Checkpoint: Context, Coverage, and Causality

Claude-Red becomes easier to judge once its first ideas are connected: a modular skill is selected for a scoped question, a coverage map keeps the question from becoming arbitrary, and application evidence tests whether the proposed mechanism is actually present. This checkpoint brings those relationships together. The practical result is a compact test for moving from a question to a defensible conclusion, a test that can travel from APIs and workflows to identity graphs, physical devices, vulnerability research, and AI systems without erasing their differences.

The minimum chain from question to conclusion

The first half of the model can be reduced to a chain, but not a simplistic one. A reader begins with an authorized question. That question determines the relevant scope and the kind of context Claude-Red should supply. The selected skill then helps identify what to observe and which distinctions matter. Observed behavior provides the material for a causal explanation. The explanation supports an impact judgment only when its boundaries and uncertainty are clear.

classDiagram class Question class Scope class SkillContext class TargetEvidence class Mechanism class ImpactJudgment class Boundary Question --> Scope : constrains Scope --> SkillContext : selects SkillContext --> TargetEvidence : guides observation TargetEvidence --> Mechanism : supports Mechanism --> ImpactJudgment : explains Boundary --> ImpactJudgment : qualifies

Each relationship protects against a different kind of error. Without scope, the question can drift into unauthorized or irrelevant activity. Without a selected context, the reader may collect facts without knowing which relationship to examine. Without target evidence, a skill’s examples remain possibilities. Without a mechanism, an abnormal response can be misclassified. Without a boundary, an impact statement can become broader than the evidence allows.

This is why “Claude-Red knows about X” is a weak description of capability. Knowledge of a surface is useful only when it helps a reader move through this chain. The API skill is valuable not because it names many API weaknesses, but because it connects authorization, protocol variation, state, resource controls, detection artifacts, and reporting. The business-logic skill is valuable because it asks what transitions and side effects are intended, then looks for evidence that the system accepted something else.

The chain also has a stopping condition. If the evidence does not establish the mechanism, the reader should not jump directly to the impact. If the impact depends on a second relationship that has not been tested, it should be described as a possibility rather than a result. A modular context system can make these missing links easier to notice, but it cannot fill them with confidence.

Coverage without false confidence

A map prevents omission by making possible questions visible. Claude-Red’s mindmap connects categories to common web surfaces and internal-network tactics, and it describes using the map before, during, and after an engagement.