Red Team Kit

Adversarial hypothesis testing framework. A council of red-team agents challenges your strategic, financial, or operational hypotheses and scores them for robustness.

Features

• Six adversarial agent roles — Devil's Advocate, Contrarian, Stress Tester, Assumption Hunter, Blind Spot Finder, Scenario Explorer — each attacks hypotheses from a distinct cognitive angle.
• Quantified scoring — four weighted dimensions (evidence strength, assumption quality, counter-resilience, internal consistency) produce a composite 0–1 score with a ROBUST / CONDITIONAL / WEAK / REFUTED verdict.
• 52 production hypotheses — curated YAML corpus spanning financial, market, strategic, operational, temporal, and cross-domain categories for benchmarking and regression testing.
• Pluggable LLM backend — defaults to OpenAI-compatible APIs; swap any endpoint via env vars (REDTEAM_LLM_BASE_URL, REDTEAM_LLM_MODEL, REDTEAM_LLM_API_KEY).
• Structured reports — export session results as JSON or dict with per-hypothesis breakdowns and verdict distribution summaries.
• Composable council — add or remove agents at runtime; run single-hypothesis sessions or batch the full hypothesis corpus.

Quick Start

pip install -e ".[dev]"
export REDTEAM_LLM_API_KEY="sk-..."   # OpenAI or compatible

from redteamkit.council import AdversarialCouncil
from redteamkit.report import ReportGenerator

# Define a hypothesis
hypothesis = {
    "id": "h001",
    "title": "Enterprise pricing at $500/seat is defensible",
    "statement": "Customers will accept $500/seat pricing given the ROI demonstrated in pilots.",
    "assumptions": ["Pilots are representative of production load", "Competitors won't reprice"],
}

# Run the adversarial council
council = AdversarialCouncil()
session = council.run_session(hypothesis)

# Score and report
report = ReportGenerator().generate(session)
print(report.summary)         # "Verdict: CONDITIONAL (score: 0.61)"
print(report.to_json())       # Full JSON report

Run the test suite:

pytest

Architecture

red-team-kit/
├── redteamkit/
│   ├── agent.py      # AgentRole enum, Critique model, RedTeamAgent
│   ├── council.py    # AdversarialCouncil — orchestrates agents, holds SessionRecord
│   ├── scoring.py    # HypothesisScorer — 4-dimension weighted scoring, Verdict enum
│   ├── llm.py        # LLMConfig, LLMClient (sync/async httpx), prompt builders
│   └── report.py     # ReportGenerator — JSON/dict export, verdict distribution
├── hypotheses/       # 52 YAML hypothesis files (financial, market, strategic, …)
└── tests/            # pytest suite — 6 modules, full unit coverage

Data flow:

Hypothesis YAML
      │
      ▼
AdversarialCouncil.run_session()
      │  spawns one RedTeamAgent per role
      │  each agent calls LLMClient → returns Critique
      ▼
HypothesisScorer.score(critiques)
      │  evidence_strength  × 0.30
      │  assumption_quality × 0.25
      │  counter_resilience × 0.25
      │  internal_consistency × 0.20
      ▼
SessionRecord  →  ReportGenerator  →  JSON / dict

Configuration

Env var	Default	Description
REDTEAM_LLM_API_KEY	—	API key (required for LLM calls)
REDTEAM_LLM_BASE_URL	https://api.openai.com/v1	OpenAI-compatible base URL
REDTEAM_LLM_MODEL	gpt-4o	Model name

Contributing

See CONTRIBUTING.md for branch conventions, test requirements, and the PR checklist.

---

Built by TechKnowMad Labs