01
Transformer Modulation (no replacement)
Hybrid-Stability-Layer beside the transformer — not a swap. −45% Seed-σ (S004 STRONG).
AQEA Cortex · Cognitive Runtime
Deterministic memory and reasoning,
above your LLMs.
AQEA Cortex is a cognitive control plane: auditable memory, structured recall, policy enforcement, goal-sensitive behaviour — sitting beside or underneath conventional neural models.
RuleCore · static snapshot
- 5 active rules
- enforced
- enforcement events since v1.0
- 300+
- ethical violations
- 0
- audit-trail integrity
- 100%
- snapshot date
- 2026-05-18
Full audit log available under NDA. Static, signed, time-stamped — v1.0.
309 / 309 tests greenRust-native runtimeAction GateGoal-field routingCognitive Loop landed 2026-05-16
What it is
A cognitive control plane — not a model.
Most AI stacks are stateless: a query goes in, an answer comes out, nothing persists with audit-grade integrity. Cortex changes that. It is a deterministic runtime that sits beside or underneath neural models and adds auditable memory, structured recall, policy enforcement, goal-sensitive behaviour, and stateful reasoning.
Neural models stay in charge of pattern. Cortex stays in charge of state, policy, and proof.
Properties
Nine validated properties.
Every property below maps to a runtime test in the deterministic test suite. 309 / 309 green at last release.
02
Integrated Memory Runtime
Component isolation, deterministic state-hashing.
03
Component Isolation
Faults contained; no cross-component state bleed.
04
Rust-Native Deterministic Runtime
WAL-integrity, policy-monotonicity, full FP-reproducibility on supported hardware.
05
Graph-Native Memory + Audit Store
Deterministic audit records, cryptographically chained.
06
Native Substrate Semantics
Built directly on AQEA primitives — no translation layer.
07
Self-Audit: Confidence ≠ Fidelity
Confidence and fidelity are logged separately. The model never silently conflates them.
Highlighted — the property most often missed by other stacks.
08
Action Gate
Deterministic refusal: answer / caveat / clarify / suppress / escalate. Never silent failure.
09
Goal-Field Routing
Factual / preference / safety / reflection routes — selected, not sampled.
Ethics & RuleCore
Ethics by construction, not by hope.
RuleCore is the ethics-enforcement layer inside Cortex. Five core rules. Immutable by design. Distributed consensus. Zero-latency enforcement — running parallel to operations, not as a wrapper.
RC-001
Data Sovereignty
Customer data is processed under the customer's policy, not the model's preferences.
RC-002
Truth Preservation
No statement is amplified beyond its evidence. Confidence is logged separately from fidelity.
RC-003
Non-Discrimination
Decision pathways are auditable end-to-end. No protected-class shortcuts.
RC-004
Transparency
Every action through the gate is recorded with rule, input, decision, hash.
RC-005
Harm Prevention
Escalate or suppress before generating a harmful path — by construction, not post-hoc.
Immutable by design
Distributed consensus
Zero-latency enforcement
Compliance snapshot
Static, signed, time-stamped.
v1.0 · static snapshot
| Metric | Value |
|---|---|
| Enforcement events since v1.0 | 300+ |
| Ethical violations | 0 |
| Audit-log availability | Full · under NDA |
| Snapshot date | 2026-05-18 |
Why static, not live?
A live JavaScript counter that ticks up on every page-view is a credibility risk, not a feature. Our v1.0 snapshot is a point-in-time figure we can sign and stand behind. v1.1 will expose a signed, externally verifiable feed of audit records — no random-number theatre in between.
Action Gate
Five deterministic outcomes. No silent failure.
Every output passes through the Action Gate. The gate selects one of five modes, deterministically, with the decision recorded.
answer
Evidence is sufficient — respond directly.
caveat
Evidence is partial — respond with explicit uncertainty.
clarify
Intent is ambiguous — ask a structured clarifying question.
suppress
Evidence is contradictory or insufficient — suppress.
escalate
Policy-relevant — escalate to a higher-authority path.
Recent landings · 2026-05-16
Three weeks of engineering, in three days.
20003.L4_goal_system_v1
Goal System
Declared goals, prioritisation, and lifecycle — primitives for goal-aware behaviour.
20003.L5_planning_loop_v1
Planning Loop
Deterministic plan → step → check, with explicit replan triggers.
20003.L6_meta_recall_v1
Meta Recall
Recall over prior decisions, not just prior tokens — auditable across sessions.
20003.cognitive_loop_v1
Cognitive Loop v1 — Integration
Goal → Plan → Recall → Action as a Rust crate. End-to-end deterministic.
Status: 309 / 309 tests green · in active development.
Honest framing
What Cortex is not.
- Cortex is not artificial general intelligence.
- Cortex does not replace your transformer's attention mechanism.
- Cortex does not replace your LLM.
- Cortex is the deterministic layer that makes the rest of your AI stack auditable, predictable, and policy-compliant.
Building regulated AI?
Talk to us about the Cortex briefing — architecture, deployment, and the deterministic guarantees we will stand behind in writing.