stress-test-methodology

Adversarial Plan Review

Adversarial plan review uses two independent agents with different roles to stress-test a planning document:

1. Red team (adversarial) -- reads the plan and surrounding artifacts, generates what-if questions targeting gaps, unverified assumptions, edge cases, and failure modes. Operates on local artifacts only to keep questions grounded.

2. Blue team (neutral analyst) -- receives the what-if questions and attempts to answer each using the plan, artifacts, and a configurable set of tools. Classifies each answer with a verdict.

The technique works because the agents have separate contexts: the red team generates challenges without knowing the answers, and the blue team answers without knowing which questions are easy or hard. Neither agent is biased toward defending or attacking the plan.

When to Use

• Before committing to an implementation plan
• Before presenting a business plan or proposal to stakeholders
• After drafting a design document and before sharing it for review
• When you feel a plan is "done" but want to pressure-test it
• When surrounding context is rich enough to ground the analysis (codebase, supporting docs, config files)

Stressor Modes

Grounded Mode

The red team generates what-if questions strictly tied to plan artifacts. Every question references a specific file, config value, dependency, or code path found in the surrounding context. This mode finds gaps in what the plan explicitly covers.

Best for: implementation plans with a full codebase, detailed technical designs with supporting configs.

Creative Mode (Residuality Theory)

Inspired by Barry O'Reilly's Residuality Theory, creative mode adds extreme, cross-domain stressors that go beyond artifact grounding. Instead of only asking "what does the plan miss?", it asks "what survives when extreme stress hits?"

Additional stressor categories:

• Extreme scale (10x/100x assumptions)
• Dependency disappearance (vendor, library, team member gone)
• Regulatory/compliance shifts
• Team/organizational disruption
• Market/competitive shocks
• Graceful degradation analysis ("what survives?" not just "what breaks?")

Best for: strategic plans, plans with external dependencies, plans where hidden coupling matters more than line-level coverage.

Both

Runs grounded questions first (artifact-tied), then creative questions (extreme/cross-domain). Most thorough. The QA report separates grounded and creative sections so you can triage accordingly.

The Verdict System

The blue team classifies each answer:

Verdict	Meaning	Action
ANSWERED	Plan or artifacts explicitly address the concern, with a quotable reference	No action needed
PARTIALLY ADDRESSED	Some coverage exists but gaps remain	Strengthen the relevant plan section
NOT COVERED	The plan has no answer -- genuine gap	Add coverage to the plan
UNCERTAIN	Cannot determine with available tools -- gap might or might not exist	Expand tool scope or investigate manually

Focus your iteration on NOT COVERED and UNCERTAIN items. These are the plan's blind spots.

Tool Scope

The blue team's tool scope determines how thoroughly it can verify claims:

Local artifacts only (Read, Grep, Glob)

• Best when: rich surrounding context (full codebase, detailed docs)
• Limitation: cannot verify claims about external systems, standards, or APIs
• UNCERTAIN verdicts may indicate the plan references things outside the artifacts

+ Web research (adds WebSearch, WebFetch)

• Best when: plan references external APIs, standards, benchmarks, or third-party services
• Allows the blue team to check documentation, specs, and industry practices
• Reduces UNCERTAIN verdicts for external dependency questions

+ System verification (adds Bash, MCP tools)

• Best when: plan makes assumptions about live systems (API responses, config values, resource limits)
• The blue team can query actual APIs, check live configurations, run diagnostics
• Most powerful -- turns assumptions into verified facts or confirmed gaps
• Requires the referenced systems to be accessible from the current environment

The red team always uses local artifacts only, regardless of scope selection.

Interpreting Results

A healthy stress test typically shows:

• 40-60% ANSWERED -- the plan covers many concerns
• 10-20% PARTIALLY ADDRESSED -- some areas need strengthening
• 10-20% NOT COVERED -- genuine gaps to fill
• 5-15% UNCERTAIN -- areas needing more investigation

If most questions are ANSWERED, the plan is solid. If most are NOT COVERED, the plan needs significant revision before proceeding.

Watch for false confidence: an ANSWERED verdict is only as good as the evidence behind it. Check that ANSWERED items include specific references (plan sections, code paths, config values), not vague reassurances.

Limitations

• The red team can only challenge what it can see. If critical context is in a separate repo, a Confluence page, or someone's head, the red team cannot generate questions about it.
• The blue team's ANSWERED verdicts depend on its tool scope. A local-only blue team marking something ANSWERED based on a code comment is weaker than a system-verification blue team confirming it against a live API.
• Neither agent understands organizational context (team capacity, political constraints, budget). These factors affect plan feasibility but are invisible to the agents.
• Creative mode questions are not grounded in artifacts. The blue team may mark many as UNCERTAIN or NOT COVERED simply because the plan was never designed to address extreme scenarios. This is expected -- the value is in identifying which extreme scenarios the plan should address.

Running a Stress Test

Use the /stress-test command:

/stress-test path/to/plan.md

The command orchestrates the full flow: reads the plan, asks about tool scope, dispatches the red team, then the blue team, and presents a summary with action items.