One window, several models: a reading of Model Functionality Monitoring

What changes in this design

You might expect each model to be watched on its own and around the clock, with a running score kept for it and a decision made about that single model whenever its score slips. Judging never stops, and anything that gets fixed gets fixed one model at a time.

Watching here is boxed into time windows handed down from the network side. One message sets up the window, or windows. A second says whether performance inside a window landed above or below a level. Acting on that, the device manages the group of models carrying out the function rather than a single one.

Two ordinary ideas end up separated: the stretch of time you look at, and the verdict on how it went. Because the verdict arrives as a message rather than being worked out on the spot, the device can respond without judging for itself, and its response covers every model doing that job.

2 August 2026

On 2 August 2026, continuous performance monitoring stopped being a nice-to-have for providers of high-risk AI systems in the European Union. They must collect, document and analyse performance data through a system's life under a written plan. That legal duty is where I start, because it tells us what network patents are trying to own: machinery for noticing failure.

That doesn't mean every network tool is suddenly regulated as high-risk. Once monitoring is a legal duty in one setting, defining windows, thresholds, reports and choices becomes valuable infrastructure elsewhere. That is the setting in which I read the U.S. application called Model Functionality Monitoring.

The terminal and the network

US20260254730A1 is assigned to Nokia Technologies Oy. Its core move is modest but revealing. A terminal receives one set of information configuring time windows for watching an AI/ML function, and another marking a window above or below a performance level. It uses that second signal in a management operation involving multiple models. The division of labor is clear to me: the network supplies monitoring terms, while the device acts on a model set.

The related international filing makes the reason plainer: a network can police performance without knowing exactly which model a terminal is using. That matters when a terminal must function through changing conditions, but it does not promise a correct diagnosis. A low score might call for selection, switching, fallback, update, or another operation in a wider lifecycle system. The application stakes out the handoff. That handoff is where, I think, durable value may sit. The count I found was 71 related entries, with filings from 2022 through 2024. This is not a quiet technical corner.

Release 19

This is not a patent appearing before its field has language for the job. In Release 19, 3GPP put AI/ML lifecycle work into an AIMLE framework covering management, training, inference and analytics. The framework includes continuous checks on performance metrics, data drift and inference accuracy, with retraining or updates when results fall below defined thresholds. The patent looks to me like a contender for the interfaces inside that agenda, especially between network policy and terminal action. An open radio project had built the same loop: training, deployment, monitoring and retraining, though early model serving was manual and file-based.

A radio-interface study from the same effort names functionality and model monitoring alongside training, delivery, selection, switching and fallback. Its debates were still testing whether monitoring belongs at the user device or network, and whether a device may make management decisions. That disagreement strikes me as more revealing than a polished architecture diagram. The split in Nokia's filing is not a settled implementation detail. It is a claim on a question engineers were still negotiating.

91 percent

Why monitor so closely? A 2022 study summarized by a monitoring company ran 20,000 history-future simulations across 128 model-dataset pairs and found temporal degradation in 91 percent of cases. Error and its variability rose as models aged. That result is, I think, a warning, not a universal law: the proposed remedy depends on retraining alerts and recent ground truth, which can arrive late or cost too much. Monitoring can identify decay. It cannot restore judgment.

That is why I am skeptical of grand claims for network AI. One industry critic argues that changing networks impose a permanent drift tax: retraining for one change may erase performance learned elsewhere, and probabilistic systems are ill-suited to deterministic five-nines operations. That isn't, in my view, the last word. But it describes a cost that a monitoring patent cannot wish away. A system that watches carefully may also discover more situations demanding expensive intervention.

Where this leaves things

The financial evidence lags engineering confidence. One industry report put global telecom operating expenditure down only 0.2 percent in 2024, while a cited telecom executive estimated AI savings at about 2 percent of a EUR 550 million cost base. A group with Nokia among its founders grew to 132 members and announced trials after 33 AI-RAN demonstrations. The sector appears to me to be preparing to deploy while still arguing over the returns.

That is the tension I take from US20260254730A1. Nokia isn't, as I read it, claiming to solve model decay, network economics, or regulatory compliance in a single stroke. It is claiming a way to organize the response once performance is measured over time. If this layer becomes common infrastructure, patents around windows, thresholds and authority may matter long after today's technical fashions change. If the promised savings never arrive, the same filings will read as evidence of an industry that built careful controls before it proved there was enough value to control.

Sources

Reporting for this piece drew on the published filing and the references below.

1. US20260254730A1: Model Functionality Monitoring
2. AIMLE (Artificial Intelligence Machine Learning Enablement) - 3GPP glossary
3. TS 38.843 Section 4.2 - Life Cycle Management
4. Article 72: Post-Market Monitoring - EU AI Act
5. 91% of ML Models Degrade in Time
6. Enabling Intelligent RAN Framework in O-RAN
7. Don't expect AI to boost telco profits
8. Why network foundation models and AI-RAN won't save telecom
9. AI-RAN Alliance Reaches Major Milestone

The record, briefly. US20260254730A1, “Model Functionality Monitoring” A terminal device receives from a network device first information configuring at least one time window for monitoring AI/ML model functionality performance, and second information indicating a window where performance is above or below a performance level; based on the second information it performs a management operation on a plurality of AI/ML models carrying out that functionality.

Identity details. The saved title-page transcription needs review, so names are not repeated here. Check the official published title page.

More angles on this

• CPC H04L: Transmission of digital information, e.g. telegraphic communicationThe official classification family this record was filed into, published in the USPTO Cooperative Patent Classification scheme.
• A Reading Taken in the Ear, and the File It Has to ReachAnother reading from the neighbouring H04R classification family, loudspeakers, microphones, gramophone pick-ups or like acoustic electromechanical transducers.
• Next Generation Prenatal Screening: A Probability Where a Verdict Was ExpectedAnother reading of a record filed by the same applicant, (52).
• Keeping Time Across a StackAnother reading from the wider H10B classification family, electronic memory devices.
• The Quiet Work of Alignment Around a TargetAnother reading in the same Objects category on this site.

Read around this one:Public patent recordUSPTO published figure