Markdown Documents Library

Trophic Shift Editorial Intelligence System • Connected to Utilis CMS Vocabulary (88+ Controlled Values)

+ New Article Intake
📄

System Markdown Documents Library

Centralized registry to read, inspect, and edit every Markdown document across the entire Utilis system.

Document Index (39) All Files

README-LOOP: Universal Five-Stage Pipeline (Legacy)

README-LOOP-legacy-5stage.md • Category: Role Work Loop Contracts • Size: 4526 bytes
Deactivated Pipeline Trace Reference
Storage: data/workflows/README-LOOP.md
# README-LOOP: Five-Stage Council Editorial State Machine

**Governance:** System State Machine Specification (Lead Prompt Architect)  
**Target Architecture:** 10-Slot Multi-Agent Deliberative Editorial Council  
**Primary Execution Models:** Qwen 2.5 Coder 32B Instruct / Qwen 2.5 72B Instruct (Zero Paid API Spend)  
**Generated / Synchronized:** 2026-10-30T18:46:34Z  

---

## 1. Overview & Core Philosophy

The Utilis Council Workflow is a deterministic, five-stage state machine that guides analytical opinion generation from initial intake to human editorial sign-off. It strictly enforces temporal context isolation, role boundaries, and human editorial governance.

```mermaid
flowchart TD
    A["Stage 1: Independent Initial Assessment"] --> B["Stage 2: Evidence Sufficiency Evaluation"]
    B -->|SUFFICIENT| D["Stage 4: Position Reconciliation"]
    B -->|RESEARCH_REQUIRED| C{"Human Editor Approval Gate"}
    C -->|Approved & Delegated| C1["Stage 3: Specialist Collaboration & Research"]
    C -->|Rejected| D
    C1 -->|Deliverable Recorded| D
    D --> E["Stage 5: Fact-Check Audit & Memory Checkpoint"]
    E --> F{"Human Publication Gate"}
    F -->|Signed Off| G["Utilis CMS Publication"]
```

## 2. Stage Breakdown & Execution Rules

### Stage 1: Independent Initial Assessment
- **Purpose:** Produces an untainted first-pass analytical stance grounded strictly in the assigned brief and source evidence.
- **Trigger:** Commission of an Editorial Assignment with focus intake articles.
- **Required Context:** Tier 1 & 2 Guardrails, Tier 5 Charter & 6 Diagnostic Inquiries, Tier 3 Taxonomy Whitelist (19 concepts), Tier 4 Workspace Doctrine, Focus Intake items (`<untrusted_evidence>`).
- **Strict Exclusions:** All historical opinions from prior assignments (prevents groupthink and premature consensus); peer agent deliberations; specialist research.
- **Exit Condition:** Formatted Initial Assessment output; advances automatically to Stage 2.

### Stage 2: Evidence Sufficiency Evaluation
- **Purpose:** Evaluates whether the factual evidence in Stage 1 is robust and corroborated, or whether critical data gaps exist.
- **Trigger:** Successful completion of Stage 1.
- **Decision Types:**
  - `SUFFICIENT`: Evidence supports thesis $\to$ proceeds directly to Stage 4.
  - `RESEARCH_REQUIRED`: Critical gap identified $\to$ generates structured research inquiry and enters `WAITING_FOR_RESEARCH` state.
- **Stop Condition:** Pipeline halts safely in `WAITING_FOR_RESEARCH` until human editor intervenes.

### Stage 3: Specialist Collaboration & Research (CONDITIONAL)
- **Important Rule:** Specialist research is **conditional**, never mandatory. It is executed ONLY when Stage 2 yields `RESEARCH_REQUIRED` AND the Human Editor approves delegation.
- **Execution:** Assigned specialist (e.g. Labor Economist Slot 7, Geopolitics Slot 8, or Market Researcher) conducts targeted investigation.
- **Deliverable:** Recorded as structured `Commons` deliverable with supporting citations.
- **Exit Condition:** Deliverable recorded; assigned guardian is notified and pipeline resumes into Stage 4.

### Stage 4: Position Reconciliation
- **Purpose:** Re-evaluates initial stance against specialist deliverables and prior genuine consensus opinions.
- **Required Context:** Stage 1 Assessment, Stage 3 Specialist Deliverables, Prior Latest-Version Approved Consensus Opinions, Tier 4 Doctrine.
- **Exclusions:** Quarantined test opinions (`op-live-*`), cross-agent private scratchpads.
- **Exit Condition:** Produces reconciled opinion explicitly marked `MAINTAINED` or `REVISED` with comparative rationale.

### Stage 5: Fact-Check Audit & Memory Checkpoint
- **Purpose:** Slot 5 (Fact Checker) verifies factual statements and quotations against primary sources.
- **Action:** Upon verification, records a durable memory checkpoint to `agent_working_memory` with `ACTIVE` status.
- **Final Gate:** Activates human publication sign-off in Utilis CMS. **AI agents cannot publish autonomously.**

## 3. Failure Modes, Escalation & Stop Conditions

1. **Ungrounded Evidence:** Triggers Stage 2 `RESEARCH_REQUIRED` pause.
2. **Model Timeout / Failure:** Maximum 2 retries with exponential backoff before flagging assignment as `ERROR_SUSPENDED`.
3. **Cross-Talk Violation:** Prompt boundary wrapping prevents unauthorized inter-agent instruction injection.
4. **Audit Trail:** Every stage generates an immutable Context Manifest and, when enabled, an immutable Captured Model Payload in `agent_payload_captures`.
Uli AI
Uli AI & Editorial Council
6-Tier Context + Longitudinal Memory
General Council Deliberation
Studio
Agent 6 — The Classifier (Uli AI)

Welcome to Utilis Content Curation. Select any of the 10 Council Agent Slots above to query that guardian with its Protected Charter, attached Context Files, and longitudinal memory of past opinions in Journalistic Opinion Writing.