# Using Jev (TypeSafe AI): Current Practice and Open Problems (November 2026)

- Author: Shuichiro Ogawa (Design Researcher / Design Consultant), https://jamjet-iqbal.com/
- Published: 2026-11-21; updated: 2026-11-21
- Canonical (HTML): https://jamjet-iqbal.com/en/notes/jev-best-practices-2026-11/
- Japanese version: https://jamjet-iqbal.com/notes/jev-best-practices-2026-11/ (Markdown: https://jamjet-iqbal.com/notes/jev-best-practices-2026-11.md)
- Language: en
- Keywords: ai, jev, typesafe, system-one, decision-model, best-practices, confidence-threshold, selective-prediction, cascade, criteria, evals, guardrail, ai-frontier

> Two months after TypeSafe AI released Jev, the material for checking how to use it has grown well beyond the few items available in September. Pre-registered independent evaluations sit on GitHub, and an arXiv search for preprints on Jev returns 63 hits. None of this material has passed peer review, however, and no record was found of running Jev in production for several weeks or more. Of the September conclusions, the guidance on writing criteria and on not delegating arithmetic or dates was supported by independent evaluations. The largest change concerns thresholds. A threshold has to be chosen again for each task, and a threshold point-estimated from a few examples tends to fall short of the target accuracy. A cascade is better read as a way to lower cost than as a way to raise accuracy, because Jev and LLM judges err in the same places. On how to provide abstention, the TypeSafe docs have moved to asking with a separate Noul, but no figures comparing the two forms exist yet. When labeled data are at hand, several cases show a small trained classifier outperforming Jev.

How Jev was used as of September is summarized in [How to Use Jev, and Where Not To](https://jamjet-iqbal.com/en/notes/jev-usage-patterns/).
The conclusion at that time was that no consensus that could be called best practice existed yet.
This note returns to the same question on 2026-11-21.
It separates the September conclusions that still hold from those that have changed.
What Jev guarantees and what it does not guarantee is covered in [TypeSafe AI's Jev: What It Guarantees and What It Does Not](https://jamjet-iqbal.com/en/notes/typesafe-ai-jev/).

## Which September Conclusions Still Hold

The table below gives the verdicts first.
The "Level of support" column sorts the strength of the backing into three levels.
An official recommendation is guidance that exists only as text in the docs of TypeSafe AI or a distribution provider.
A single implementation is a measurement that one implementer made on their own.
An independent replication is a measurement that someone other than the provider made with published code.

| September item | November verdict | Level of support | Main evidence |
|---|---|---|---|
| Score: describe situations, not degrees | Kept | Official recommendation | Docs, Score |
| Noul: one yes/no question per Noul | Kept | Official recommendation | Docs, Noul |
| Noul: avoid negated phrasing | Revised (frame the question so that a high value means yes) | Official recommendation and independent replication | Docs, Noul; Rafe and Das |
| Do not delegate arithmetic or date calculations | Kept | Official recommendation and single implementation | Model jaggedness; GC AI's record |
| Start from confidence 0.6 and 0.85 | Revised (no fixed starting point) | Independent replication | Li et al.; Brom; MLflow |
| Provide an abstain or "none of the above" option | Unsettled (the vendor has moved to asking with a separate Noul) | Official recommendation | Docs, Choice and companion Noul |
| Raise accuracy with a cascade | Revised (the main effect is on cost) | Independent replication | Rao and Callison-Burch |
| Bundle questions into one request (12.2x cheaper) | Kept (still an official recommendation only) | Official recommendation | Cookbook, parallel_questions |
| Check both input and output with guardrails | Kept (with an added condition) | Official recommendation and single implementation | Spring AI's record |

## What Evidence Has Grown, and What Is Still Missing

In September there were a few independent implementations, with samples of 12 to 60 items.
By November, several pre-registered independent evaluations were on GitHub, with samples of hundreds to thousands of items.
A search of arXiv for Jev together with "TypeSafe AI" or "System One" returned 63 hits, and the collector checked all of the titles.

Some things are still missing.
Zero peer-reviewed papers were found, and within one web search, no record of publication appeared in the ACL Anthology or OpenReview.
Zero records of running Jev in production for several weeks or more were found either.
In other words, no record reports the volume, the automation rate, the misjudgment rate, and the cost of rework.
No record was found from anyone who removed Jev and moved to another method.
None of the following three figures cited in September has been re-measured by anyone other than its source.

- Langfuse's agreement rate of 91.5%
- 91.7% on operation-risk judgment
- 12.2x from batching

For analyst reports, Gartner, Forrester, IDC, McKinsey, Menlo, and ICONIQ were checked.
The full texts that could be reached did not mention Jev (Gartner's full text could not be reached).
Documents from public bodies that name Jev also numbered zero.

## Choose Thresholds Again on Your Own Data

The September note introduced the three-tier thresholds in the docs as a starting point.
In that scheme, a judgment goes to a person when confidence is below 0.6, and it is executed automatically when confidence exceeds 0.85.
The confidence page of the docs, retrieved again for this note, no longer contains these three-tier figures.
Instead, TypeSafe describes what confidence measures.
"Confidence measures how concentrated the answer's probabilities are, which is different from the probability that the selected answer is correct".

Independent evaluations produced results that argue for caution about setting any fixed starting point.
Li et al. tested, across several tasks, a setup that accepts only confident Jev judgments and escalates the rest to an LLM judge.
The authors conclude that the threshold must be set on the target task.
As a procedure, they recommend labeling about 100 items of the target task and choosing the threshold from them.
When choosing the threshold, they use the lower bound of a confidence interval on accuracy rather than a point estimate.
They report that this lowers the risk of losing 2 points of accuracy or more from about 45% to about 5%.

It has also become clear that a threshold set on a small number of examples fails to keep the target accuracy.
Brom reported that, in 190 evaluation runs, thresholds chosen by point estimates fell short of the target accuracy 147 times.
With a selection method that statistically guarantees an upper bound on the error rate (Learn then Test), the target was met in every run.
Šuppa and Kopčan examined thresholds carried across languages.
When a threshold tuned to a 5% error rate on English data was carried over to German and French XNLI, the error rate was between 7.8% and 12.4%.

Records from implementers point in the same direction.
Yayoi and Watanabe of MLflow calculated what would happen if only judgments with probabilities between 0.2 and 0.8 were routed to another judge (GPT-OSS-120B).
Fewer than 20% of the judgments would be routed, and in both runs all 72 cases would then match the human labels.
However, the two authors write that this range was set by looking at the same data, so it should be treated as a starting point and not as a rule.
When the same input was sent repeatedly, one judgment was wrongly accepted at 0.82, outside the range.
In an evaluation by the independent implementer ickma2311, confidence was 1.0 for 102 of 200 items, and 6 of those were wrong.

Three practices follow from these results.

- Choose the threshold on labeled data from your own task, and do not carry it over to another task, language, or model version
- Treat a threshold set on a few dozen items as a starting point
- Re-select the starting threshold with a method that guarantees the error rate, or check it on held-out data

TypeSafe itself writes on the Noul page: "Avoid fitting thresholds closely to a handful of test cases".

## How to Provide Abstention Is Still Unsettled

Jev's Choice always selects one of the options, even for an input that none of the options fits.
The September note recommended providing an abstain or "none of the above" option.

The recommendation in the TypeSafe docs has moved during these two months.
The Choice page still says "Add an `other` or `none of the above` option when the list might not cover every input".
The annotation directly below it says "A `none of the above` option may not work well in many use cases", and it points to the companion Noul instead.
The companion Noul page states that adding `other` used to be the standard advice.
It then says that `other` competes with the real options for probability.
The pattern it describes therefore asks, in a separate Noul within the same request, whether any of the options applies.
The only basis given for this form is "In our tests, that usually works better", with no figures attached.

Independent results do not converge either.
Chen and Le ran a pre-registered evaluation of a gate that judges whether a memory can be retired, and they changed the two options to three.
The added option was "needs confirmation".
Wrong retirements fell from 39.2% to 15.5%, but detection of items that really should be retired dropped by 15.3 points.
This exceeded the 10-point tolerance that the authors had set in advance.
Watanabe of MLflow writes that if abstention is needed, one should use a Choice that has an "unsure" option.

No evaluation was found that compares the abstain-option form and the separate-Noul form on the same data.
For now, what can be done is to try both forms on your own data.
When doing so, count both wrong automatic actions and missed cases.

## Use Cascades to Cut Cost

A cascade processes inputs with a cheap model and passes only the uncertain ones to an expensive model.
The vendor has recommended this form since September.
What the November independent evaluations found is that its effect on accuracy is small.

Rao and Callison-Burch compared Jev and Perplexity's Decider with flash-tier LLM judges on nine evaluation panels.
The LLM judges cost 16 to 357 times as much as Jev or the Decider.
On the judgments where Jev was most confidently wrong, however, 96.0% of the LLM judges' verdicts gave the same answer as Jev.
If the errors were independent, the expected rate of matching answers would be 50.3%.
The authors describe Jev as "a cheap substitute for a flash-tier LLM judge but a poor complement to one".
Even when the routing threshold was chosen after the fact to be the most favorable, the gain over the strongest single judge did not exceed 2.7 points.

The TypeSafe cookbook also presents the effect on the cost side.
On a calendar task of 1,000 items, the cheap model alone reached 92.6% accuracy.
The reasoning model alone reached 98.1%, at $1.89 per 1,000 items.
Routing to the reasoning model only the items that Jev judged to violate a rule gave the following results.

- Routing about 13%: 96.5% ($0.49)
- Routing about 54%: 98.1% ($1.31)
- Routing more than 64% gave no further gain in accuracy (98.3% at $1.50 when routing 64%)

Two things should be checked before building a cascade.
The first is whether the later model answers correctly where the earlier stage errs.
The second is the share of items routed, together with the cost and latency that routing adds.
Measure both on your own data, and record them together with accuracy.

## Independent Evaluations Also Support the Criteria Guidance

The ways of writing criteria introduced in the September note remain in the November TypeSafe docs.
For Score the docs say "Describe situations, not degrees", and for Noul they say "Ask one yes/no question per Noul".

On negated Noul questions, TypeSafe's guidance has become more detailed.
The docs now ask that each question be framed so that a high value means yes, in place of a blanket rule against negated questions.
The docs give "Is the message free of personal data?" as an example of a question whose meaning can be misread when code reads the answer later.
Rafe and Das ran a test that swapped the direction of yes and no.
Jev's answer changed in 50.5 of every 100 items, whereas the four trained judgment models changed 6.5 or fewer.

Writing rule boundaries with examples and counterexamples has also gained independent support.
Beno studied a task in which rules stop coding-agent actions, and added examples and counterexamples that describe the rule boundaries.
According to the author's abstract, missed detections decreased for all 23 models that could be evaluated.
The recall of Jev 1.14 rose from 89.3% to 98.2%.
In the condition where GPT-6 wrote the rule text, however, two models had confidence intervals for the gain that included 0.

The order and wording of options also need attention.
TypeSafe's model jaggedness page says that Jev can lean toward the first option.
As a countermeasure, it recommends checking whether the answer changes when the options are reordered.
Song et al. report that reordering the options moved mean accuracy by less than 2 points.
Even so, the set of problems answered wrongly under some ordering grew by between 27.3% and 28.5% relative to the original order (the range across two benchmarks).
In the evaluation by Li et al., repeating the same request 96 times did not change Jev's judgment.
On the other hand, 4 of 48 paraphrases of the rubric changed the judgment.
Test the criteria on paraphrased versions and on versions with the options reordered, and keep the wording unchanged once it has passed these checks.

## The Effect of Bundling Questions Rests on the Vendor's Measurement Only

The practice of bundling questions that may be related into one request is still recommended in the November docs.
The cookbook figures were measured again on jev-1.14.0.
Compared with sending 13 requests sequentially, the result is "12.2x cheaper and 11.8x faster with no change in answers" (in September, on jev-1.12, the figure was 10.0x faster).
The same page also says that the speed gap narrows when single calls are sent concurrently.
The cost gap remains as it is.

No record was found of anyone other than the vendor measuring this figure again.
Pourvakil of GC AI describes an attempt to pack many similar items into one array for Jev to judge.
Answers mixed between adjacent rows, and accuracy fell from 96% to 72%.
Because the conditions differ from bundling separate questions into the same state, this result is counted neither as a refutation nor as support.

## Tasks Where Trained Classifiers Win

In September there was almost no material for comparing Jev with alternatives when deciding whether to adopt it.
By November, several comparisons with small classifiers trained on labeled data had appeared.

In ickma2311's pre-registered evaluation, Jev reached an accuracy of 0.832 on Banking77, which has 77 classes.
A small encoder trained on about 10,000 labels from the same distribution reached an accuracy of 0.933.
The author notes that a result from one dataset and one label budget does not make a general rule.
Iqbal et al. report that on hate-speech detection, DeBERTa trained on 100 to 250 labels outperformed Jev.
On CLINC150, DeBERTa did not clearly outperform Jev at any label budget.
Jev also does poorly on tasks where the answer cannot be known without deriving it.
In the evaluation by Li et al., accuracy on logic puzzles was 68.4% for Jev and 95.9% for GPT-6.

Conversely, Jev wins where work starts without labels, and on speed and cost.
The implementer anisselbd tested Jev on phishing detection.
Jev's accuracy as a direct judge was 62.6%, below the 91.6% of a two-line rule that uses a list of link destinations.
In the same call, however, Jev was also asked to return five signals.
When these signals were fed into a logistic regression, accuracy on held-out data was 95.0%.
The author notes that the classes in this dataset are easy to separate because of how it was built.
There was no statistically significant difference from having Claude Haiku 4.5 answer the same five signals (p = 0.063).
The advantage left to Jev was that it was about 27 times cheaper and 5 times faster.
Passing Jev's answers as features to another mechanism that makes the final decision is consistent with TypeSafe's "Jev makes judgments, code makes decisions".

## What Happened in Production

There are no operational records, but there are reports of outages and defects.
From around September 26, Jev accessed through Vercel's AI Gateway kept returning 429 for some teams.
Vercel staff explained that, as a mechanism to protect service reliability, use may be temporarily limited for some accounts.
The TypeSafe status page lists two incidents in October.
On October 14, latency was elevated for about 45 minutes in the us-east-1 region, and on October 30 the service was overloaded for about 1 hour 15 minutes.
In September, the page lists three incidents of 2 to 12 minutes.

Another report describes Cloudflare blocking valid requests with an HTML 403 when the state contains a curl command.
Because this looks like an authentication failure, the cause is hard to track down.
In the implementation that Tzolov of Spring AI published, a passage passes through unchecked when the Jev API cannot be reached.
The only signal at that point is a warning log.
If you delegate input guardrails to Jev, configure them to stop when Jev cannot be reached (fail-closed).
Test that behavior as well.

The model version has also changed.
On 2026-10-14, jev-1.14.0 was released, and `jev-latest` now points to 1.14.0.
TypeSafe writes that it intends to give 60 days' notice before retiring 1.13, but it has not yet published a retirement policy.
If you chose thresholds on your own data, call the model with an ID that pins the version.
Choose the thresholds again every time you change versions.

## Competitors and the Scale of Adoption

OpenAI and Google have also released APIs of the same kind as Jev.
On 2026-11-11, Simon Willison wrote about Google's Gemini Decision 1.
According to him, it has almost the same three types as Jev, and it even uses Jev's term "Noul" for the yes/no type.
On price, Jev is the cheapest of the three APIs.
Willison compares Jev as a model that handles only text and JSON.
Version 1.14 added image input as an experiment, but TypeSafe itself writes on its blog that these experimental features, including images, are not yet at a level it can recommend for production use.

The only figures on the scale of adoption come from gateway operators.
In Vercel's AI Gateway figures, decision models accounted for 26% of requests in October.
Decision models here mean models that, like Jev, return only classifications, scores, and selections.
Their share of spending in the same month was 0.53%.
The cost per request was about 1/70 of that of language models, so the requests are many but the payments are small.
Vercel writes that most requests to decision models went to Jev.
These are request counts, and they do not show production operation.
They may include trial and evaluation requests.

No job posting that lists Jev as a requirement was found, apart from TypeSafe's own postings.

## External Conditions for Production Use

Three of the norms cited in the September note have changed.

The application of the EU AI Act's obligations for high-risk AI systems (Annex III) was postponed from 2026-08-02 to 2027-12-02 (Regulation (EU) 2026/1744).
The text of Art 14, which sets out human oversight, has not changed.
Art 14(4)(b) requires that the people who oversee a system be enabled to remain aware of the tendency to over-rely on its output (automation bias).

Colorado's SB24-205 was repealed and re-enacted as SB26-189.
Technologies that assist decisions with classifications or scores now fall within its scope, and developers' obligations begin on 2027-01-01.
People who receive an adverse outcome can request meaningful human review and reconsideration.
Under the statute's definition, the person who conducts that review must not treat the system's output as the default answer ("does not default to the system output").
Developers notify deployers of significant updates.
Both developers and deployers keep records for three years or more.
The practice of pinning the model version and choosing thresholds again after a version change also fits the obligations to notify updates and to keep records.

The OMB memorandum M-24-10 for US federal agencies had been replaced by M-25-21.
The September note cited the version from before the replacement.

## The Academic Lineage

The questions behind many of the evaluations cited above did not originate with Jev.
Peer-reviewed research has accumulated on the following three problems.

- Classification with abstention
- Calibration of probabilities
- Performance degradation on long inputs

The problem of choosing thresholds has been studied as selective classification.
Geifman and El-Yaniv added abstention to trained networks and presented a method that keeps a user-specified error rate with high probability.
Brom's Learn then Test result builds on this line of work.

The caution against reading confidence directly as the probability of being correct is also old.
Guo et al. showed that the probabilities of modern neural networks tend to deviate from their actual accuracy (that is, they are poorly calibrated).

Research on long inputs supports the practice of keeping irrelevant content out of the state.
Levy et al. ran an experiment that lengthened inputs with text unrelated to the task.
The added text alone lowered the reasoning performance of LLMs at lengths far shorter than the maximum the models can handle.
No peer-reviewed study has yet measured this on Jev itself.

Enforcing an output format and getting the judgment right are separate matters.
Tam et al. report that the stronger the format restriction on output, the lower the reasoning performance of LLMs.

## Reading the Tiers

The T1v vendor primary sources are TypeSafe AI's docs, cookbook, status page, and blog.
They were treated as the primary source on usage and constraints.
The cookbook figures are TypeSafe's own measurements.
The method is disclosed, but there is no third-party verification.
The claims about a Fortune 500 adoption rate and customer savings have no methodology, so they are not used in the body.

The T2 public bodies and research are the primary texts of the EU AI Act, the Colorado statute, and the OMB memorandum, together with Vercel's usage figures.
The number of analyst reports that name Jev was zero.
Figures from investor materials were excluded because the investors' interests cannot be separated from them.

The T3 individuals and independent implementations are records from individuals whose authority could be confirmed, and pre-registered evaluations on GitHub.
The individuals are the MLflow maintainers, the Spring AI lead, and Simon Willison.
The GitHub evaluations are measurements by single implementers, and their authority was not confirmed.

All of the arXiv papers that evaluated Jev are preprints that have not been peer reviewed.
The figures cited here from papers were checked against the papers' full texts, but figures can change during peer review.

The sources agree on the following three points.

- How to write criteria
- Not delegating arithmetic and dates
- The order of magnitude of Jev's speed and cost

They diverge on how to provide abstention and on accuracy when questions are bundled.
Until long-term operational records appear, the effects of thresholds and cascades can only be measured task by task.

Related notes: [The Structure of a Claude Code Environment with a Jev Judgment in Every Exchange](https://jamjet-iqbal.com/en/notes/claude-code-ecosystem-jev-review/), [How to Use Jev, and Where Not To](https://jamjet-iqbal.com/en/notes/jev-usage-patterns/), [TypeSafe AI's Jev: What It Guarantees and What It Does Not](https://jamjet-iqbal.com/en/notes/typesafe-ai-jev/), [When Judgment Becomes a Function Call, What Moves in Design](https://jamjet-iqbal.com/en/notes/jev-typed-judgment-design/), [Recent Currents in LLM-as-a-Judge: A Literature Map of 53 Core Studies and a Standalone Chapter on Creativity Evaluation](https://jamjet-iqbal.com/en/notes/llm-as-a-judge-literature/), [Choosing Between Frontier and Cheaper Models: Claude Fable 5 and GPT-5.6 Sol](https://jamjet-iqbal.com/en/notes/model-tier-selection-ai-frontier/), [Has RAG Fallen Out of Use? What Replaced It, and What Remains](https://jamjet-iqbal.com/en/notes/rag-grounding-2026-ai-frontier/)

## Unverified Items

- No record of running Jev in production for several weeks or more was found. The count was zero within web searches and searches of GitHub issues and PRs, which does not prove that no such record exists
- The full texts of Gartner and CB Insights could not be reached (403), so it cannot be said for certain that they do not mention Jev
- The statement "P(noul) ≠ 1 - P(not_noul)" cited in the September note was not found in the docs retrieved this time. Whether it was removed has not been confirmed

## References

All accessed 2026-11-21.

### Vendor primary (T1v)

- TypeSafe AI. Docs: Confidence. https://docs.typesafe.ai/confidence.md
- TypeSafe AI. Docs: Choice. https://docs.typesafe.ai/primitives/choice.md
- TypeSafe AI. Docs: Noul. https://docs.typesafe.ai/primitives/noul.md
- TypeSafe AI. Docs: Score. https://docs.typesafe.ai/primitives/score.md
- TypeSafe AI. Docs: Pattern, companion Noul. https://docs.typesafe.ai/patterns/companion-noul.md
- TypeSafe AI. Docs: Model jaggedness. https://docs.typesafe.ai/model-jaggedness.md
- TypeSafe AI. Docs: Models. https://docs.typesafe.ai/models.md
- TypeSafe AI. Cookbook: Parallel questions. https://docs.typesafe.ai/cookbooks/parallel_questions.md
- TypeSafe AI. Cookbook: SDE cascade. https://docs.typesafe.ai/cookbooks/sde_cascade.md
- TypeSafe AI. "Jev: For the builders AND dreamers". 2026-10-14. https://typesafe.ai/blog/jev-1-14
- Laabs, A. (TypeSafe AI). "Don't Call It a Decision Model". 2026-11-03. https://typesafe.ai/blog/dont-call-it-a-decision-model
- TypeSafe AI. Status. https://status.typesafe.ai

### Public bodies and research (T2)

- EU. Regulation (EU) 2024/1689 (AI Act, Art 14). https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202401689
- EU. Regulation (EU) 2026/1744 (amendment to the AI Act). https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=OJ:L_202601744
- Colorado General Assembly. SB26-189, Concerning the Use of Automated Decision-Making Technology in Consequential Decisions. 2026-05-14. https://leg.colorado.gov/bills/SB26-189
- Colorado General Assembly. Session Laws 2026, Chapter 131 (SB26-189 as enacted). https://leg.colorado.gov/laws/session-laws/4864/download
- OMB. M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust. 2025-04. https://www.whitehouse.gov/wp-content/uploads/2025/02/M-25-21-Accelerating-Federal-Use-of-AI-through-Innovation-Governance-and-Public-Trust.pdf
- Charles, A., & Arora, H. (Vercel). "AI Gateway Production Index: November 2026". 2026-11-09. https://vercel.com/blog/ai-gateway-production-index-november-2026

### Individual opinion and independent implementation (T3)

- Yayoi, T., & Watanabe, Y. (Databricks). MLflow blog, Jev judge part 2. 2026-09-25. https://mlflow.org/blog/jev-llm-judge-part-2/
- Watanabe, Y. (Databricks). MLflow blog, Jev judge. 2026-09-22. https://mlflow.org/blog/jev-llm-judge
- Tzolov, C. Spring blog, Spring AI modular RAG with TypeSafe Jev. 2026-10-02. https://spring.io/blog/2026/10/02/spring-ai-modular-rag-typesafe-jev
- Willison, S. "Gemini Decisions API". 2026-11-11. https://simonwillison.net/2026/Nov/11/
- ickma2311. jev-baselines-eval. 2026-09-25. https://github.com/ickma2311/jev-baselines-eval
- anisselbd. jev-phishing-bench. 2026-10-20. https://github.com/anisselbd/jev-phishing-bench
- Pourvakil, B. (GC AI). "What we learned building 38 projects with Jev in 12 hours". 2026-10-06. https://gc.ai/blog/what-we-learned-building-38-projects-with-jev-in-12-hours
- Vercel Community. "TypeSafe AI Jev requests shed with 429 and providerAttemptCount 0". 2026-09-26. https://community.vercel.com/t/typesafe-ai-jev-requests-shed-with-429-and-providerattemptcount-0-per-team-throttling/49779
- josebreijo. typesafe-sdk-js issue #15, Cloudflare blocks valid /v1/systemone requests. 2026-09-23. https://github.com/typesafe-ai/typesafe-sdk-js/issues/15

### Academic (preprints)

- Li, Y., Miao, Y., Krishnan, R., & Padman, R. (2026). JEV-as-a-Judge: Accept When Confident, Escalate When Unsure. arXiv:2609.26550v6. https://arxiv.org/abs/2609.26550
- Rao, D., & Callison-Burch, C. (2026). Decision Models vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places. arXiv:2609.29769v3. https://arxiv.org/abs/2609.29769
- Brom, P. C. (2026). Confidence Is Not a Promise: A Reference Audit of Typed Decision Models. arXiv:2610.36311. https://arxiv.org/abs/2610.36311
- Šuppa, M., & Kopčan, J. (2026). XDecideBench: Evaluating Typed Decision Models and Confidence Policies Across Languages. arXiv:2611.01370v2. https://arxiv.org/abs/2611.01370
- Chen, X., & Le, C. (2026). Suggested Is Not Superseded: A Pre-Registered Evaluation of a System-One Model as a Memory-Retirement Gate. arXiv:2610.15710. https://arxiv.org/abs/2610.15710
- Rafe, A., & Das, S. (2026). Benchmarking System One decision models against trained classifiers and language models for automated decision gates. arXiv:2610.00346v2. https://arxiv.org/abs/2610.00346
- Beno, J. P. (2026). Decision Models as Policy Gates for Coding Agents: Explicit Policy Boundaries Improve Detection and Reduce False Alarms. arXiv:2611.02155. https://arxiv.org/abs/2611.02155
- Song, M., Zheng, M., Ma, Z., & Zhou, X. (2026). Understanding Jev-Style System One Models: Does Structure Reduce Selection Bias? arXiv:2610.35826v3. https://arxiv.org/abs/2610.35826
- Iqbal, A. R., Nasim, M. F., & Saleem, A. (2026). Typed Decision Models under Prevalence Shift and Label Budgets: Decision Training, Calibration, and Implicit Priors. arXiv:2610.26325v2. https://arxiv.org/abs/2610.26325

### Academic (peer-reviewed)

- Geifman, Y., & El-Yaniv, R. (2017). Selective Classification for Deep Neural Networks. NeurIPS 2017. https://arxiv.org/abs/1705.08500
- Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On Calibration of Modern Neural Networks. ICML 2017. https://arxiv.org/abs/1706.04599
- Levy, M., Jacoby, A., & Goldberg, Y. (2024). Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models. ACL 2024. https://arxiv.org/abs/2402.14848
- Tam, Z. R., Wu, C.-K., Tsai, Y.-L., Lin, C.-Y., Lee, H.-y., & Chen, Y.-N. (2024). Let Me Speak Freely? A Study On The Impact Of Format Restrictions On Large Language Model Performance. EMNLP 2024 Industry Track. https://aclanthology.org/2024.emnlp-industry.91
