Bring the AI key you already use. Add a few lines of code. Your AI gives one clear answer and how sure it is.
Which team should handle this email?Billing, 97% sure
Curva vs Jev on a public phishing benchmark
Curva's best run on each measure, with the model named under each chart. Jev is a hosted AI decision service.
Jev
18 points more accurate
Accuracy on phishing emails
gemini-flash-lite-latest · PhishNChips · n = 100 · 2026-10-01 · 95% range 73% to 89%
10% lower error
Calibration error: how far its confidence is from reality
gemini-flash-lite-latest · PhishNChips · n = 100 · 2026-10-01
61 ms faster
Typical response time
groq qwen3.8-27b · PhishNChips · n = 100 · 2026-10-01
Your code needs one clear answer and a confidence number.
Which team should handle this ticket?
Without Curva
It’s probably billing, but it could be technical. I’m fairly sure, maybe 0.95.
With Curva
{"choice": "billing", "confidence": 0.97}
Your code can act on this.
Options in one order, then reversed, and averaged. The order of options can no longer sway the answer.
“I was charged twice, please refund me.” Which team?
Read from the model itself, so you see how sure it is of each one.
Tell Curva the right answer. After 30, it tunes its confidence, and keeps the change only if it helps.
30/ 30 labels
That is calibration: confidence that matches how often it is right. Curva is named after the curve that checks it.
0.149
Calibration error: how far confidence is from reality, lower is better. Before tuning: 0.404.
Gemini flash-lite, AITA, n = 100, tested on rows it did not learn from, 2026-10-01. Curve illustrative.
Your app calls Curva on your own server. Curva asks your model, reads a probability for every option, and sends back one of your labels.
Rules catch the obvious ones at no cost; unsure tickets go to a person.
The Route node gives one output per option, plus Needs review.
Typed answers and an abstain flag instead of a free-text judgement.
Pass, revise or block an answer before a user sees it.
A probability of phishing, a risk level and a sender-mismatch flag in one call.
Curva accepts the criteria request shape, so a client needs a new base URL and key.
One support ticket and one question: which team should handle it? Here is what each one hands back to your code.
ticketI was charged twice for order A-104. Please refund the duplicate!
PromptWhich team should handle this ticket? Reply with one of billing, technical, sales.
This looks like a billing issue, fairly sure.
{"decision_id": "dec_…", "answers": {
"department": {"choice": "billing", "probabilities": {"billing": 0.97, "technical": 0.02, "sales": 0.01, "none_of_these": 0.0},
"confidence": 0.97, "abstain": false, "calibrated": false},
"refund": {"noul": 0.93, "calibrated": false}}}billing0.97technical0.02sales0.01none_of_these0.0choice is always one of your labels, checked before it reaches you.none_of_these for “none of them fit”.abstain turns true below the confidence you set (0.8 here), so a person can take it.refund is a second question, answered in the same call: P(yes) is 0.93.Nuance. The ticket and the numbers are the docs’ example; real numbers depend on the model you choose. A probability matches how often Curva is right only once that exact question is calibrated: after 30 feedback labels, and only if the calibrator beats the raw numbers on held-out labels. Until then you get the model’s own probabilities, debiased, which is what calibrated: false says. Debiasing asks in two option orders, so it costs two model calls unless you set debias: "auto".
Set how sure Curva must be (min_confidence). Below that, it says it isn't sure (abstain: true) and hands the case to a person.
Pick one to see what Curva sends back.
Send every ticket to the team that should handle it.
Flag suspicious emails, with how sure Curva is.
A policy call on every post, text and images together.
Check an AI answer before your user sees it.
Pull values out of invoices and receipts, each with its own confidence.
Linked questions about one alert, answered in one request.
“Charged twice, refund please”
Which team should handle this ticket?
billing0.97
technical0.02
sales0.01
none_of_these0.00
refundP(yes)0.93
Under 0.8 Curva abstains. A person picks the team, and that answer calibrates the question.
Its link opens a webmail login page on port :2096, on a domain that has nothing to do with the sender.
Is this email phishing?
phishingP(phishing)
tacticseach tactic found, with a probability
riskSafe · Low · Medium · High
sender_mismatchP(yes)
With Gemini flash-lite it got 80.8% right on a public phishing set (n = 100). Smaller models did worse, so test yours first.
A caption and an image, from an account 3 days old.
Does the image break the content policy?
Not sure enough? A moderator decides.
oknothing wrong
nudity
violence
spamads or scams
Up to 8 images per decision. The audit log keeps a hash of each image, never the image.
request the user’s messageanswer your model’s draftcontext the source documents
Pass, revise or block this answer?
Unsure? The draft waits for a person.
verdictpass · revise · block
answers_questionNot at all · Partly · Mostly · Fully
groundedP(every claim is supported)
unsafeP(yes)
leaks_dataP(yes)
Ask two or three models at once. Where they disagree, confidence drops: your cue to look closer.
ACME Inc. / Invoice 2291 / 3 x Widget @ 4.10 / Shipping 2.00 / Total due 14.30 EUR
What are the vendor, invoice number, total and PO number?
vendorACME Inc.
invoice_no2291
total14.30.94
po_numbernull
The invoice has no PO number, so Curva returns null instead of inventing one.
stage 0 which system is failingstage 1 how severe, and did a deploy cause itstage 2 roll back or not
Should the last deployment be rolled back now?
systemdatabase · api · network
severityminor · major · critical
deployment_relatedyes or no
rollback0.78
Each question sees the earlier answers and how sure they were. The decision is logged once.
Real decisions come in steps, and each answer depends on the one before. Curva runs the whole chain in one request, stage by stage. Each question sees the earlier answers and how sure they were.
systemWhich system is failing?
database · api · networkseverityHow severe is the incident?
minor · major · criticaldeployment_relatedDid a recent deployment cause it?
yes or norollbackShould the last deployment be rolled back now?
yes or no, withthinknoul0.78stage2from curva import Curva, Choice, Noul, Score
d = Curva().decide(alert, {
"system": Choice("Which system is failing?", ["database", "api", "network"]),
"severity": Score("How severe is the incident?", ["minor", "major", "critical"]).depends("system"),
"deployment_related": Noul("Did a recent deployment cause it?").depends("system"),
"rollback": Noul("Should the last deployment be rolled back now?", think=True)
.depends("severity", "deployment_related"),
})The rollback question gets think=True: the model reasons in its own call there, while the other questions answer fast. The decision is logged once.
ruleswhen · @keycascadecouncilText · Number · IntegerNuance. A request runs at most 8 stages, and latency and cost add up across them. think is slower and costs the reasoning tokens, so keep it for the hard question. A council or a cascade means more model calls. And like the models it uses, Curva is bad at counting, arithmetic and comparing dates: compute those in code and put the result in the state.
Curva never reads a label out of a sentence. It reads the probability of each label you declared, or asks for JSON limited to them, and checks every answer before it reaches you.
The answer is always one of these keys. When none of your options fit, the model can say so with none_of_these instead of being forced into a wrong pick.
In TypeScript, the answer’s type is built from your options, so a typo is a compile error:
d.answers.team.choice; // "billing" | "technical" | "none_of_these", typed from the optionsinvoiceACME Inc. / Invoice 2291 / 3 x Widget @ 4.10 / Shipping 2.00 / Total due 14.30 EUR
totalNumber, min=0
14.3confidence 0.94
A number, inside its bounds.
po_numberText, nullable
None
Not in the document, so it says so instead of making one up.
any valuewrong type or out of range
nullconfidence 0
Rejected. Never a silent bad value.
Text, Number and Integer questions each come back as a value with a confidence, so a cut-off works on extraction too.
Nuance. A valid type is not a right answer: a well-formed label can still be wrong, which is what the probabilities, abstain and calibration are for. none_of_these is on by default and you can turn it off; the LLM output QA recipe’s verdict has no escape option. Text, Number and Integer questions are answered in verbal mode. Keep arithmetic in code: ask for the numbers printed on the document, then add them up yourself.
How often each one picks the right answer, on six public test sets. We call it a win only when even the low end of Curva's range beats Jev. Otherwise it matches.
Jev (published)PhishNChipsgemini-flash-lite-latest · n = 100
80.8%Jev 62.6%win
PhishNChipsqwen3.8-27b on Groq · n = 100
72.8%Jev 62.6%win
HellaSwagqwen3.8-27b on Groq · n = 100
86.0%Jev 86.1%matches
OpenBookQAgemini-flash-lite-latest · n = 100
92.2%Jev 94.2%matches
BoolQgemini-flash-lite-latest · n = 127
88.9%Jev 89.7%matches
BANKING77gemini-flash-lite-latest · n = 120
79.8%Jev 75.3%matches
| Asking an LLM directly | Curva | |
|---|---|---|
| Output | A sentence you parse | One of your labels, validated |
| Confidence | A self-reported guess | A probability for every option. Once you send corrections, it matches how often it is right |
| Stability | Can change when options are reordered | Asked in both orders and averaged |
| When unsure | It guesses | Says it is not sure (abstain: true), so a person can decide |
| Audit | Nothing to check | Every decision logged |
| Runs on | Their servers | Yours, with any model |
Jev is still ahead on some runs. Curva vs Jev, every number ↗
Jev and TypeSafe are trademarks of their respective owners. Tarkova is not affiliated with them. Jev numbers are their published figures.
We ran eight public test sets through Curva with six models from four providers. Each line is one model. The pink diamonds are Jev's published scores.
| Model | PhishNChips | BANKING77 | OpenBookQA | CommonsenseQA | HellaSwag | AITA | BoolQ | Yelp |
|---|---|---|---|---|---|---|---|---|
| gemini-flash-lite-latest | 80.8%73% to 89%n = 100 | 79.8%73% to 87%n = 120 | 92.2%87% to 97%n = 100 | 82.0%74% to 90%n = 100 | 78.6%70% to 87%n = 98 | 67.2%58% to 76%n = 100 | 88.9%83% to 94%n = 127 | 54.0%44% to 64%n = 100 |
| qwen3.8-27b on Groq | 72.8%64% to 82%n = 100 | 74.7%64% to 85%n = 69 | 85.0%78% to 92%n = 100 | 84.0%77% to 91%n = 100 | 86.0%79% to 93%n = 100 | 55.7%46% to 65%n = 100 | not run | not run |
| Claude Haiku 4.5 (early) | 75.0%56% to 94%n = 20 | 65.0%44% to 86%n = 20 | 94.5%84% to 100%n = 20 | 75.0%56% to 94%n = 20 | 85.0%69% to 100%n = 20 | 61.2%40% to 83%n = 20 | 90.0%77% to 100%n = 20 | 50.0%28% to 72%n = 20 |
| gpt-4.1-mini (early) | 70.0%50% to 90%n = 20 | 75.0%56% to 94%n = 20 | 91.7%80% to 100%n = 20 | 80.0%62% to 98%n = 20 | 80.0%62% to 98%n = 20 | 41.0%19% to 63%n = 20 | 90.0%77% to 100%n = 20 | 50.0%28% to 72%n = 20 |
| gpt-4o-mini (early) | 50.0%28% to 72%n = 20 | 60.0%39% to 81%n = 20 | 86.4%71% to 100%n = 20 | 90.0%77% to 100%n = 20 | 55.0%33% to 77%n = 20 | 40.9%19% to 62%n = 20 | 95.0%85% to 100%n = 20 | 55.0%33% to 77%n = 20 |
| gpt-4.1-nano (early) | 55.0%33% to 77%n = 20 | 50.0%28% to 72%n = 20 | 77.0%59% to 95%n = 20 | 75.0%56% to 94%n = 20 | 60.0%39% to 81%n = 20 | 40.9%19% to 62%n = 20 | 75.0%56% to 94%n = 20 | 50.0%28% to 72%n = 20 |
| Jev (published) | 62.6% | 75.3% | 94.2% | 88.1% | 86.1% | 75.4% | 89.7% | none |
qwen3.8-27b on Groq is faster and more accurate than Jev on this set: 72.8% in a median 178 ms, against 62.6% in 239 ms (n = 100). Gemini is the most accurate and the slowest. Every other model is slower than Jev.
Cost per 1,000 decisions is what each provider charged across 160 decisions per paid model, 20 on each of the eight sets (2026-09-30). Gemini and Groq ran on their free tiers, within the daily limits. Two more free models ran on PhishNChips only (gpt-oss-20b, n = 36; Nemotron 3 Super, n = 20) and are not charted. Model and company names and logos belong to their owners.
Use the AI API key you already have: OpenAI, Anthropic, Gemini, Groq, OpenRouter, or a local model. Add a few lines of code, and your AI makes quick decisions you can trust. Curva itself is free.
pip install curva-ai
export OPENROUTER_API_KEY=sk-or-v1-... # or any one provider keyimport curva
d = curva.decide("I was charged twice, please refund me",
{"team": ["billing", "technical"], "refund": "Asks for a refund?", "total": float})
print(d.team, d.refund, d.total) # e.g. billing True NoneIt also works from TypeScript, n8n, MCP and Docker. See the docs ↗
© 2026 Tarkova Private Limited. Curva and the Curva logo are trademarks of Tarkova Private Limited; all rights reserved. Curva is free to use under the Curva Free License; its source code is not published.