SOP-AIMS-AGENTOS-001 v1.3 | All Phases | ISO/IEC 42001:2023

AgentOS Control Centre

Complete eagle-eye view of the UNUS London Agentic Operating System — 7 agent groups, 13 registered agents, local model inference architecture, and all four memory sub-layers.

Wiki Pages

—

Sub-layer 4E

HITL Approved

—

Total all-time

HITL Pending

—

Awaiting review

Learning Events

—

experience_log

Prompt Examples

—

Active in registry

System Health

—/7

Checking...

Agent Execution — 7 Days

Live

Wiki Status

Self-Learning Summary

Recent Activity

◈

Loading...

Quick Actions

Infrastructure Monitoring

System Health

Live status of all AgentOS infrastructure. Auto-pinged on load and every refresh.

Checking...

Infrastructure Reference

Service	Address	Role in AgentOS	Notes
CouchDB	100.96.35.122:5984	Sub-layer 4E — Semantic Wiki	Docker: obsidian-sync \| DB: unus-london-brain
n8n	n8n-shach.unusailabs.com	Orchestration Layer	Self-hosted VPS \| Schedule Trigger pattern only
Supabase / PostgreSQL	Configured in Settings	Sub-layers 4C + 4D	agent_experience_log, prompt_registry, approval_queue, kb_documents
Gemma4 26b — AIMS-SYS-005	Local inference — UNUS VPS	Primary inference — Report, Intake, Intelligence agents	Google DeepMind \| Self-hosted \| No data exits infrastructure
Gemma4 e4b — AIMS-SYS-006	Local inference — UNUS VPS	Fast classification — QA, Audit, Memory, Knowledge agents	Google DeepMind \| Self-hosted \| Lightweight — fast throughput
Qwen3-coder-next — AIMS-SYS-007	Local inference — UNUS VPS	Code & SQL generation — Schema, Workflow agents	Alibaba / Qwen Team \| Self-hosted \| Code-specialised model
Mistral (local) — AIMS-SYS-002	Local inference — UNUS VPS	Document Agent — PDF extraction & classification	Mistral AI \| Self-hosted weights \| No API data transfer
Anthropic Claude API — AIMS-SYS-001/004	api.anthropic.com/v1	Content Studio pipeline (SYS-004) \| Dev tooling only	Build-phase only — not used for client data inference in production. DPA assessment in progress.
Obsidian / LiveSync	Local machine	Sub-layer 4A — Knowledge Authoring	Syncs to unus-london-brain \| E2E encryption OFF (compensating control logged)
Qdrant	Self-hosted	Sub-layer 4B — Vector Retrieval	RAG queries from all agents \| Updated by Knowledge Ingestion Agent

SOP-AIMS-AGENTOS-001 v1.3 §8 | 13 Registered Agents

Agent Registry

All agents across all AIMS system IDs. Authority boundaries, risk classifications, and real-time status. Local model architecture active — AIMS-SYS-005/006/007 replace API inference for production workloads.

Group 1 — Client Onboarding Pipeline | AIMS-SYS-005/006/007 | ISO/IEC 42001:2023 Cl. 8.2

Client Pipeline

End-to-end client onboarding: sector classification → document processing → schema deployment → QA validation → report generation. All five agents operational on local inference.

Pipeline Runs

—

All-time onboardings

Documents Processed

—

kb_documents total

Reports Generated

—

report_history

Avg QA Confidence

—

Schema validation score

Pipeline Entry Points

Open Matter Checklist → Open Sign-Off Portal →

Stage 1 — Intake Agent | AIMS-SYS-005 — Gemma4 26b (local)

Client Sector Classification

Live

Classifies inbound client sector (Legal / Finance / Healthcare / Manufacturing / ITSM). Reads client intake form data from kb_documents. Routes to appropriate Schema Agent. Confidence-gated — sub-threshold outputs routed to HITL queue before proceeding.

Inference: Local READ — intake tables Webhook: /webhook/client-intake

↓

Stage 2 — Document Agent | AIMS-SYS-006 — Mistral (local)

PDF Compliance Document Processing

Live

Extracts and classifies uploaded compliance documents. Processes PDFs through local Mistral inference — no data transmitted externally. Writes classified records to kb_documents. Triggers Knowledge Ingestion Agent for Qdrant embedding on APPROVED output.

Inference: Local READ/WRITE — kb_documents Confidence gate: 0.75

↓

Stage 3 — Schema Agent | AIMS-SYS-007 — Qwen3-coder-next (local)

PostgreSQL Schema Deployment

Live

Selects and deploys sector-appropriate compliance schema to the client Supabase instance. Uses Qwen3-coder-next for SQL generation accuracy — code-specialised model. Creates client-specific tables, RLS policies, views, and stored procedures. Triggers QA Agent on completion.

Inference: Local WRITE — schema layer RLS enforced post-deploy

↓

Stage 4 — QA Agent | AIMS-SYS-006 — Gemma4 e4b (local)

Schema Validation — ISO/ITIL Compliance Check

Live

Validates deployed schema against ISO 9001:2015, ISO/IEC 27001:2022, and ITIL 4 rules. Checks: foreign key integrity, RLS policy coverage, audit trigger presence, mandatory fields, and regulatory clause mapping. Sub-threshold outputs flagged to MD HITL queue before client access is granted.

Inference: Local READ — schema validation HITL gate: confidence < 0.85

↓

Stage 5 — Report Agent | AIMS-SYS-005 — Gemma4 26b (local)

Compliance Report Generation

Live

Generates SRA Readiness Reports, Client Matter Checklists, and FCA Evidence Packs. Pulls live data from the deployed client schema. All outputs require principal sign-off via the Report Sign-Off Portal before filing to report_history. HITL quality gate — override category captured for AIMS metrics.

Inference: Local WRITE — report_history Principal sign-off required

Document ID	Title	Sector	Status	Confidence	Created
Loading kb_documents...

Publisher Pipeline

GitHub deployment and post-deploy automation. Both workflows are live and webhook-triggered.

GitHub Deploys

—

Last: —

Publisher Runs

—

Sitemap + Slack + Audit

Webhook Status

LIVE

Both endpoints active

Group 2 — GitHub Deploy Agent

Site Deployment Trigger

Live

Triggers a full GitHub Pages deployment. Posts to the GitHub Deploy Agent webhook which initiates the build-and-deploy pipeline. Use after any content or configuration changes that require a live site update.

Webhook: n8n-shach.unusailabs.com/webhook/github-deployment

Method: POST

Last Run: —

Group 2 — Publisher Agent

Post-Deploy Automation

Live

Runs post-deployment tasks: sitemap generation, Slack notifications, audit log entry. Fires automatically after GitHub Deploy or can be triggered independently to re-run post-deploy tasks without a full redeployment.

Webhook: n8n-shach.unusailabs.com/webhook/publisher-agent

Method: POST

Last Run: —

Group 3 — Regulatory Intelligence Layer | AIMS-SYS-005 — Gemma4 26b (local)

Regulatory Intelligence

Continuous monitoring of SRA, FCA, CQC, ISO, and ITIL publication feeds. AI-classified by impact and urgency. Schema impact assessed per active client. Findings routed to MD HITL queue.

Publications Monitored

—

Last 90 days

High Impact Findings

—

Awaiting schema review

Last Scan

—

Scheduled: daily 06:00

Client Schemas Assessed

—

Active client count

Open Reg Intel Portal →

AIMS-SYS-005 — Gemma4 26b | Group 3 | Schedule: Daily 06:00 UTC

Regulatory Feed Monitor — 5 Jurisdictions

Live

Continuously monitors publication feeds across five regulatory bodies. On detection of a new publication, classifies by impact tier (Critical / High / Medium / Low) and assesses which active client schemas may require updates. Generates a structured finding record and routes HIGH/CRITICAL findings to the MD HITL queue within 15 minutes of publication.

SRA: Code of Conduct | Thematic Reviews | Warning Notices | Guidance

FCA: Handbook Updates | Policy Statements | Dear CEO Letters | SMCR Guidance

CQC: Key Lines of Enquiry | Inspection Frameworks | Safety Alerts

ISO: ISO 9001 | ISO/IEC 27001 | ISO/IEC 42001 | Amendment Notices

ITIL: ITIL 4 Practice Guides | Axelos Updates

Impact Classification — Current Cycle

◈

Click Refresh to load findings

Schema Impact Assessment

◈

Click Refresh to load schema assessments

Recent Regulatory Findings

Publication	Body	Impact	Schema Action	Status	Detected
Click Refresh to load regulatory findings

AIMS-SYS-008 | Sub-layer 4E | Group 6

Knowledge Graph

Semantic wiki, ingestion pipeline, confidence scores, regulatory mappings, and lint reports.

Total Wiki Pages

—

unus-london-brain

Published

—

MD approved

Draft

—

Pending review

Review Required

—

Low confidence

Classifications

Confidence Distribution

Recent Wiki Pages

Page ID	Title	Status	Confidence	Classification	Created
Loading...

Latest T-03 Lint Report

◈

Click Load to fetch the latest wiki lint report

Unprocessed in raw/

Document ID	Title	Created	Action
Click Refresh

SOP-AIMS-AGENTOS-001 v1.2 | AIMS-SYS-006 | Sub-layer 4C

Memory & Learning

Episodic memory (agent_experience_log), procedural memory (prompt_registry), and the full self-learning loop status. ISO 9001:2015 Cl. 10.3 Continual Improvement evidence.

Total Experiences

—

agent_experience_log

Approval Rate

—%

APPROVED / total

Active Prompts

—

prompt_registry

Avg Confidence

—

All agents all-time

Self-Learning Loop — SOP-AIMS-AGENTOS-001 v1.2 §6.1

Step 1

Task Execution

Agent executes task. Confidence score computed. Output written to pipeline execution log.

Step 2

HITL Checkpoint

MD reviews output. Marks APPROVED, CORRECTED, or REJECTED. Adds correction_detail where applicable.

Step 3

Memory Manager

Triggered by HITL signal. Writes to agent_experience_log. Correction notes to Obsidian if applicable.

Step 4

Re-embedding

If Obsidian modified: Knowledge Ingestion Agent triggers Qdrant re-embedding cycle.

Step 5

Prompt Promotion

If APPROVED + confidence ≥ 0.90 + MD countersignature: promotes to prompt_registry as few-shot example.

Step 6

Enriched Next Run

Next invocation queries Qdrant + prompt_registry + experience_log. Agent performs measurably better.

HITL Decision Breakdown

Correction Categories

Correction Rate by Agent — SOP §10.1 Trigger: >30% flags review

Agent Role	AIMS System	Total Events	Approved	Corrected	Rejected	Correction Rate	Status
Loading from agent_experience_log...

Active Prompt Registry — Procedural Memory

Agent Role	Task Type	Sector	Example Label	Use Count	Approved
Loading prompt_registry...

SOP-AIMS-AGENTOS-001 v1.3 §3.3 | All Groups

Workflow Control

Trigger and monitor all AgentOS workflows across all seven groups. Configure credentials in Settings before triggering.

ℹ

n8n API Key required for Workflow ID triggers (Groups 5–6). Webhook triggers (Groups 2, 4) fire directly. Placeholder workflows are pre-wired and will activate when URLs are added in Settings.

Group 2 — Publisher Pipeline

Webhook | Group 2

GitHub Deploy Agent

Live

POST to GitHub Deploy webhook. Initiates full site deployment pipeline.

URL: n8n-shach.unusailabs.com/webhook/github-deployment

Last: —

Webhook | Group 2

Publisher Agent

Live

POST to Publisher Agent webhook. Runs sitemap, Slack notification, and audit logging post-deploy tasks.

URL: n8n-shach.unusailabs.com/webhook/publisher-agent

Last: —

Group 3 — Intelligence Layer

Webhook | Group 3 | AIMS-SYS-005

Regulatory Intelligence Agent

Building

Scans SRA, FCA, and CQC publication feeds. Assesses impact against all active client schemas. Routes changes to MD HITL queue.

URL: Not configured

Group 4 — Knowledge Management

Webhook | Group 4 | AIMS-SYS-006

Knowledge Ingestion Agent

Live

Triggers Obsidian → Qdrant re-embedding cycle. Runs nightly automatically and on CORRECTED HITL signal. Manual trigger forces an immediate re-embed of all modified vault content.

URL: n8n-shach.unusailabs.com/webhook/knowledge-ingestion-trigger

Last: —

Group 5 — Memory & Learning

Workflow ID: 1jz64WLTaixZwxG3 | AIMS-SYS-006

Memory Manager Agent

Polling

Polls approval_queue every 1 minute. Converts HITL decisions into governed memory writes across agent_experience_log and prompt_registry. Three branches: APPROVED / CORRECTED / REJECTED.

Schedule: Every 1 minute

Status: —

ID: 1jz64WLTaixZwxG3

Group 6 — Knowledge Architect Agent (AIMS-SYS-008)

T-01 | ATqQNoeODsMePftf

CouchDB Poll — Automated Ingestion

Scheduled

Polls CouchDB every 5 minutes for UNPROCESSED documents in raw/. Processes up to 5 per run through the full AIMS-SYS-008 pipeline. Confidence gate at 0.75.

Schedule: Every 5 minutes

Status: —

Last: —

T-02 | ZHfQVQ5y1jLHTeAr

Manual Ingestion — Priority Document

MD Only

Priority-process a specific document in CouchDB raw/ path. Document ID must begin with raw/.

Status: —

Document ID (CouchDB _id)

T-03 | KqRbXhZNwBKmaeye

Wiki Lint — Weekly Audit

Weekly

Weekly audit (Monday 07:00) of all UNUS_Wiki_* pages. Checks orphans, stale escalations (>7 days), stale drafts (>30 days), outdated regulatory mappings. Sends MD report email.

Schedule: Monday 07:00 UTC

Status: —

Last Issues: —

Group 7 — Orchestration Controller

Group 7 | Phase 4 Step 6D

Orchestration Controller

Not Built

The central n8n workflow that receives task requests from all upstream triggers, routes to the correct agent based on task classification, manages confidence thresholds, and feeds the HITL approval_queue on sub-threshold outputs. This is the single entry point that ties all agents into one governed system.

URL: To be built in Step 6D

Functions: Task routing | Confidence management | HITL queue management | Agent authority enforcement

Group 1 — Client Onboarding Pipeline

Group 1 | Phase 5

Client Onboarding Pipeline — 5 Agents

Building

Five-stage client onboarding: Intake Agent → Document Agent → Schema Agent → QA Agent → Report Agent. All placeholder controls are pre-wired.

Human-In-The-Loop | approval_queue + CouchDB

HITL Review Queue

All outputs pending MD review. Approval gates both wiki publication (AIMS-SYS-008) and memory writes (AIMS-SYS-006). This is the primary input mechanism for the self-learning system.

Pending Review

—

Across all agents

Approved Today

—

Memory writes triggered

Avg Queue Age

—

Hours pending

◈

Loading HITL queue...

ISO/IEC 42001:2023 | ISO 9001:2015 | ISO/IEC 27001:2022 | ITIL 4

Compliance & Audit

Regulatory readiness scores, standards compliance matrix, SOP review trigger monitoring, and the monthly learning performance report required by ISO/IEC 42001:2023 Cl. 9.3.

ISO/IEC 42001 Score

—

Cl. 8.4 / 8.5 / 9.1 / 9.3

ISO 9001 CSI Score

—

Cl. 10.3 Continual Improvement

HITL Compliance

—%

% outputs reviewed

Audit Records

—

Immutable log entries

Standards Compliance Matrix

v1.3

STANDARD

STATUS

IMPLEMENTATION

ISO/IEC 42001:2023

Cl. 8.4

AIMS-SYS-008 registered with authority boundaries, risk classification, and HITL requirements. Deployed in production per v1.3 amendment.

ISO/IEC 42001:2023

Cl. 8.5

Governed write-back protocol (v1.2 §6) defines operational evolution within HITL boundaries. No ungoverned behavioural drift permitted.

ISO/IEC 42001:2023

Cl. 9.1

agent_experience_log provides AI performance dataset. Confidence scores, HITL override rates, and correction frequencies are computable from live data.

ISO/IEC 42001:2023

Cl. 9.3

Monthly learning performance report generated from agent_experience_log. Covers approval rates, correction categories, prompt_registry changes, confidence trends.

ISO 9001:2015

Cl. 10.3

Self-learning loop (v1.2 §6) is the technical implementation of CSI. Corrections → prompt_registry promotion = documented, traceable improvement cycle.

ISO 9001:2015

Cl. 7.5

Wiki pages governed as documented information assets. DRAFT → PUBLISHED lifecycle with HITL gate enforces document control.

ISO/IEC 27001:2022

Cl. A.12.4

agent_experience_log is immutable (no UPDATE/DELETE policies). Every memory write traceable to specific HITL decision, reviewer, and timestamp.

ISO/IEC 27001:2022

Cl. 8.1

E2E encryption OFF exception recorded as compensating control (LiveSync agent access requirement). Logged in risk register.

ITIL 4

KM Practice

AIMS-SYS-008 operationalises Knowledge Management. DRAFT/PUBLISHED lifecycle maps to ITIL Knowledge Management status model.

ITIL 4

CSI Practice

Experience → HITL → prompt_registry cycle = ITIL CSI loop: measure (confidence) → analyse (correction categories) → improve (prompt_registry) → measure again.

SOP Review Trigger Monitor — §10.1

◈

Loading SOP review trigger status...

Monthly Learning Performance Report — ISO/IEC 42001:2023 Cl. 9.3

◈

Click Generate Report to produce the monthly performance summary

Audit Trail — kb_audit_logs

Timestamp	Agent	Action	Record ID	Outcome
Loading kb_audit_logs from Supabase...

AIMS-SYS-008 | log.md | CouchDB

Activity Log

Full agent execution log from CouchDB. Every ingestion, every memory write, every workflow trigger.

Agent Execution Log

◈

Loading activity log...