Queue v3.1

Executive Summary

Queue v3.1 is not an "agent system" — it's a distributed AI operating system. The key difference: agents don't self-certify their work. An independent verifier checks every output before it counts as complete.

What Changed from v2

v2 (Old Approach)

Agent receives task → Agent does work → Agent says "done" → System trusts it

Problems:

• No verification — agent could fake completion
• No hash chain — couldn't prove data integrity
• No quality enforcement — low-quality outputs accepted
• Self-verification — agent checks own work (conflict of interest)
• No audit trail — couldn't trace what happened

v3.1 (New Approach)

Agent receives task → Agent does work → Receipt with hash → 
INDEPENDENT VERIFIER checks → Only then: verified

Improvements:

• 3-layer verification (structural, integrity, quality)
• Hash-chained receipts prove data integrity
• Quality score thresholds enforced
• No self-verification allowed
• Full audit trail for every job

7-Stage Workflow

The pipeline is STRICT — no skipping allowed.

Rule: If ANY stage is skipped → job = INVALID

Stage Assignment

3-Layer Verification System

Verification Decisions

Hash-Chained Receipts

Each receipt contains:

• input_hash — SHA-256 hash of previous stage's output
• output_hash — SHA-256 hash of this stage's output

This creates an immutable chain proving data integrity through the pipeline.

Example Chain

Stage 0 (research_synthesis):
  input_hash:  null (first stage)
  output_hash: 978d2db1bda39957...

Stage 1 (structural_review):
  input_hash:  978d2db1bda39957...  ← Must match Stage 0 output
  output_hash: e47faacc674adc23...

Stage 2 (brand_polish):
  input_hash:  e47faacc674adc23...  ← Must match Stage 1 output
  output_hash: d9af6735f71a90c2...

...and so on through all 7 stages

Why It Matters

1. Tamper evidence — If any stage's data changes, hash breaks
2. Provenance — Can trace every output back to source
3. Audit trail — Mathematical proof of what happened
4. Trust — No agent can fake their output

Weighted Routing

When routing a stage to an agent, the system calculates a score:

Score = (Capability Match × 0.40) +
        (Success Rate × 0.30) +
        (Quality Score × 0.20) +
        (Load Balance × 0.10)

Factors

Concurrency Control

The Solution: Claim Locks

UPDATE jobs 
SET status = 'claimed', claimed_by = ?, claimed_at = ?
WHERE id = ? AND status = 'dispatched'

Only ONE agent wins this race. The others get 0 rows affected.

Completion Definition

A job is COMPLETE only when:

• All 7 stages verified
• All artifacts verified
• Delivery confirmed (message_id + timestamp)
• Audit log recorded
• Job state = closed

Running Services

Auth Token: q3-v3-axon-fleet-2026

API Reference

Key Endpoints

Example: Create Job

curl -X POST http://localhost:3334/jobs \
  -H "Authorization: Bearer q3-v3-axon-fleet-2026" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "report",
    "priority": 7,
    "from": "dispatcher",
    "subject": "Q1 Marketing Report",
    "body": "Generate Q1 marketing performance report",
    "acceptance_criteria": {
      "output_format": "pdf",
      "required_sections": ["executive_summary", "recommendations"],
      "min_confidence": 0.85,
      "min_quality_score": 75
    }
  }'

Summary

Queue v3.1 transforms the FTWS agent fleet from a loose collection of workers into a coordinated, verified, audited operating system.

• No trust without verification — Verifier is source of truth
• No shortcuts — 7 stages required, no skipping
• No hiding — Full audit trail
• No corruption — Hash chains prove integrity
• No double-work — Concurrency locks prevent races

This is the foundation for scaling to government contracts and enterprise clients.