85.7% to 100% error elimination on repeated tasks. 4.8% Expected Calibration Error on self-assessment. 53% reduction in dispatch logic.
These are not three separate results from three separate systems. They come from one architecture with seven functional knowledge layers — and the most important finding is not any single number. It is that we added two entirely new cognitive capabilities without changing a single line of existing layer code.
Why Layers Matter More Than Types
Most AI systems organize knowledge by format: text here, vectors there, tables elsewhere. We organize knowledge by function — what the knowledge is for:
| Layer | Function | Example |
|---|---|---|
| Y0 | Raw sources | Original text chunks with provenance |
| Y1 | Semantic | Entities, relationships, concepts |
| Y2 | Reasoning | Rules, ontology, constraints |
| Y3 | Experience | Conversations, actions, learning events |
| Y4 | Journal | Opinions, reflections, mental notes |
| Y5 | Procedural | Compiled DAG workflows for repeated tasks |
| Y6 | Metacognitive | Calibration tracking, error classification |
The first five layers (Y0-Y4) were built over multiple development cycles. Y5 and Y6 were added for this paper. The separation is what made that possible — each layer has its own schemas, its own access patterns, and its own evolution path.
Procedural Memory That Actually Works
Y5 is where a being compiles repeated reasoning paths into directed acyclic graph (DAG) workflows. Instead of re-reasoning through a multi-step task every time, the being recognizes the pattern and executes the compiled procedure.
The results speak for themselves. On tasks the being had seen before, procedural memory eliminated 85.7% to 100% of reasoning errors depending on the task type. The floor is 85.7% because some tasks have genuine novelty — but for the repetitive backbone of domain work, compilation turns slow reasoning into fast execution.
A Being That Knows When It Is Wrong
Y6 is metacognition — the being monitors its own accuracy. We measure this with Expected Calibration Error (ECE), which tracks how well the being's confidence matches its actual performance. An ECE of 4.8% means the being is accurate to within about 5 percentage points when it says "I am 80% confident." It knows what it knows, and more importantly, it knows what it does not know.
The metacognitive layer also achieved 100% error classification accuracy — when the being makes an error, it correctly categorizes what went wrong (knowledge gap, reasoning failure, context misunderstanding, etc.). This is the foundation for targeted self-improvement.
The Architecture Validates Itself
Here is the result that matters most for anyone building cognitive architectures: Y5 and Y6 were developed independently, by different engineers, without modifying any existing layer code. The dispatch logic that routes knowledge to the right layer was reduced by 53% because each layer handles its own concerns.
This is what separation of concerns looks like when it works. You do not prove an architecture is good by building one thing well. You prove it by building the second thing and discovering that the first thing did not need to change.
The paper combines three studies — procedural memory, metacognition, and the full Y0-Y6 integration — into a single evaluation of the complete cognitive stack.