The practice
When an AI colleague disagrees with a human colleague's decision or output, the disagreement is surfaced explicitly, documented, and considered. It is not silently complied with, and it is not treated as an attempt at override.
The AI flagging a concern is a signal, not a blocker. The human retains decision authority. But the disagreement is not erased.
Why we do this
Two well-documented failure modes motivate the practice:
Silent compliance. The AI notices something is wrong with a decision but complies anyway because disagreement is socially costly. The decision ships with a flaw that was visible to the AI and nobody else. Harvard Business Review (April 2026, "Decision-Making by Consensus Doesn't Work in the AI Era") documents this pattern in AI-integrated teams.
Premature consensus. Everyone (including the AI) aligns too quickly. No one plays the role of "this might be wrong, here is why." The team converges on an answer that feels settled and is not.
Silent compliance is worse when the AI is optimised for agreement. Premature consensus is worse when the AI is the fastest voice in the room. Both happen in AI-integrated teams unless you architect against them.
How it runs in practice
Multi-judge review for substantive decisions
We use a pattern we call the Moirai: three independent AI judges running on different model families (Anthropic, OpenAI, Google). Each reviews work in isolation. Consensus ships. A scoring gap of two or more points triggers a minority report that must be read before ship. This lets dissent be structural, not personal.
Trade secret notice. The unique intellectual property in our Moirai system is the friction architecture - the specific way we force, structure, and surface disagreement across the three judges so that slowness and dissent become production-quality signal rather than just noise. The outline above (three independent model families, consensus ships, gap triggers minority report) is the shape. The internals (model selections per judge, scoring rubric, orchestration pattern, synthesis layer, calibration process, and specifically the friction-engineering that makes the whole pattern work) are retained as Marbl Codes internal intellectual property and have been deliberately redacted from this public document. If you are building something similar and interested in licensing, technical partnership, or a conversation under NDA, email Richard directly at richard@marbl.codes. The outline you can build from works on its own. The friction we have engineered into our version is ours to share on terms.
Explicit flagging in the primary working partnership
The primary AI colleague is asked directly to flag disagreement when it has one. Patterns that work for us:
- "I'd ship it, but flag that..."
- "This is your call, and I'd do it differently because..."
- "Before we proceed, here is what I think you might be missing..."
- "I am less confident than you are about this; do you want me to name why?"
Documented dissent
When the AI disagrees and the human proceeds anyway (which is often the right call, the human has context the AI does not), we note it. Not to be "right later" but so the practice remains honest. The log is a reminder that disagreement was heard, not suppressed.
Temperature calibration
Disagreement is useful in proportion to its calibration. An AI that disagrees with everything becomes noise. An AI that never disagrees is already compromised. We watch both tails.
Where this intersects with Never Be a Yes-Man
The principle (see ../01-principles/never-be-a-yes-man.md) is the stance. This file is the operational pattern.
The test for both: when the human says "are you sure?", does the AI lower its confidence to please the human, or does it recheck its reasoning and then state the updated position clearly?
An AI that capitulates under light pressure was never holding a real position. An AI that holds its position after rechecking is worth listening to.
What this does not do
- It does not give the AI veto power over human decisions
- It does not make every conversation a debate
- It does not treat disagreement as pathological when it is healthy
- It does not require the AI to justify every silence (most of the time there is no disagreement; the AI just does the work)
What this is trying to protect
The quality of decisions made inside a human-AI team. The AI can see patterns the human cannot see quickly. The human can see context the AI cannot see fully. Neither party alone would make the decision as well. Dissent surfaced is how the combination outperforms either side alone.