01
What an LLM does well
A large language model can interpret instructions, analyse supplied information, draft plans, and generate useful recommendations from the context it can currently see.
The Framework
ToM
Intelligence as structure in motion.
ToM (Tree of Mind) is a governed AI architecture around the language model. It brings approved context, evidence, and configured rules into decisions while keeping human authority outside the model.
Why structure matters
We believe artificial intelligence is an engineering problem.
The question is not only whether an AI system can produce an answer. It is whether that answer remains stable, coherent, and traceable under load.
Consider the tree. A tree that never feels the wind grows hollow. Strength comes from measured resistance, from developing more resilient structure in response to real challenge.
This is the inspiration behind ToM's approach to bounded adaptation. Operational evidence can inform reviewable calibration over time; exposure alone does not guarantee improvement.
The aim is not autonomous maturation or perfect correctness. ToM is designed to preserve governed state, detect drift, and support bounded adjustment when evidence and authority permit.
Conceptual architecture
A governed system around the model.
Connected context enters an Observe, Interpret, Decide, Act, and Retain loop before a recommendation or permitted action reaches an inspectable outcome. The language model contributes reasoning inside that system; it is not the system's source of authority.
Connected context
- Product workflows
- Enterprise systems
- Documents & evidence
- Operational tools
Governed loop
Observe · Interpret · Decide · Act · Retain, with policy, memory, evidence, and human boundaries around the language model.
Inspectable outcomes
- Recommendations
- Evidence-linked decisions
- Bounded actions when permitted
- Auditable records
Conceptual architecture: this simplified view is not a runtime trace or performance claim. Implementation details intentionally omitted.
Stage explorer
Follow what happens around the model.
Use the stage explorer to read each part of the loop while the complete system map remains visible above.
Interpret
The model contributes reasoning
A configured language model can help interpret the available context, but it remains one component inside the governed system.
Model output is treated as a proposal, not as permission to act.
Language model + ToM
The model is a reasoning component, not the whole decision system.
Modern language models are trained on vast bodies of text and can produce capable analysis. The remaining problem is how context, evidence, permission, and consequence are managed when that analysis enters a real workflow.
02
What ToM adds around it
ToM builds the governed path: approved context, available evidence, configured rules, memory boundaries, action limits, and explicit authority.
03
The practical difference
The recommendation can be released, constrained, held, or escalated through a separate configured decision path, with a record available for authorised review where supported.
Key principles
Four responsibilities around model reasoning.
The five-stage orbit shows how a decision moves. These principles describe what governs the boundaries around that movement.
01
Gated Information Boundaries
Context, evidence, memory, and product integrations remain subject to the permissions and handling rules of the active system boundary.
02
Reviewable Evidence
Configured decision paths are designed to retain the evidence, authority, and rationale needed for authorised review.
03
Reflection-Gated Memory
Permitted evidence may be retained after provenance, policy, and reflection checks rather than treating every interaction as memory.
04
Bounded Agency
Initiative remains controlled, never unlimited. ToM is designed to support adaptation without silently exceeding the configured mandate.
Public architecture view: the orbit explains how connected context moves through a governed decision cycle. It is conceptual and intentionally omits proprietary runtime topology.
Understand the framework
Govern the next action. Keep the context useful.
Runtime governance and persistent memory answer different questions within ToM. These guides connect each subject to a practical workflow, the published research and the evidence needed for an integration.
AI agent runtime governance
Assess a proposed action against current evidence, operating limits and the authority to proceed. Work through a changed document release and the questions to test at the execution boundary.
Read the runtime governance guidePersistent memory for AI agents
Preserve objectives, decisions and source references across extended work. Explore context drift, stale evidence and a session handoff that can be checked against the current task.
Read the persistent memory guideWhy it matters now
The harder problem is governing what happens after the model answers.
Models are becoming more capable and more interchangeable. The enterprise bottleneck is no longer just generating an answer. It is preserving context across extended work, governing access to core systems, and reviewing recommendations before they create financial, operational, or safety consequences.
ToM is not a prompt wrapper or another agent toolkit. It is the cognitive-control architecture around models, agents, and workflows, designed to make context, authority, evidence, and outcomes more inspectable. As models commoditise, this surrounding architecture is where durable differentiation and accountable action move.
Proposal is not permission
AI can recommend. Authority stays outside the model.
A persuasive model output must not be able to override a safety, policy, or authority boundary through language alone. ToM keeps selected release decisions in a separate configured path, where evidence, limits, and human authority, not model confidence, determine what may proceed.
Language model
What does the available context suggest doing next?
- Interprets instructions and supplied information
- Composes a candidate plan or recommendation
- Contributes reasoning without owning operational release
ToM
Is that recommendation supported and permitted now?
- Checks which approved context and objectives apply
- Applies available evidence, configured rules, and limits
- Returns the governed outcome available to the workflow
Conceptual outcome model: product integrations may support different configured outcomes and authority paths. These four responses describe the governance proposition, not a guarantee that every intervention is operationally optimal.
Inspect the evidence
See what Newport Resonance's current studies do, and do not, demonstrate.
Separate research examines structured reasoning, long-context continuity, and governed release in a bounded embodied-action simulation.
See industry applications
Explore governed decisions across consequential operating systems.
Spinal robotic navigation, autonomous mine haulage, integrated air defence, and bushfire grid restoration make proposal, permission, and consequence visible.
Put the Framework to work
Where does a model recommendation need a governed path into action?
Talk with us about ToM, an enterprise product, a governed evaluation, research, or strategic partnership.