Hand routine decisions to software only as fast as your team's evidence allows. Your reviewers confirm past outcomes, the Cognitive Engine turns them into rules, and no rule decides before it clears an exam on cases it never saw. Qualified rules need no AI call; unfamiliar cases return to your review queue, or go to AI where your team allows it. Start in shadow mode. Published results use synthetic data.
Inputs from your records
Rule in plain English
Mark it OK when the value stays in its normal band, with no frozen sensor and no signal loss.
Exam
1,453 of 1,453 right
Held back from training
Reasoner consensus
6 of 6 agree
From the Factory Floor simulated business. Measured in October 2026 on synthetic data.
Nx·Studio builds the digital workers your teams run day to day. Nx IDE gives your engineers governed AI help. The Cognitive Engine, a separate product, takes over routine decisions an AI agent would otherwise send to a model case after case, once each rule clears its exam. Nxgyle supplies all three.
When every routine case goes to an AI model, your team waits on each answer, pays for each call and sees answers vary. Over a full shift that adds up case by case. The practical question for a manager is which decision types already have enough confirmed outcomes to run on a qualified rule.
Every incoming case is checked against the rules your team has allowed to decide. Matches are settled in milliseconds, with no AI call, and the same rule release gives the same answer each time. Cases no qualified rule covers go where your mode sends them: to a reviewer's queue or, where you permit it, to AI within budget.
New cases
Rulebook
First check
Routed to
Illustration only. The tallies are generated by the animation, not by testing; measured results appear lower on the page.
Your team owns the facts and the final say on every decision type. The engine learns only from what your team has confirmed, and each rule must qualify before it acts. AI helps only where you allow it. Your reviewers settle the conflicting and the unfamiliar, so accountability never blurs.
Keeps the records and confirms what really happened. Nothing else counts as evidence for the engine.
Offers suggestions during training. If your team permits it, it also handles cases beyond the engine's training. Its suggestions never count as proof.
Builds rules from confirmed outcomes. A decision type becomes its job only after the exam is passed.
Works the confirmation queue, settles cases where past decisions conflict and, in the strictest mode, takes every case no rule covers.
Every decision type moves through the same five steps, in this order. Your team grants authority one decision type at a time, and only on evidence.
Training uses past cases whose outcome is confirmed, read from a source system or signed off by a named reviewer. AI suggestions may inform training but are never accepted as proof.
The maintenance log shows which past readings were normal and which ones needed action.
Each candidate rule is tested on past cases kept out of training. It must answer 400 or more of them correctly, with no misses, before it gets decision rights. Your team may raise that bar but never lower it.
On a simulated factory's synthetic data, a rule that missed 1 case of 515 was refused decision rights. After more cases were confirmed, it reached 1,453 of 1,453.
A qualified rule settles each case in milliseconds without a model, and the same release gives the same answer to the same inputs. All six built-in reasoners must agree on the rule; a single objection stops it.
A normal reading from a healthy sensor is marked OK, and the record shows which rule decided.
No rule is stretched over a case outside its training. Your chosen mode decides what happens next: reuse a precedent with identical inputs, consult AI within budget, or hand the case to a reviewer. Conflicting precedents are always handed to a reviewer.
A reading outside every learned range is never guessed: a reviewer takes it, or AI where you allow that.
Each rule keeps its authority for 30 days by default. If a confirmed outcome proves it wrong, the rule loses its rights at once and must pass a new exam. A policy change works the same way: the old rule retires and its successor qualifies afresh.
When your team redefines a normal reading, the first confirmed exception retires the old rule, and a successor has to qualify through a fresh exam.
Most of a sensor-monitoring shift is routine readings. The engine settles only the reading types its rules have qualified for. Other readings are matched to your team's earlier rulings or passed to budgeted AI, and a named reviewer receives every hand-off.
Nine readings are waiting in the queue.
Exam-qualified rules; millisecond answers, no AI.
Reuses a past ruling on identical inputs, or consults AI under a budget.
Hand-offs, plus cases where past rulings conflict.
An illustration based on Factory Floor, a simulated business of ours built on synthetic data, where routine sensor readings have qualified. In your operation the split depends on which decision types qualify and which mode your team selects. In every mode, rule decisions never call AI, AI use stays within your budgets, and people handle every hand-off.
A qualified decision type runs with no model charge. Remaining AI use stays inside a per-run budget and an optional daily cap.
How our fees workYour team gets consistent calls: under the same rule release, the same inputs return the same result, whereas an AI-only agent can drift from day to day. A re-run shift repeats every rule decision exactly.
Each engine decision records its rule in plain English, and all six built-in reasoners must agree on that rule first. Every decision is logged, Studio activity is recorded in an audit trail, and hand-offs always land with a reviewer.
Controls, in detailYour team sets the engine's reach for each application and each run. In our simulated test businesses, a comparison run places it beside an AI-only agent.
Day-to-day work is untouched, and your existing process still makes every decision. The engine handles the same cases alongside it, purely for scoring, so your team can review its record before it affects anything.
Sequence of authority
Qualified rules take the first pass. Next comes precedent, your team's own ruling in exactly the same situation, unless past rulings conflict, when a reviewer decides. AI sees only genuinely new situations, within budget.
Sequence of authority
No AI takes part in any decision: qualified rules settle what they can, and every other case goes to the fallback your team names, usually a reviewer.
Sequence of authority
Operations and risk staff use the Studio to gather evidence, train and test the engine, compare it with an AI-only agent and sign off before live use. Access comes in three roles: viewer, trainer and administrator. Reviewers work from a confirmation queue of their own.
Lists its agents, the hard limit set for each, outcomes that must never occur and the decision types qualified so far.
On our simulated businesses, one action scores the engine beside an AI-only agent for accuracy, reviewer workload, speed and AI calls. On your own system it is scored against the decisions that system already makes.
Read every learned rule in plain English, next to its exam record and the verdicts of all six built-in reasoners.
Reviewers clear one queue of cases awaiting confirmation. In blind review the suggested answer stays hidden, so each judgment is the reviewer's own.
AI drafts each decision's inputs, limits and policy from your documents and systems, citing the source of every point. A person on your team must approve a draft before it applies.
Each application comes with a checklist for going live, covering a security review, load tests, restore drills, a handover pack and the owner's sign-off.
Viewer, trainer and administrator roles keep access apart, identifiers are masked, and Studio activity is kept in an audit trail.
Give each run an AI budget and, if you want, each application a daily cap. One switch decides whether AI ever sees case facts.
Flags expiring or withdrawn rules, slow decisions, AI spend and low disk space for the team running it.
Nxgyle tests the engine in twenty simulated businesses, all on synthetic data, with cases drawn from claims handling, payments, plant monitoring and energy markets. The comparison point is a typical agent that asks an AI model for every decision.
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.
One simulated factory decision, two days: given the same model and instructions, an AI-only agent scored 97% on the first day and 20.5% on the second. Over the same 865 readings, the engine's proven rule made no mistakes on either day.
All figures were measured by Nxgyle in October 2026 on simulated businesses with synthetic data. “Right” means matching each simulated business's own policy. They show the method, not a customer's results, and do not by themselves prove the engine is ready for live use.
Nothing your team runs today has to be replaced: the engine adds a qualified decision step beside your agents and systems. Your records remain the source of what actually happened.
Agent platforms ask the Studio's web API for a decision, each with an access key of its own.
Send the engine a copy of the decisions your current system makes. It works the same cases without acting and is compared with your results; new situations become cases your reviewers confirm.
Your AI assistants can consult the engine through a local MCP server (MCP is a common protocol between AI tools and software). It can advise; it never acts. For direct embedding there is a Python SDK.
A Docker Compose package installs it on servers your team controls. After an upgrade, every rule must requalify through its exam before acting.
Careful with your systems. Each action the engine takes through your tools is recorded as an intent before it runs. The engine never retries an action on its own, and an uncertain result keeps the case blocked until your team resolves it.
Controls in the Trust CenterWhat the prototype does today, and the options your team can enable.
Only where you permit it. The engine handles only routine decision types it has qualified for; new cases go to your reviewers or, with your permission, to budgeted AI.
Its exam. A rule has no decision rights until it passes, and it loses them when contradicted or expired. Anything without a qualified rule follows your mode: precedent, budgeted AI or a reviewer.
Any confirmed result that disproves a rule withdraws it immediately. Every rule also lapses after a 30-day default term, and its successor must pass a fresh exam.
Yes. Every engine decision records its rule in plain English, and the Studio's Transparency screen adds the rule's exam record and the verdicts of all six reasoners.
By default, the facts of cases where AI helps, after names and identifiers from your sources are masked. Your team can set any application or data source to share nothing with AI, and turn off drafting from your documents.
Not yet. Every figure here comes from synthetic data in simulated businesses. Shadow mode is how your team measures the engine on its own cases before granting any decision rights.
