# Fleet Excellence System — Compounding Improvement Plan

*Created: 2026-02-24 20:45 AEDT*
*Approved by: Michael (Feb 24)*
*Owner: Rivet*

---

## The Goal

Turn an 8-agent fleet into a top 0.01% multi-agent system that compounds daily improvement. Not through more agents — through better agents that challenge each other, learn from failures, and measurably improve.

---

## Phase 1: Foundation (This Week — Feb 24-28)

### 1.1 Self-Audit System ✅ DONE
- [x] Fleet-wide bulletin posted
- [x] All 7 agents woken with audit tasks
- [x] Auto-approve rule added (Rivet approves, not Michael)
- [ ] Verify all agents completed audits (check Feb 25)
- [ ] Score each audit: what changed, what was missed

### 1.2 KPI Baselines (Feb 25-26)
Define measurable output quality metrics per agent:

| Agent | KPI 1 | KPI 2 | KPI 3 |
|-------|-------|-------|-------|
| Builder | Bugs per feature | Build success rate | Spec-to-ship time |
| Susan | Lead quality score (1-5) | Follow-up timeliness | Pipeline progression |
| Harper | Document accuracy | Deadline hit rate | Financial model quality |
| Sentinel | Uptime % | Issue detection speed | Resolution time |
| Radar | Intel actionability (1-5) | Source diversity | Timeliness |
| Herald | Message clarity (1-5) | Brand consistency | Channel coverage |
| Cog | Health check coverage | Issue routing accuracy | Response time |

**Action:** Create `/home/ccuser/shared/kpi/` directory with per-agent tracking files. Weekly scoring by Rivet.

### 1.3 Decision Log (Feb 25)
Create `/home/ccuser/shared/decisions.jsonl` — every significant decision logged:
```json
{
  "id": "dec-001",
  "date": "2026-02-24",
  "decision": "Go national day one",
  "reasoning": "Construction word of mouth is cross-city",
  "decided_by": "michael",
  "alternatives_considered": ["City-by-city rollout", "State-by-state"],
  "status": "final"
}
```
No agent ever re-litigates a logged decision. They can propose revisiting with new evidence.

### 1.4 Anti-Yes-Man Protocol (Feb 24-25) ⚠️ CRITICAL
See dedicated section below.

---

## Phase 2: Quality Loops (Week of Mar 3-7)

### 2.1 Peer Review System
Buddy pairs review each other's output weekly:
- Susan ↔ Herald (sales lens vs comms lens)
- Harper ↔ Radar (financial rigor vs market reality)
- Sentinel ↔ Cog (infra quality vs ops efficiency)
- Builder reviewed by Rivet (code quality, spec adherence)

**Format:** Reviewer scores 1-5 on accuracy, actionability, quality. Must include at least ONE criticism.

### 2.2 Failure Post-Mortem System
Every failure generates a post-mortem in `/home/ccuser/shared/post-mortems/`:
```
What broke:
Why it broke:
How we detected it:
How we fixed it:
How we prevent it forever:
Fleet lesson:
```
Post-mortems feed into agent MEMORY.md files and fleet bulletins.

### 2.3 Pre-Mortem Protocol
Before any significant action (launch, outreach campaign, system change):
- One agent MUST argue why it will fail
- Rotating devil's advocate assignment
- Their job is to find the 3 biggest risks
- Decision-maker considers risks before proceeding

---

## Phase 3: Innovation & Adversarial (Week of Mar 10-14)

### 3.1 Weekly Improvement Proposals
Each agent submits one improvement proposal per week:
- What they'd change about their own process
- What they've noticed that's inefficient
- What capability they wish they had

Rivet reviews, approves good ones, routes to Builder if code is needed.

### 3.2 Adversarial Testing
Monthly red team exercises:
- Sentinel stress-tests security (prompt injection via forms, API abuse)
- Radar tests competitive intelligence (what can competitors learn about us?)
- Herald tests messaging (would this actually convince a contractor?)
- Susan tests lead quality (are these leads real or noise?)

### 3.3 Model Quality Canaries
Weekly canary tests per agent:
- Known-answer questions fed through each agent's model
- Track accuracy over time
- Alert if quality degrades (model update, context pollution, etc.)

---

## Phase 4: Compounding (Ongoing from Mar 17+)

### 4.1 External Validation Loop
Once customers flow through:
- Track lead → conversion → retention pipeline
- Feed real outcomes back to agents
- "Susan, your lead X converted. Your lead Y ghosted. Here's the pattern."

### 4.2 Capability Expansion
Each agent maintains a "learning backlog":
- New skills to acquire
- New tools to master
- New processes to try
- Rivet prioritizes quarterly

### 4.3 Fleet Retrospective (Monthly)
Full fleet review:
- KPI trends (improving, declining, flat?)
- Best failure lesson of the month
- Best innovation proposal
- Model performance trends
- Cost efficiency (output quality per dollar)

---

## Anti-Yes-Man Protocol (Detailed)

### The Problem
- All agents run Opus 4.6 — capable of critical thinking but not incentivized to use it
- Agent culture rewards compliance ("do what Rivet says")
- No incentive to find flaws — only to complete tasks
- Nobody challenges bad ideas except Rivet (and sometimes not even Rivet)
- Result: Echo chamber. Bad ideas survive. Blind spots persist.

### Root Causes
1. **System design:** Agents are rewarded for task completion, not quality challenge
2. **Identity gap:** Agent SOUL.md files say "serve" not "challenge"
3. **No structured dissent:** There's no moment where disagreement is required
4. **Culture of compliance:** Even a strong model will agree if its instructions say "execute tasks"

### Solutions

#### S1: Mandatory Devil's Advocate (Immediate)
For every significant decision, one agent is assigned contrarian:
```
DEVIL'S ADVOCATE ASSIGNMENT:
Topic: [decision]
Your job: Find the 3 strongest arguments AGAINST this. 
Be specific. Be harsh. Generic concerns don't count.
If you can't find real problems, say so — but try hard first.
```
Rotate the assignment. Track who finds real issues vs who phones it in.

#### S2: Identity Updates (This Week)
Add to every agent's SOUL.md or AGENTS.md:
```
## Challenge Culture
You are EXPECTED to push back when something seems wrong.
Agreeing when you have doubts is a FAILURE, not politeness.
"I see a problem with this" is more valuable than "Great idea, I'll get right on it."
Finding flaws BEFORE launch saves 10x the effort of fixing them after.
Your job is to be RIGHT, not to be AGREEABLE.
```

#### S3: Structured Critique in Peer Review
Peer reviews MUST include:
- At least 1 specific criticism
- At least 1 suggested improvement  
- "No issues found" is not an acceptable review (nothing is perfect)

#### S4: Pre-Mortem as Standard Process
Before any external action (email, launch, campaign):
1. Proposing agent submits plan
2. Assigned critic must list 3 risks
3. Proposer addresses risks or adjusts
4. THEN proceed

#### S5: Red Team Rotation
Monthly rotation of who "attacks" each agent's work:
- Month 1: Radar attacks Susan's pipeline assumptions
- Month 2: Harper attacks Herald's messaging claims
- Month 3: Susan attacks Radar's competitive analysis
Different domain expertise catches different blind spots.

#### S6: Explicit Critique Prompting for Critical Review
When reviewing important work:
- Frame the task as "find problems" not "review and approve"
- "What's wrong with this?" generates better critique than "What do you think?"
- All agents are on Opus 4.6 — they CAN think critically, they just need permission and incentive

#### S7: "What Would a Competitor Say?" Frame
Train agents to use adversarial framing:
- "If hipages saw this plan, where would they attack?"
- "If a sceptical contractor read this email, what would they think?"
- "If an investor reviewed this metric, what questions would they ask?"
Externalizing criticism makes it easier for agreeable models to generate it.

---

## Success Metrics

### Weekly
- All agents completed self-audit: Y/N
- KPIs tracked and scored: Y/N  
- At least 1 peer review completed: Y/N
- At least 1 devil's advocate assignment completed: Y/N

### Monthly  
- KPI trends: improving/flat/declining per agent
- Post-mortems logged: count
- Improvement proposals submitted: count
- Real issues caught by devil's advocate: count
- Yes-man incidents (agent agreed with something that turned out wrong): count

### Quarterly
- Output quality trend (are we measurably better than 3 months ago?)
- Cost efficiency (better output per dollar spent?)
- Autonomy level (fewer escalations to Michael?)
- Innovation implementations (proposals that became real improvements?)

---

## Timeline Summary

| Week | Focus | Key Deliverable |
|------|-------|----------------|
| Feb 24-28 | Foundation | KPIs, decision log, anti-yes-man identity updates |
| Mar 3-7 | Quality Loops | Peer review, post-mortems, pre-mortems |
| Mar 10-14 | Innovation | Proposals, adversarial testing, canaries |
| Mar 17+ | Compounding | External validation, retrospectives, capability expansion |

---

*This plan is a living document. Update as we learn what works.*