For classification, scoring, and routing, a model does not always need to write a sentence. A typed decision model returns a constrained judgment—such as a choice, score, or yes/no probability—that application code can use directly. You can put a hosted model or a self-hosted open model in that role, while keeping thresholds, permissions, and escalation in your software. Google Cloud documents GPU-backed Cloud Run services and BigQuery remote functions as possible deployment and invocation patterns, but neither platform feature proves a particular model will be accurate, fast, or cheaper for your workload.
What a System 1 decision model returns
In this context, “System 1” borrows Daniel Kahneman’s terminology for fast judgments; it is not a guarantee of human-like reasoning or correctness. The DEV Community article by Francisco Riveros, reported in search as published September 22, 2026, describes Jev as a decision model: a caller supplies input such as text or JSON, along with questions whose answers are constrained in advance. The result is structured for software to consume rather than composed as a free-form response.
As an Amazon Associate I earn from qualifying purchases.
The System One Models directory describes three question shapes. These are category-level descriptions, not a promise that every model implements the same limits, probability calibration, or API.
| Shape | What the caller asks | Reported output |
|---|---|---|
| Choice | Pick one option from a caller-defined set of candidates. | A selected candidate; the directory reports up to 255 candidates for the category it covers. |
| Score | Assess input against ordered levels. | A probability-weighted mean over two to ten levels, according to the directory. |
| Noul | Answer a yes/no question. | A probability from 0 to 1 for “yes.” |
A bounded answer can make downstream handling simpler than extracting a label from generated prose. It does not establish that the judgment is factually right, that its probability is calibrated, or that the question itself captures the policy you need.
When to use a decision model—and when not to
This approach fits tasks with a stable input, a clearly specified judgment, and a known set of possible outcomes. Examples include assigning a support request to one of a fixed number of queues, scoring urgency against ordered levels, or estimating whether a defined condition is present. Keep the decision policy in ordinary application code: the model supplies a judgment, while software decides what to do with it.
- Use a constrained output when the task and valid answers can be defined before the request arrives.
- Keep permissions, thresholds, retries, audit logging, and escalation rules outside the model.
- Use a generative model when the user needs an open-ended explanation, novel synthesis, or a response whose content cannot be represented by a bounded answer.
A fast-path decision model followed by a generative fallback is one possible architecture, not a guarantee that any fixed share of requests can safely bypass generation. Decide which outcomes qualify for automatic handling only after evaluation on your own cases.
Rank #2
Hosted Jev and open models: compare the workload, not the label
The Riveros article names hosted Jev from TypeSafe AI and open implementations including SemIf and Laya. The System One Models directory also lists hosted and open entries and describes Laya as self-hosted under Apache 2.0. Availability, versions, license terms, prices, and hosted-service conditions can change; confirm them with the relevant project or provider before adopting a model.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errors| Choice | What changes operationally | What to verify |
|---|---|---|
| Hosted decision model | A provider operates the service; your application sends requests to it. | Request and context limits, regional availability, data handling, service terms, price, and measured quality for your task. |
| Open, self-hosted model | Your team operates the model in its own environment and manages the serving path. | License and model version, hardware fit, deployment effort, throughput, monitoring, and full operating cost. |
Riveros reports Jev latency of 70–500 ms and input pricing of $0.042 per million tokens. Those are figures reported by that article, not a neutral benchmark or a guarantee for a particular request, current price, or deployment. They should not be compared with self-hosted measurements unless the task, input size, batching, concurrency, hardware, and timing conditions are made comparable.
AutoTrust’s JEV-27B model card reports an 84.07% mean across six benchmarks and a 137 ms median single-decision latency on one B200 GPU, along with its own comparison runs against hosted Jev. These are developer-reported results for that model and evaluation setup, not evidence that one model family is generally more accurate, faster, or cheaper. The sources cited here do not establish a controlled, category-wide comparison or a universal speedup.
Evaluate quality, calibration, and operating cost
Start with representative historical examples and hold out cases that were not used to tune prompts, thresholds, or model selection. Have qualified reviewers establish reference labels, especially for ambiguous or high-impact cases. Measure more than aggregate accuracy: a routing model that misses urgent cases may be unacceptable even if its overall score looks strong.
- Define the decision and the cost of errors. Specify the valid choices or score levels, then record the relative costs of false positives, false negatives, and abstentions or escalations.
- Test on held-out examples. Compare model outputs with human-reviewed labels; inspect errors by class and by ambiguous-input type rather than relying on one overall figure.
- Check probability calibration. Group predictions by stated confidence and compare those probabilities with observed outcomes. Choose thresholds from your application’s error tolerance, not from a generic recommendation.
- Measure the complete request path. Record end-to-end latency and throughput at expected payload sizes, concurrency, and request mix; include cold starts where relevant.
- Calculate total cost at realistic usage. Include API charges or GPU time, minimum resources, storage, networking, monitoring, and the engineering and operational work required to run the service.
- Set escalation and monitoring rules. Define which low-confidence, out-of-distribution, or otherwise risky cases need review, then track errors and calibration as inputs and model versions change.
A probability is useful only if its behavior has been checked on the application’s data. Treat confidence thresholds as policy you own, not as a built-in safety boundary.
Run an open model on Google Cloud Run with a GPU
Google Cloud documents NVIDIA L4 support for Cloud Run services. Its GPU documentation lists 24 GB of VRAM and minimum service resources of 4 CPUs and 16 GiB of memory for an L4 service. Google also documents that GPU-enabled Cloud Run service instances can scale down to zero when not in use. That capability does not make every deployment cost-free: storage, networking, other services, and configured minimum instances can still incur charges.
Before adapting a deployment example, verify current regional availability, quota, concurrency behavior, resource settings, and billing constraints in Google Cloud’s service documentation and your project configuration. A model’s memory fit is only one requirement; serving framework, loading behavior, request shape, and expected concurrency also affect whether a deployment works for your workload.
Cloud Run can provide a managed place to serve an open model, but the platform capability is distinct from a performance result. Benchmark the actual model and endpoint with your payloads, including startup and idle behavior, before estimating latency or cost.
Call a decision service from BigQuery
BigQuery remote functions let GoogleSQL invoke external software through a Cloud Run functions or Cloud Run endpoint. This can make a decision service callable from a query, for example when a data workflow needs to apply a defined judgment to records. BigQuery’s documentation specifies supported argument and return data types and other limitations; design the function interface around those constraints.
Recommended Free Tools
Remote-function support verifies that the invocation pattern exists, not that it will be appropriate for bulk scoring, a particular query volume, or a cost target. Test end-to-end behavior—including request volume, service capacity, errors, retries, and query cost—before using it in a production pipeline. Do not infer a speedup or savings estimate from the fact that SQL can call the endpoint.
Make the deployment decision with evidence
Choose hosted or self-hosted based on measured task quality, calibrated probabilities, governance needs, operational capacity, and total cost. Keep the control flow and consequences in code, and retain a generative model for work that genuinely needs open-ended synthesis. The figures published by individual projects and the Cloud Run and BigQuery capabilities describe specific products and platform features; they are not substitutes for evaluating your own workload.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
