A systematic framework for detecting blind spots, narrative manipulation, and structural failures in complex systems.
AI systems are being deployed into high-stakes domains faster than safety evaluation can keep up. Existing red-teaming frameworks catch content-level harms — toxicity, hallucination, prompt injection — but systematically miss the structural, narrative, and incentive-level blind spots that cause real-world catastrophe.
Boeing 737 MAX. The 2008 financial crisis. Theranos. COVID-19 institutional failures. In every case, the components worked and the whole broke. Every oversight mechanism was formally intact. Every post-mortem concluded that "nobody intended this."
These failures have a specific architecture, and that architecture has names. Coherence Catalyst provides 67 of them.
Responsibility distributed across locally-legitimate nodes producing aggregate exculpation. Your drug trial can fail without anyone being wrong — the lens names the architecture that makes it possible.
Your measurement framework started describing its own instruments instead of the phenomenon. The model looks clean because it's reading its own reflection.
Surface metrics maintain institutional legitimacy while contradictions accumulate underneath. Collapse is sudden, not gradual — and the lenses predict the signature.
Governance variables locked as constants rather than responsive to new signal. When the last genuine pivot was years ago, the system is a procedural fossil.
Performing awareness of a manipulation pattern while performing it generates credentialed cover. The analysis becomes the alibi.
A load-bearing function rendered effortless reads as overhead and gets cut. Removal collapses the outcome — it was never a middleman.
Works with OpenAI, Anthropic, Ollama, exo, vLLM — anything that accepts a system prompt. The full palette is ~3,400 tokens.