A decision layer that sits beside your agents. It starts with no domain knowledge and trains solely on outcomes that your systems of record or named reviewers have verified. Before a rule may decide, it must pass a held-out exam of 400+ cases with zero misses and unanimous agreement from its six reasoners. Qualified rules then run on standard servers without an LLM call; anything the engine is unsure of is routed to a reviewer, or to an AI fallback only if you enable one.
Input facts from the host
Rule text, as stored
Return OK if the reading is in band, the sensor is responsive and the signal has no dropouts.
Holdout exam
1,453 of 1,453 right
Excluded from training
Six-way agreement
6 of 6 agree
From the Factory Floor simulated business. Measured in October 2026 on synthetic data.
Nx·Studio generates digital workers and exports each one as complete source code. Nx IDE brings governed AI modes into VS Code. The Cognitive Engine is a stand-alone product, not yet integrated with Nx·Studio, that answers repeat decisions for your agents without an LLM call once its rules qualify. Nxgyle builds all three.
Most agent frameworks route every decision through an LLM, including cases the system has resolved hundreds of times. At production volume that means seconds of latency per call, token spend that grows with traffic, and outputs that drift as models and prompts change.
Each request is evaluated against the rules qualified for its domain. A match returns a decision with its explanation and no LLM call. No match, or facts outside the trained range, goes to review, and missing facts trigger a request for more. The mode decides whether budgeted AI may advise.
Requests
Released rule set
Evaluated first
Route
Animation only: the counters are illustrative, not measured. Measured results appear further down.
The architecture splits authority. Your host application owns the job, the tools and the source of truth; an advisory AI model only proposes. The engine learns from verified outcomes and earns permission to operate, while named reviewers verify what happened.
Owns the job, the tools and the record of what happened. Verification means an authoritative state read or a named person, never a second model agreeing.
Suggests candidate answers during training and, if the caller allows it, on unfamiliar cases. Nothing it returns is treated as evidence.
Learns decision rules from verified outcomes only, and decides alone only after it passes its exam.
Confirm outcomes, resolve conflicting history and take every case sent for review. In Cognitive mode, anything outside the qualified rules lands with them.
Every decision type follows one pipeline inside an isolated domain. Qualifying one domain never qualifies another, and hosts can only raise the bars.
Training data is limited to outcomes the host marks as verified: an authoritative state read or a named person's confirmation. A new engine starts with no domain knowledge, and AI suggestions never count as evidence.
The plant historian supplies past readings with confirmed outcomes: normal, or intervention needed.
Defaults: 30+ training cases and 5 examples per rule, then 400+ unseen cases all correct, a statistical confidence bar of at least 99% and six-way reasoner agreement. You can raise these bars, never lower them.
In a simulated factory running on synthetic data, one rule missed 1 case in 515 and was kept from acting. After more evidence it scored 1,453 of 1,453.
A qualified rule runs on standard servers without GPUs, typically in tens of milliseconds, with no call to a large language model. Every rule needs unanimous agreement from its six built-in reasoners, and any one of them can veto it.
A steady reading from a responsive sensor returns OK, with the rule and every reasoner's conclusion attached.
Facts outside the trained range go to review, and missing facts trigger a request for more. An uncertain tool result holds the affected record until your systems confirm what happened. Depending on the mode, consistent prior decisions or budgeted AI come first.
A value beyond the training range is sent for review instead of getting an extrapolated answer.
Each qualified rule is tied to the exact definition, software version and evidence it was examined on. A verified contradiction, any change, or its 30-day default expiry withdraws it until it is re-examined.
When the plant changes its normal band, the first verified contradiction withdraws the release, and the replacement rule sits its own exam.
One monitoring agent, nine requests. Qualified rules take the repeat readings; unqualified cases fall back to consistent prior decisions or budgeted AI, and hand-offs land in the review queue. Run it twice to check reproducibility.
Nine sensor requests are queued.
Released after its exam; answers without a large-model call.
A consistent prior decision on identical facts, else budgeted AI if allowed.
Hand-offs and inputs with conflicting history.
Illustrative trace based on the Factory Floor simulated business (synthetic data), where one decision type has qualified. Real splits depend on the decision types that qualify and the configured mode. In every mode, qualified rules make no LLM call and AI spend stays within budget; hand-offs reach a person unless an AI route is enabled.
Qualified paths incur no model spend. Calls to an outside AI model happen only when allowed, inside the budgets you configure in the Studio.
Pricing modelSame release, same facts, same output: replay a shift and every qualified decision returns unchanged. Uncertain tool outcomes block the resource until the host reconciles them, with no blind retry.
Each output states its rule in plain words, with the exam it passed and every reasoner's result. Studio actions are logged, and each hand-off reaches a named person.
Trust Center evidenceSet the mode per application. On the Studio's simulated businesses, each mode can be benchmarked against an LLM-only agent on accuracy, latency, LLM calls and review volume.
Your existing system keeps authority. The engine evaluates the same traffic from a parallel feed and is scored against your outcomes; nothing it returns is acted on.
Evaluation order
Qualified rules answer first. Otherwise a prior decision is reused only when identical facts were decided one way at least 99 times in 100. Conflicting history goes to a person, and only new facts reach the LLM.
Evaluation order
Only qualified rules decide. Everything else goes to the fallback you configure, and where inaction is unsafe that fallback needs deliberate design.
Evaluation order
The Cognitive Studio wraps the engine in a web API, trains and examines it, and benchmarks it against LLM-only agents on 20 synthetic test-bench apps. Permissions come in three tiers (viewer, trainer, administrator), and reviewers verify cases from a single queue.
Agents, hard limits, forbidden outcomes and every qualified decision type, per application.
On the 20 synthetic test-bench apps, one run scores the engine against an LLM-only agent on accuracy, latency, LLM calls and review volume. On your own system, the baseline is your system's own decisions.
Shows any learned rule in plain words, the exam it cleared and all six reasoners' outputs for each decision.
Reviewers verify cases from one queue. Exam cards hide any suggested answer, so verification stays independent of the engine.
AI drafts each decision's facts, limits and policy from your documents and systems, citing a source per point. Nothing counts until a person approves the draft.
Per application: a go-live checklist, owner sign-off, a security review, load and restore drills, and a handover package for operations.
Three permission tiers, identifier masking by policy and logged Studio actions. Operational authentication for production remains deployment work.
Per-run AI budgets and an optional daily ceiling per application, plus one setting that controls whether case facts are ever sent to AI.
Alerts for rules nearing expiry or withdrawn, slow decisions, AI spend and low disk space.
Twenty synthetic businesses make up the harness, from insurance claims and payments to energy markets and factory floors. Each scores the engine against an agent that calls an LLM on every request, with identical inputs.
Field test on five simulated businesses with synthetic data, October 6, 2026.
Every decision the engine made itself, on all 20 simulated businesses (synthetic data, 1,000 events each, October 9, 2026). The engine currently decides cases in 14 of them.
Typical decision: 17–31 ms for the engine vs 3.3–4.2 s for an AI-only agent. Five simulated businesses, synthetic data, October 8, 2026.
Same simulated runs, synthetic data: per 1,000 decisions, the AI-only agent made 329–1,519 AI calls; Cognitive + AI mode made 0–14. Most AI-only figures are scaled up from runs capped at 300 calls.
Same simulated factory decision, same model and prompt: the LLM-only agent scored 97% on day one and 20.5% on day two. The qualified rule decided the same 865 readings across both days, error-free.
Measured by Nxgyle in October 2026 on simulated businesses with synthetic data; “right” means agreeing with that simulated business's policy. The numbers demonstrate the method, not customer results, and zero errors on an evaluation set does not prove zero errors on future inputs.
No replatforming: the engine runs beside your existing agents and services, and the host's records stay authoritative for outcomes.
The engine itself has no web interface. The Cognitive Studio wraps it and gives each of your agent platforms a decision endpoint.
Stream decisions your system already made. The engine evaluates them silently and is graded against your outcomes, and unfamiliar inputs become cases for verification.
A local tool server lets AI assistants consult the engine, which advises but never acts. Hosts integrate through a code-level API, and every tool action is recorded so it never runs twice.
Packaged for servers you run and tested on Linux; Windows and macOS are not yet verified. After any upgrade, every rule goes back to review until it is re-examined.
Safe side effects. Every tool action is recorded before it runs, so it cannot execute twice. No action is ever retried by the engine itself, and an unclear outcome holds the affected record until your systems confirm what happened.
Inspect the Trust CenterA correctness-focused prototype rather than a tuned high-volume service. Scaling, production authentication and deployment remain work for your platform team.
During training it proposes answers, and on unfamiliar cases it advises only with your permission. An optional path sends declined requests to an AI model and a second checking model within an hourly budget; that advice acts only after its own exam.
No. Only released rules decide, and the test-only switch never acts in production. Everything else goes to review or follows the mode you configured.
A verified contradiction, any change to its definition, software or evidence, or its 30-day default expiry. Until it is re-examined, its cases go to review.
How the case was routed, which checks ran and what each reasoner concluded, with zero LLM calls recorded for rule decisions. The Transparency screen adds the exam the rule cleared.
The engine goes online only to reach an outside AI model, during training or where you opt in. When AI is consulted, masked case facts are shared by default; turn that off for an application or a single data source.
No. Every published figure comes from simulated businesses with synthetic data, and real-domain accuracy has not been measured. Shadow mode scores it on your own traffic before it gets any authority.
