roll-outs are permitted until a change completes its minimum observation window and KPIs remain stable.

Change Type	Minimum Window	Key Metrics to Watch
Schema migration	7 days	DB errors, P99 latency, replica lag
New LLM model	4 days	Token cost, alignment %, response latency
Voice SDK upgrade	3 days	Call drop rate, Word-Error-Rate (WER)
Minor UI tweak	24 h	Front-end error rate, user feedback

---

🔗 Related Docs

• 01_System_Independence_Model.md
• ../../02_Brainframe_Cognitive_Architechture/02_Reasoning_&Memory/02_Cognitive_Alignment&_Review.md