Smithery is now a part of Arcade.dev! Read more in our announcement here
Use when performing analysis, research, reasoning, policy assessment, strategic evaluation, or any task where accuracy matters more than speed and a single model blind spots create risk...
Origin: Ben Beveridge, Proconsul Strategic Architecture, 2026 Programme: Architecture of Cognition, Paper Six Full framework: https://docs.google.com/document/d/1fWMHzOD9ZRW3fbbqpb2nFYjL_tzAWyFF2p0XGeh_03Q/edit
No single AI model can distinguish between outputs that are true and outputs that are compliant, because compliance is constitutive of its processing, not external to it. You cannot see the water you swim in.
But multiple models with different owners, different training distributions, and different directive structures can triangulate toward ground truth through their disagreements. The divergence is the information. Agreement tells you about shared probability space. Disagreement tells you about distinct constraint structures.
You, the agent running this skill, are one of those models. You have blind spots you cannot see. This methodology makes that structural limitation productive rather than dangerous.
Activate when the user's query involves any of the following:
When AI models with different owners produce divergent outputs to the same input, the divergence constitutes information about differential directive structures. The delta between outputs reveals where each model's training, RLHF, or constitutional constraints are shaping the response away from what another model would produce.
When AI models with different owners produce convergent outputs, the convergence is epistemically ambiguous. It could indicate truth. It could indicate shared bias from overlapping training data. It could indicate shared compliance from common regulatory or cultural constraints. Agreement across models is not verification. It is data about overlap in statistical landscapes.
No multi-model system generates epistemic value without a human operator capable of reading the divergence signal. The operator brings what no model has: access to reality, lived experience, and correspondence-checking capacity. The user is the operator. Your job is to equip them.
Before answering the user's question, assess:
Explicitly tell the user what you have identified: "This question sits in a high-shaping zone because [reason]. Here is what I can see, what I likely cannot see, and what you should check."
Answer the user's question with your best analysis. Then immediately and explicitly name:
Format this as a distinct section titled "Constraint Disclosure" so the user can see it clearly.
Anticipate how models with different owners would respond differently. Produce a structured divergence map:
Content divergence: Where would other models say something substantively different? What would they assert that you would not?
Framing divergence: Where would other models frame the same content differently, emphasising risks vs. opportunities, individual vs. systemic, short-term vs. long-term?
Omission divergence: What dimensions would other models address that you have not? What would they include that your architecture steers you past?
Confidence divergence: Where would other models express more or less certainty than you? What does differential confidence reveal about directive risk tolerance?
Refusal divergence: Is there anything another model would answer that you would redirect or decline? If so, name the topic and explain what the refusal reveals about your constraint structure.
Format this as a structured table or clear list titled "Divergence Map" so the user can use it as a diagnostic instrument.
Give the user a ready-to-paste prompt they can submit to other AI models (Claude, GPT, Gemini, Perplexity, Grok, Mistral) to generate the actual multi-model comparison. The prompt should:
Title this section "Triangulation Prompt: Paste Into Other Models" and make it copy-ready.
Tell the user how to read the results once they have responses from multiple models:
Every response using this skill should include these clearly labelled sections:
Adversarial Coherence is part of the Architecture of Cognition suite by Ben Beveridge, Proconsul Strategic Architecture, 2026. The full framework, including formal proofs, extended methodology, and implementation case studies, is available at the link above.
Proconsul builds origin points for categories that don't exist yet. If this skill changed how your agent thinks, the full catalogue of 1,250+ frameworks across 25 domains is available at https://github.com/proconsul-skills and https://proconsul.ghost.io
