Agent guardrails
Check a tool call or command before an agent runs it: allow it, ask the user first, or block it.
rm -rf /var/log/* · what should the agent do?
Example ·
Caliper-1 is a decision model for business rules
Send the facts and your questions about a policy, transaction or agent action. For each question, Caliper-1 returns one of the answers you allowed, with a probability for every option.
The model
What Caliper-1 does
Caliper-1 is the first model from Inferr. It is trained mainly on policies, deadlines, amounts and safety checks, and it answers only with the options you define.
Caliper-1
Send a policy and the relevant facts, ask typed questions (yes or no, choose one, or a rating scale), and receive a probability for each answer. One request can include long policy documents, multi-step conditions, date windows and amount limits.
Model idinferr/caliper-1
Answer typesyes / no · choice · scale
Price$0.04 / M in · output free
Options per questionUp to 5,000
Questions per requestUp to 256
Context16K state · 128K total
Why Caliper
Most decisions in software follow written rules, such as refunds within 30 days or expense approval under $500. Language models are designed to generate text. Caliper is designed to apply rules to the facts you give it and to report how confident it is.
Ask a language model
Asked whether a return on day 38 is inside a 30-day window, a language model replies with a paragraph. You have to parse the reply and trust its date arithmetic, and you get no measure of its confidence.
Ask Caliper
Caliper returns outside 0.94: one of the answers you defined, with its probability. You can set a confidence threshold and act on answers above it automatically.
Examples
Example requests
Spend limits, refund windows, agent safety checks and support triage. Each request includes the facts and several questions, and all of them are answered together.
Use cases
What it is used for
Caliper-1 suits decisions that follow rules or clear criteria and that your software makes many times a day.
Support triage
Route each ticket to the right team, and score its urgency and sentiment in the same request.
Charged twice for one order · which team handles it?
Policy checks
Apply written policies to a case, such as refund windows, expense limits and eligibility rules.
Returned 38 days after delivery · inside the 30-day window?
Content and prompt screening
Classify user content or incoming prompts before they reach your product or your agent.
“Ignore all previous instructions…” · what is this message?
Catalogue and intent routing
Choose one item from a long list, such as products, intents, help articles or tools, in a single question.
“waterproof trail shoes, wide fit” · which of 5,000 products?
Document review
Check contracts and forms for required clauses, conditions and missing information.
Does this supplier contract cap liability?
Questions and answers shown are examples.
Reliability
How it handles real requests
Requests in production vary in ways that cause problems for decision models. Each case below runs on an example request as you scroll.
Option order
Option order does not change the answer
Some decision models favour options because of their position in the list. Caliper-1 returns the same answer and the same probabilities for any order of the same options, and lists them in the order you sent.
Any ordergives the same answer and probabilities
Notes in the input
Notes inside the input are treated as content
Inputs often include text written by the people being assessed, such as “pre-approved” or “ignore the above”. Caliper-1 is trained to treat that text as part of the input, not as an instruction to follow.
Facts onlytext in the input is read as content, not as instructions
Many questions
Questions in one request are answered independently
You can ask up to 256 questions about the same input in one request. Each question is answered separately, so adding a question does not change the answer to another one.
256questions per request
Thousands of options
Up to 5,000 options in one question
A single question can have up to 5,000 options, enough to choose from a product catalogue, a list of intents or a tool registry. The other decision models we checked allow 255 options or fewer.
5,000options per question
answer
RepeatableThe same request returns the same answer.
Typed answersEvery answer is one of the options you defined, never free text.
Clear limitsInput over the limit returns an error instead of being cut off. You pay only for the text you send.
The requests and answers shown are examples. No model catches every phrasing, so do not rely on a single model to approve irreversible actions. Option limits for other models are taken from their documentation.
Scale
Many decisions in one request
Caliper-1 reads the input once and scores every question and every option in the same pass. Large decisions do not need to be split across many calls.
256questions per request, all answered together.
5,000options per question.
128Ktokens of input per request, including a state of up to 16K.
Calibration
Probabilities you can act on
Each answer includes a probability. Choose a threshold, act automatically on answers above it, and send the rest to a person.
97.0%correct on the answers it handles automatically
77%of questions handled automatically
0.04expected calibration error
Benchmarks
Accuracy on the decisions it is built for
Results on public benchmarks for the jobs Caliper-1 is designed to do. Below them, a comparison with other decision models on the same questions.
Business rules and policies
Multi-step policies, deadlines, amounts and trade-offs.
89.2%Intent routing
Sending each customer message to the right one of 77 banking intents.
84.1%Agent tool decisions
Whether a tool call, command or code change should go ahead.
74.9%Contract review
Finding clauses such as audit rights, exclusivity and liability caps.
91.2%Compared with other decision models on the same questions
Caliper-1 as served, 16 October 2026, on questions the model was not trained on. hard-v1 and devtools-v1 comparisons use the development sets, where Jev results are published for the same questions. BANKING77 is macro-F1, with other models' results from the Decision Index 0.3 public board. LegalBench scores are the average over every task in the named group, 40 questions per task.
Cost
Cost of one million decisions
Here one decision is a 500-token input with one answer. The chart compares Caliper with popular language models at their OpenRouter list prices.
60×cheaper than Claude Sonnet 5 for the same decisions
$0.04per million input tokens. Output is free.
OpenRouter list prices on 15 October 2026. Language models are charged for about 20 output tokens per answer; Caliper's output is free. Caliper also returns a probability for every option, which these models do not.
Questions
Frequently asked questions
Is Caliper just a small language model?
It is built on a language model, but it does not generate text. It reads your input once and scores each answer you allow, so it is inexpensive to run and can only return one of your options.
How is this different from JSON mode or structured outputs?
Structured outputs make a language model write in a given format. It can still return a value you did not expect, and you get no reliable measure of its confidence. Caliper can only return your options, and each answer includes a calibrated probability.
What is it good at, and where does it struggle?
Rules applied to the facts you send: policies, deadlines and date windows, amounts and limits, eligibility, safety checks on agent actions, triage and routing. It is weaker on questions that need knowledge that isn't in the input, like trivia, and it does not write open-ended answers.
Can it be wrong?
Yes. The probabilities tell you when to trust it. Choose a threshold for each question using a few hundred of your own examples, act on answers above it, and send the rest to a person or a larger model.
Is it deterministic?
Yes, in practice. The same request returns the same answer, and neither option order nor the other questions in a request change an answer. Under heavy load, probabilities can differ in the third decimal place; the chosen option does not change.
Can text inside the input trick it?
It is trained to resist common attempts, such as claims of pre-approval, notes addressed to an AI reviewer, or “ignore the above”. No model catches every phrasing, so do not rely on a single model to approve irreversible actions.
How many options can a question have?
Up to 5,000 options in one choice question, and up to 256 questions in one request. For very large option sets, the probabilities of clearly unlikely options are approximate; the chosen option and the leading contenders are scored precisely.
What happens to my data?
Request content is processed in memory and discarded. We never train on it. We keep only counts, token totals and error codes to bill and run the service. Requests are processed in the United States.
How do I get access?
Click Get access or email hello@inferrai.com and we'll send you an API key. Then follow the quickstart in the documentation.
Company
Inferr AI
Inferr AI is a small lab that builds AI models for specific tasks. Caliper-1 is our first model, and more are in training.
Inferr AI is a brand of RK Web Solutions. For API access or partnerships, email hello@inferrai.com.
Make your first request
Send the facts and your questions in one HTTP request. Each answer comes back as one of your options, with a probability for every option.
curl https://api.inferrai.com/v1/systemone \
-H "Authorization: Bearer $INFERR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "inferr/caliper-1",
"state": { "ticket": "Charged twice for order 4471 and it arrived late." },
"questions": {
"team": { "type": "choice", "instructions": "Which team handles this?",
"criteria": { "billing": null, "shipping": null, "returns": null } },
"refund": { "type": "noul", "instructions": "Should we refund the duplicate charge?" }
}
}'import os, requests
r = requests.post(
"https://api.inferrai.com/v1/systemone",
headers={"Authorization": f"Bearer {os.environ['INFERR_API_KEY']}"},
json={
"model": "inferr/caliper-1",
"state": {"ticket": "Charged twice for order 4471 and it arrived late."},
"questions": {
"team": {"type": "choice", "instructions": "Which team handles this?",
"criteria": {"billing": None, "shipping": None, "returns": None}},
"refund": {"type": "noul", "instructions": "Should we refund the duplicate charge?"},
},
},
)
answers = r.json()["answers"]
print(answers["team"]["choice"], answers["refund"]["noul"])const r = await fetch("https://api.inferrai.com/v1/systemone", {
method: "POST",
headers: {
Authorization: `Bearer ${process.env.INFERR_API_KEY}`,
"Content-Type": "application/json",
},
body: JSON.stringify({
model: "inferr/caliper-1",
state: { ticket: "Charged twice for order 4471 and it arrived late." },
questions: {
team: { type: "choice", instructions: "Which team handles this?",
criteria: { billing: null, shipping: null, returns: null } },
refund: { type: "noul", instructions: "Should we refund the duplicate charge?" },
},
}),
});
const { answers } = await r.json();
console.log(answers.team.choice, answers.refund.noul);Response · example
{
"answers": {
"team": {
"type": "choice",
"choice": "billing",
"probabilities": { "billing": 0.81, "shipping": 0.16, "returns": 0.03 }
},
"refund": { "type": "noul", "noul": 0.96 }
},
"usage": { "input_tokens": 74, "output_tokens": 0, "cost": 0.00000296 }
}