AI DECISION ENGINE
Principles
Principles
Principles
Otonomii’s AI/ML capabilities are not built by applying the latest techniques from the research literature. They are built by decomposing problems to their fundamental truths and discovering solutions from first principles.
This is not anti-academic, it is anti-herd. The difference between applying known solutions and discovering new ones is the difference between following the market and understanding it.
Is this what everyone does? If so, why would it give us an edge? The default assumption is that popular approaches produce average results. Before adopting any technique, algorithm or architecture, the system asks who else is doing this and what are their results. If the answer is everyone and the results are mediocre, the approach is rejected regardless of theoretical elegance.
Start from what we actually know to be true, not what the literature says. Most AI/ML knowledge is herd-derived, biased toward prediction and leads to the same fragile systems everyone else builds. Otonomii decomposes every problem to its fundamental truths and builds upward from there.
Relentless questioning exposes hidden assumptions. Every design decision, parameter choice and architectural pattern is subjected to structured questioning: why this approach, what are the alternatives, what assumption does this depend on and what happens if that assumption is wrong.
The goal is to discover new ways, not apply existing solutions. Most platforms apply known techniques to new data. Otonomii discovers new techniques from first principles, validated against data. The system is a discovery engine, not an application engine.
Before investing resources in solving a problem, the system challenges the problem statement itself. Many AI/ML failures trace back not to poor solutions but to poor problem formulation. A perfectly solved wrong problem is worse than an approximately solved right problem.
The Semantic Data Layer is the Brain. The domain agnostic memory substrate that stores, retrieves, associates and compresses all information the system encounters.
It handles multi-modal data: structured data, unstructured data, streaming data and geospatial data. The data layer is cross cutting as it imports nothing from any cognitive layer. This ensures that memory is not contaminated by interpretation. The Brain stores what happened; the Mind decides what it means.
Commits new observations to memory with source, timestamp, confidence and regime context.
Retrieves relevant patterns using similarity matching that accounts for regime context.
Creates associative connections between patterns that co-occur, enabling contextual recall.
Reorganizes memory, merges redundancy, strengthens validated links and compresses rarely accessed information.
The Dynamic Logic Layer is the Mind, where raw memory becomes understanding. This layer transforms stored patterns into predictions, measures the gap between predictions and reality and uses that gap as the primary learning input.
It is inspired by predictive processing theory from neuroscience: the brain is fundamentally a prediction machine and experience is shaped by prediction error. The difference between what was expected and what actually happened.
Converts raw observations into internal representations optimized for prediction.
Generates expectations about the next state based on current context and historical patterns.
Computes prediction error across magnitude, direction and attribution.
Updates internal models continuously and incrementally based on prediction error.
Allows multiple internal models to express independent opinions.
Determines how much attention each input receives based on regime and reliability.
The Kinetic Action Layer is the Machine, where understanding becomes behavior. This layer translates the Mind’s assessments into concrete actions in the real world.
It answers four questions: when to act, when to stop, what to do and what resources are available. Self healing operations emerge from the interaction between Machine and Mind: identify the discrepancy, trace it to its source, generate a corrective action, execute the correction and verify the result.
Determines when confidence is sufficient to act. It rises in uncertainty and drops in well understood reversible regimes.
Determines when to stop searching, close a position or accept the current answer as good enough.
Translates assessments into concrete operations in the real world.
Manages computational budgets, operational limits, API quotas and time constraints.
Every agent, model and pipeline operates under a defined role with explicit permissions. An agent can read market data without being able to place orders or generate recommendations without executing them.
Data and decisions are classified by sensitivity and reversibility. One Way Door decisions require higher confidence thresholds and additional review gates.
Every data access must declare its purpose and access is limited to what is necessary for that declared purpose. Purpose declarations are logged and auditable.
Before any output is finalized, the system asks whether this is the obvious conclusion and what may be missing. This prevents groupthink among model ensembles.
Unexpectedly perfect results trigger automatic investigation for data leakage, circular reasoning or overfitting.
Every decision, data access, model invocation and action is logged with complete provenance, confidence levels, alternatives considered and governance approvals.
Scenario simulation applies the Possibility Ladder to hypothetical futures. Before any significant decision, the system generates a comprehensive set of scenarios, not just the most likely outcomes, but the full range of what can happen.
Pre-mortem analysis asks: assume this decision failed, what went wrong? By starting from the failure state and working backward, the system identifies risks that forward looking analysis misses.
The Risk Assessment gate is non-negotiable. No action proceeds without it. The system evaluates the worst case outcome, whether the decision is a One Way Door or Two Way Door, the cost of being wrong the cost of not acting and what information would change the decision.
Every action the system takes generates feedback. This feedback is not just right or wrong, it is a structured decomposition of what happened and why.
Credit assignment traces the outcome back through the Possibility Ladder to identify the specific step where the error occurred: observation, possibility space, context, probability estimate, decision logic or execution.
Memory uses Regime Indexed Recall rather than temporal decay. Experience is permanent, nothing is forgotten. Activation depends on regime similarity, not recency. A lesson learned three years ago in a matching regime is more relevant than a lesson learned yesterday in a different regime.
The Hub and Spoke architecture is the convergence mechanism that enables multi-source intelligence. Each spoke is an independent processor with its own data source, normalization, memory and recall.
Spokes do not communicate with each other; they only communicate with the hub. This independence ensures that no spoke is contaminated by another’s biases.
Confidence indicator. Independent spokes converge on the same conclusion. Strength scales with spoke independence and historical reliability.
Red flag. Disagreement between spokes indicates unusual conditions. Conflict is the important information it triggers investigation not averaging.
Caution indicator. One or more spokes cannot form a view. The current situation is outside their experience, so the system proceeds with reduced confidence.
Every decision the system makes or recommends is classified along a reversibility spectrum. This classification determines the confidence threshold required, the governance review process and the speed of execution.
Extra scrutiny. Higher confidence threshold. Board review required. Full rollback plan documented before execution. Post-implementation review mandatory.
Moderate confidence sufficient. Bias toward action. Standard governance review. Automatic rollback if metrics degrade. Post-hoc review acceptable.
Virtual Tables allow you to connect existing data sources without replication. Instead of copying data into Otonomii’s storage, Virtual Tables create a semantic layer on top of your existing databases, data lakes and APIs.
The system queries your data where it lives, applying its cognitive layers to data in place. Deployments that would take weeks with migration can take days with Virtual Tables. Data sources can be added, removed or modified without rebuilding the cognitive pipeline.
Otonomii’s AI/ML capabilities are not built by applying the latest techniques from the research literature. They are built by decomposing problems to their fundamental truths and discovering solutions from first principles.
This is not anti-academic, it is anti-herd. The difference between applying known solutions and discovering new ones is the difference between following the market and understanding it.
Is this what everyone does? If so, why would it give us an edge? The default assumption is that popular approaches produce average results. Before adopting any technique, algorithm or architecture, the system asks who else is doing this and what are their results. If the answer is everyone and the results are mediocre, the approach is rejected regardless of theoretical elegance.
Start from what we actually know to be true, not what the literature says. Most AI/ML knowledge is herd-derived, biased toward prediction and leads to the same fragile systems everyone else builds. Otonomii decomposes every problem to its fundamental truths and builds upward from there.
Relentless questioning exposes hidden assumptions. Every design decision, parameter choice and architectural pattern is subjected to structured questioning: why this approach, what are the alternatives, what assumption does this depend on and what happens if that assumption is wrong.
The goal is to discover new ways, not apply existing solutions. Most platforms apply known techniques to new data. Otonomii discovers new techniques from first principles, validated against data. The system is a discovery engine, not an application engine.
Before investing resources in solving a problem, the system challenges the problem statement itself. Many AI/ML failures trace back not to poor solutions but to poor problem formulation. A perfectly solved wrong problem is worse than an approximately solved right problem.
The Semantic Data Layer is the Brain. The domain agnostic memory substrate that stores, retrieves, associates and compresses all information the system encounters.
It handles multi-modal data: structured data, unstructured data, streaming data and geospatial data. The data layer is cross cutting as it imports nothing from any cognitive layer. This ensures that memory is not contaminated by interpretation. The Brain stores what happened; the Mind decides what it means.
Commits new observations to memory with source, timestamp, confidence and regime context.
Retrieves relevant patterns using similarity matching that accounts for regime context.
Creates associative connections between patterns that co-occur, enabling contextual recall.
Reorganizes memory, merges redundancy, strengthens validated links and compresses rarely accessed information.
The Dynamic Logic Layer is the Mind, where raw memory becomes understanding. This layer transforms stored patterns into predictions, measures the gap between predictions and reality and uses that gap as the primary learning input.
It is inspired by predictive processing theory from neuroscience: the brain is fundamentally a prediction machine and experience is shaped by prediction error. The difference between what was expected and what actually happened.
Converts raw observations into internal representations optimized for prediction.
Generates expectations about the next state based on current context and historical patterns.
Computes prediction error across magnitude, direction and attribution.
Updates internal models continuously and incrementally based on prediction error.
Allows multiple internal models to express independent opinions.
Determines how much attention each input receives based on regime and reliability.
The Kinetic Action Layer is the Machine, where understanding becomes behavior. This layer translates the Mind’s assessments into concrete actions in the real world.
It answers four questions: when to act, when to stop, what to do and what resources are available. Self healing operations emerge from the interaction between Machine and Mind: identify the discrepancy, trace it to its source, generate a corrective action, execute the correction and verify the result.
Determines when confidence is sufficient to act. It rises in uncertainty and drops in well understood reversible regimes.
Determines when to stop searching, close a position or accept the current answer as good enough.
Translates assessments into concrete operations in the real world.
Manages computational budgets, operational limits, API quotas and time constraints.
Every agent, model and pipeline operates under a defined role with explicit permissions. An agent can read market data without being able to place orders or generate recommendations without executing them.
Data and decisions are classified by sensitivity and reversibility. One Way Door decisions require higher confidence thresholds and additional review gates.
Every data access must declare its purpose and access is limited to what is necessary for that declared purpose. Purpose declarations are logged and auditable.
Before any output is finalized, the system asks whether this is the obvious conclusion and what may be missing. This prevents groupthink among model ensembles.
Unexpectedly perfect results trigger automatic investigation for data leakage, circular reasoning or overfitting.
Every decision, data access, model invocation and action is logged with complete provenance, confidence levels, alternatives considered and governance approvals.
Scenario simulation applies the Possibility Ladder to hypothetical futures. Before any significant decision, the system generates a comprehensive set of scenarios, not just the most likely outcomes, but the full range of what can happen.
Pre-mortem analysis asks: assume this decision failed, what went wrong? By starting from the failure state and working backward, the system identifies risks that forward looking analysis misses.
The Risk Assessment gate is non-negotiable. No action proceeds without it. The system evaluates the worst case outcome, whether the decision is a One Way Door or Two Way Door, the cost of being wrong the cost of not acting and what information would change the decision.
Every action the system takes generates feedback. This feedback is not just right or wrong, it is a structured decomposition of what happened and why.
Credit assignment traces the outcome back through the Possibility Ladder to identify the specific step where the error occurred: observation, possibility space, context, probability estimate, decision logic or execution.
Memory uses Regime Indexed Recall rather than temporal decay. Experience is permanent, nothing is forgotten. Activation depends on regime similarity, not recency. A lesson learned three years ago in a matching regime is more relevant than a lesson learned yesterday in a different regime.
The Hub and Spoke architecture is the convergence mechanism that enables multi-source intelligence. Each spoke is an independent processor with its own data source, normalization, memory and recall.
Spokes do not communicate with each other; they only communicate with the hub. This independence ensures that no spoke is contaminated by another’s biases.
Confidence indicator. Independent spokes converge on the same conclusion. Strength scales with spoke independence and historical reliability.
Red flag. Disagreement between spokes indicates unusual conditions. Conflict is the important information it triggers investigation not averaging.
Caution indicator. One or more spokes cannot form a view. The current situation is outside their experience, so the system proceeds with reduced confidence.
Every decision the system makes or recommends is classified along a reversibility spectrum. This classification determines the confidence threshold required, the governance review process and the speed of execution.
Extra scrutiny. Higher confidence threshold. Board review required. Full rollback plan documented before execution. Post-implementation review mandatory.
Moderate confidence sufficient. Bias toward action. Standard governance review. Automatic rollback if metrics degrade. Post-hoc review acceptable.
Virtual Tables allow you to connect existing data sources without replication. Instead of copying data into Otonomii’s storage, Virtual Tables create a semantic layer on top of your existing databases, data lakes and APIs.
The system queries your data where it lives, applying its cognitive layers to data in place. Deployments that would take weeks with migration can take days with Virtual Tables. Data sources can be added, removed or modified without rebuilding the cognitive pipeline.
2026 © Otonomii LTD. All rights reserved.
TOP
2026 © Otonomii LTD. All rights reserved.
TOP
2026 © Otonomii LTD. All rights reserved.
TOP