9 out of 9 hypotheses validated. Zero hallucination. 98% routing accuracy. P95 latency of 7.7 milliseconds — 13x under our 100ms target. And 100% LLM call elimination for known facts.
These aren't cherry-picked numbers from a demo. They come from Paper 108, where we tested a unified cognitive architecture across beings ranging from 14,000 to 525,000 knowledge triples. The architecture held at every scale.
Every Input Follows the Same Path
The core design principle is simple: every input — question, statement, command — flows through a single 7-step perception cycle.
perceive → classify → route → respond → crystallize → validate → commit
There are no special cases. No "if it's a medical query, do this; if it's chitchat, do that." The routing is emergent — determined by the being's confidence and the novelty of the input, not by hand-coded rules.
This is informed by four decades of cognitive architecture research. ACT-R gave us the production system model. SOAR showed us how chunking works. CLARION demonstrated the symbolic-neural split. LIDA provided the perception-action cycle. We didn't invent these ideas — we unified them into a single routing model that actually runs in production.
The 5-Path State Plane
The perception cycle doesn't just route to "LLM" or "knowledge graph." It routes across five paths based on two dimensions: confidence (high/low) and novelty (known/novel).
| Path | Confidence | Novelty | What Happens |
|---|---|---|---|
| Symbolic retrieval | High | Known | Answer from graph. No LLM. Fast. |
| Confident generation | High | Novel | Graph has related knowledge, LLM synthesizes with provenance. |
| Gap-driven crystallization | Low | Known | Source prose has the answer — crystallize it into the graph. |
| Exploratory learning | Low | Novel | Neither graph nor prose covers it — defer, don't fabricate. |
| Metacognitive reflection | Any | Any | The being assesses its own calibration and coverage. |
The critical paths are 1 and 4. Path 1 is the fast lane — the being knows the answer, returns it from the graph, done. Path 4 is the honest lane — the being doesn't know and doesn't pretend. Together, these two paths eliminate hallucination by construction.
Path 3 is where continuous learning happens. The being detects a gap, finds the answer in source prose, and crystallizes it — permanently converting "I don't know" into "I know this, and here's where I learned it."
The Routing Accuracy Number
98% routing accuracy means the perception cycle correctly classified inputs into the right path 98% of the time. The 2% misclassification is almost entirely between paths 2 and 3 — the being thought it could synthesize from related knowledge when it should have crystallized from source prose instead.
This is a recoverable error. The answer is still grounded (path 2 requires related graph knowledge). It's just suboptimal — the being used the LLM when it could have used the graph directly. Future calibration runs will close this gap.
Scale Independence
We validated across a wide range of knowledge graph sizes:
- 14K triples — a freshly created being with minimal training
- 148K triples — a domain-specialized being after curriculum study
- 525K triples — a heavily trained being across multiple domains
The architecture held at all three scales. Routing accuracy was consistent. Latency stayed under 8ms at P95. Hallucination rate stayed at zero.
This matters because it means the architecture scales with knowledge — the symbolic path gets better as the graph grows (more queries hit path 1 instead of paths 2-4), but the routing logic doesn't need to change.
82% Task Autonomy
One of our 9 hypotheses tested whether the perception cycle could handle multi-step tasks without human intervention. We gave beings tutoring tasks — explain a concept, assess understanding, provide practice problems, evaluate responses. They completed 41 of 50 task loops autonomously, routing through the perception cycle for each interaction, against a target of 80%.
82% task autonomy doesn't mean the beings were perfect teachers. It means 41 of the 50 assigned loops ran end to end without human intervention for routing decisions — no "I don't know what to do next" failure, no getting stuck, no hallucinated task step. The other nine wanted a human in the loop, which is what an 80% target exists to allow for: the claim is that the architecture routes multi-step work on its own most of the time, not that it never needs help.
What Crystallization Does Here
In the unified architecture, crystallization isn't a separate system — it's path 3 of the state plane. When the perception cycle detects low confidence on a known topic, it routes to crystallization. The being finds the answer in source prose, extracts it into a triple, and the next time that topic comes up, it routes to path 1 instead.
Knowledge grows monotonically. Every crystallization event moves future queries from the neural paths (2-4) toward the symbolic path (1). The being gets faster and more reliable over time because its graph keeps growing from verified sources.
This is the feedback loop that makes the architecture self-improving — not in the "the model fine-tunes itself" sense, but in the "the knowledge graph accumulates verified facts" sense. Every interaction is either answered correctly or converted into an opportunity to learn from source material.
Read the full paper: Unified Cognitive Architecture — Zenodo