/RedTeam

name: RedTeam
evals:
  manual:
    reason: "attack quality requires judgment that findings target the supplied claims and include evidence-backed remediation; world-model scenario analysis must be judged for plausible evidence-based futures and whether it tests the supplied strategy across horizons"
    falsifier: "the reviewer finds a material claim untested, a finding unsupported, or remediation incompatible with the stated constraints, or a horizon ignores a material driver, misstates evidence, or fails to evaluate the supplied strategy"
metadata:
  short-description: "Stress-test ideas and plans with adversarial agents, and test ideas against 11 world-model horizons (6 months to 50 years). Use Council for collaborative debate."
version: 1.2.1
description: "Adversarial analysis deploying parallel expert agents to stress-test ideas, strategies, and plans — attacks atomic claims, then steelmans and counter-argues, with severity-ranked findings and remediation; also tests an idea or investment against 11 persistent world models from 6 months to 50 years (Fast/Standard/Deep), and refreshes and reports on those models. USE WHEN red team, attack idea, counterarguments, critique, stress test, devil's advocate, find weaknesses, break this, poke holes, strongest objection, world threat model, world model, test idea against the future, how will this hold up in 10 or 20 years, time horizon, long-term risk, update world models, show world models. NOT FOR collaborative debate to find best path (use Council), or threat modeling your own estate (ThreatModel)."

RedTeam

**Contents:** Ideal State · Intent Routing · Data & Config · Gotchas · Examples

Attacks ideas, strategies, plans, designs and judgment-layer question sets to find their weak points before reality does. It breaks an argument into atomic claims, attacks each from several adversarial angles, and ranks what survives by severity. A full run splits the attackers across model vendors and ends in a steelman plus the strongest counter-argument; a quick run returns the few strongest objections, often inline inside a subagent that may not spawn agents. Targets arguments, not network vulnerabilities. The two intents compose: a synthesis from AdversarialValidation can be stress-tested with ParallelAnalysis.

It also attacks an idea through time. A persistent world-model harness (ported from the retired WorldThreatModel skill, 2026-10-10) stress-tests an idea, strategy, or investment against 11 time horizons from 6 months to 50 years. Each horizon is a deep (~10 page) world model covering geopolitics, technology, economics, society, environment, security, and wildcards. Three speed tiers: Fast (~2 min, single synthesizer), Standard (~10 min, 11 parallel agents plus RedTeam and FirstPrinciples), Deep (~1 hr, adds Research and Council). TestIdea returns a probability-weighted scenario matrix across all 11 horizons; UpdateModels refreshes the models and ViewModels reports their status and freshness.

Ideal State

• Every attack lands on the argument's load-bearing claims, and the headline finding is the flaw that would collapse the structure, never a pile of nitpicks.
• Findings are ranked by severity with noise discarded, each names the claim it breaks, and each weakness carries a remediation that fits the stated constraints.
• Length follows the requested depth: a quick or inline pass returns the 4-6 strongest objections; a full stress test returns a steelman and a counter-argument (8 points each) from attackers split across model vendors; designing something new returns one synthesized solution from competing proposals.
• When the caller says inline or no agents, nothing is dispatched: no `Inference.ts` call and no `Agent` spawn.
• An idea tested through time gets a verdict at every one of the 11 horizons, stated as scenarios with probability ranges, never as predictions, at the tier asked for (Standard when none is named); any horizon or capability that failed is named in the result instead of vanishing.
• Every world model it reads or writes lives under `<configRoot>/LIFEOS/MEMORY/RESEARCH/WorldModels/` with a `last_updated` date, so how fresh a model is shows before an idea is tested against it.

Intent Routing

| Intent | Trigger | File | |----------|---------|------| | **ParallelAnalysis** | "red team this", "poke holes", "strongest objection", "quick adversarial pass", "inline, no agents": stress-test existing content | `Intents/ParallelAnalysis.md` | | **AdversarialValidation** | "battle of bots", "which approach is better", "help me design/decide X": produce new content via competition | `Intents/AdversarialValidation.md` | | **TestIdea** | "test idea", "test strategy", "test investment", "how will this hold up", "stress test against the future", "test against future", "threat model my bet", "world model": test any input against all 11 world models (6 months to 50 years), Fast/Standard/Deep | `Intents/TestIdea.md` | | **UpdateModels** | "update world model", "update world models", "refresh models", "new analysis": refresh world model content with new research/analysis | `Intents/UpdateModels.md` | | **ViewModels** | "view world model", "show models", "current models", "model status": read and summarize current world model state and freshness | `Intents/ViewModels.md` |

| File | Read when | |------|-----------| | `Philosophy.md` | framing what a fundamental flaw looks like | | `Integration.md` | pairing RedTeam with Research, FirstPrinciples or ExtractWisdom | | `References/WorldModelTemplate.md` | creating or updating a world model document (UpdateModels) | | `References/WorldModelOutputFormat.md` | presenting TestIdea results, at any tier | | `Setup.md` | first-time setup: building the 11 world models the time-horizon lane needs, and what the cross-vendor runs use |

World Model Storage

Models are stored at `<configRoot>/LIFEOS/MEMORY/RESEARCH/WorldModels/`: `INDEX.md` (summary of all models with last-updated dates)