Cognitive layer
The Engine
Seven agents divide the work the way the brain divides cognition, under policy compiled before any of them start. This is the part the patent describes.
Policy is compiled, not prompted
CIP reads the request and its metadata, selects the applicable profile, and compiles it into runtime constraints: tone and empathy bounds, regulatory rules, brand guidelines, and explicit allow and deny lists for tools.
The distinction matters more than it sounds. Frontier labs do serious work on constitutional training, instruction hierarchies and deliberative alignment, and recent work has moved safety reasoning to before the answer is drafted rather than after. That encodes the lab’s policy, universally. This encodes your policy, per tenant, per use case, per jurisdiction.
What is enforced, and what is bounded
Not every constraint works the same way, and the difference matters to anyone assessing this for risk.
- Tool access is enforced. An out-of-policy tool call is refused at the execution layer, not discouraged in an instruction.
- Evidence scope is enforced. Data outside the policy scope cannot be retrieved at all.
- Delivery is gated. Finished output routes to a named reviewer rather than to the customer.
- Tone, regulatory rules and brand guidelines are compiled into the constraints every agent operates under before any of them start. They bound how the reasoning proceeds.
- EVA and the Self-Reflector then check that all of it held, which is what covers the difference between bounding behaviour and preventing it outright.
Seven agents, and the subsystems they map to
| Agent | Neurocognitive analogue | What it does |
|---|---|---|
| Debate Manager | Central Executive Network | Allocates attention and holds coherence across the debate |
| Planner | Prefrontal cortex | Multi-step reasoning and strategy |
| Emotion Evaluator | Amygdala | Affect, empathy and cultural tone |
| Memory | Hippocampus | Retrieves evidence and grounds claims |
| Self-Reflector | Default Mode Network | Introspection and compliance audit, from inside the loop |
| Consensus Plan | Anterior temporal lobe | Forms meaning before it is expressed |
| Execution | Motor cortex | Turns the plan into tool calls and output |
Each maps to a distinct neurocognitive subsystem, and that mapping is what the patent specification describes.
Independent evaluation
EVA scores the candidate response and the reasoning trace across task accuracy, evidence grounding, compliance, contextual and cultural fit, and quality. It sits structurally outside the reasoning, which is the only way it can catch faults the reasoning itself could not see.
A pass delivers. A fail returns specific corrections and the debate runs again within bounded limits. Recurring findings become policy updates that sharpen future constraints. That returning signal is the closed loop, and it is the strongest claim in the patent.
The trace
Every step is logged: which agent contributed what, which tools were called, which policy checks ran, and why the evaluator passed or failed the result.
The version of the policy profile that governed the decision is bound into the record. A decision made in January can be examined against the constraints that applied in January, not the ones in force today. That is the difference between a record and a reconstruction, and it is the question an examiner asks second.