Frequently asked questions

Short answers about Litmus, Sentinel and how your data is handled. Can't find what you need? Read the guides, or ask our team.

About Litmus and Sentinel

What are Litmus and Sentinel?

Two tools from GovTech's AssuranceAI team for teams building or running AI systems. Litmus tests your AI system against criteria you set, before you deploy and after changes. Sentinel checks prompts and responses with guardrails while your system is running.

Do we need both Litmus and Sentinel?

No. Each works on its own: use Litmus to test a system before launch, or Sentinel on a system that is already live. You can also use both: test with Litmus before launch, then check live traffic with Sentinel. Work out which you need.

Who are Litmus and Sentinel for?

Teams in the Singapore public service that are building or running an AI system.

To be confirmed: which agencies can use Litmus and Sentinel after the 19 October 2026 launch.

Is there a cost to use Litmus and Sentinel?

There's currently no charge for Litmus or Sentinel. Plans based on how much each team uses them will follow.

Testing with Litmus still costs your team something: Litmus sends each test case to your system, so your system's usual running costs, such as its AI model usage, apply. A smoke run sends a small sample; a full run sends every test case. How runs work.

Can we start using Litmus and Sentinel straight away?

Litmus, yes: sign in with TechPass and add your system. Sentinel, not straight away: fill in the onboarding form⁠ (opens in a new tab), and our team will register your team and give you an API key. See the steps to get started.

Do we need technical help to use Litmus or Sentinel?

Yes, for setup. For Litmus, someone who knows your system's API needs to connect it; after that, your team works in the Litmus web app, which has guided tours of each step. For Sentinel, an engineer needs to add the Sentinel calls to your application and decide what it does with each result.

Litmus

What kinds of AI systems can Litmus test?

Text-based systems, such as chatbots, question-and-answer and RAG systems, and tools that summarise, classify or extract, one question and answer at a time. Litmus calls your system through its HTTPS API endpoint. It doesn't yet test multi-turn conversations, or image or audio input. Check what Litmus needs from your system.

Do we need evaluation expertise to use Litmus?

It helps to have someone who understands how your system should behave. Litmus has help built in: a short primer, guided tours of each step, and a fictional example system to practise on. It can suggest risks from your description, criteria from those risks, and test cases. Your team checks those suggestions, adds its own criteria alongside the required ones, and reviews the results.

Can we use our own test cases in Litmus?

Yes. Import them from a CSV file or write them by hand, alongside the test cases Litmus generates. Datasets are versioned.

Can Litmus grade results we ran ourselves?

Yes. If you ran the test cases in your own environment, upload your system's answers as a CSV, JSON or Excel file for Litmus to grade, and the run is labelled as self-reported. You still connect your system's endpoint when you first add it.

Who grades the answers in Litmus, and can we check them?

AI judges, which are large language models, grade each answer against your criteria and explain their reasoning. You can choose the judge model, and spot-check the judges' verdicts yourself to see how far to rely on them. See which AI models Litmus uses.

What happens when a Litmus test fails?

Litmus lists the issues each run finds, graded Critical or Major, and shows the AI judge's reasoning for each answer. Fix what you need to, then run the tests again and compare the scores.

Can we compare Litmus runs?

Yes. Each run is kept, so you can re-run the tests after a change and compare pass rates with any earlier run of the same system. Litmus doesn't record which version of your system a run tested, so keep a note of that yourself.

Does Litmus approve our system?

No. Litmus doesn't decide whether your system can launch. Litmus provides evidence to support deployment decisions. It does not replace the judgement, governance processes or approvals required by the relevant agency.

Can Litmus test a system on an intranet?

Not at the moment. Your system's endpoint must use HTTPS and be reachable at a public internet address. Litmus can't currently connect to systems that are only reachable on an intranet or private network.

Does Litmus have an API or CI/CD integration?

Not yet. The new Litmus is used through its web app. It doesn't yet offer API keys or CI/CD integration.

Does the new Litmus do red-teaming?

Not yet. The new Litmus doesn't run automated red-teaming attacks; Litmus classic still does. You can include a built-in safety benchmark of probe prompts in any run, covering hateful, sexual, self-harm, violent, toxic and misconduct content.

Sentinel

What does Sentinel check?

Harmful content in a Singapore context (LionGuard 2, in three versions), prompt attacks, off-topic requests, system prompt leaks and refusals. Guardrails that run on Amazon Bedrock also check for hate, insults, sexual content, violence, misconduct, prompt attacks and personal data. You choose which guardrails to run, and save them in a policy: read the guardrail reference.

What kinds of applications can use Sentinel?

Any application that can send text to the Sentinel API over HTTPS. Call Sentinel from your server, not from code in a user's browser, so your API key stays secret. Sentinel checks text only, not images, audio or files. If your application runs on GCC without internet access, contact our team about connecting through a private endpoint.

To be confirmed: whether Sentinel accepts requests only from Singapore IP addresses.

How do we call Sentinel?

First create a policy: the guardrails to run, and the scores at which each one warns or fails. Then send a POST request to /api/v1/check with the text to check and your policy's name, with your API key in the x-api-key header. Sentinel returns passed, warning or failed for each guardrail and for the check as a whole. The Sentinel getting-started guide shows how, with examples in Python, JavaScript and cURL.

What does a Sentinel score mean?

Each guardrail returns a score from 0 to 1, where a higher score means the text is more likely to fail that check. Some guardrails only return 0 or 1. Your policy sets the score at which each guardrail warns or fails.

Does Sentinel block messages?

No. Sentinel tells your application what it found; it doesn't block or change anything itself. Your application decides whether to let the text through, block it or replace it with a safe message.

Will Sentinel catch every problem?

No. Guardrails can miss harmful text and can flag safe text, so try them on your own examples in staging first. Sentinel reduces specific runtime risks. It should be used alongside appropriate system design, access controls, monitoring and human oversight.

Is there a Sentinel test environment?

Yes. Send requests to https://sentinel.stg.aiguardian.gov.sg/api/v1/check while you build, and to https://sentinel.aiguardian.gov.sg/api/v1/check in production. Staging and production use different API keys and keep separate policies, so create your policy in each. Staging requests are recorded in the same way as production ones.

How do we get a Sentinel API key?

Our team gives you API keys for staging and production after your team is onboarded; you can't create keys yourself. Start by filling in the onboarding form⁠ (opens in a new tab).

How fast is Sentinel?

It depends on which guardrails you run. Each response reports how long each guardrail took, and the total, in time_taken. Try your own traffic in staging to see what to expect.

Security and data

Where do Litmus and Sentinel run?

On Amazon Web Services (AWS) in the Singapore region, in an environment GovTech manages. Some features send text to AI model services, which may process it outside Singapore. See which AI services Litmus and Sentinel use.

To be confirmed: where each outside AI service processes text (including Azure OpenAI, Google Gemini, the OpenAI and Anthropic models Litmus's judges can use, the model behind Sentinel's daily quality check, and the PlatformAI and GovText gateways), and whether the providers keep it or use it to train their models.

What does Litmus keep, and where does it send our data?

Litmus keeps what you enter about your system, the text of documents you upload, your criteria, personas and test cases, and the full response your system returns to each test. It also keeps the AI judges' grades and reasoning, Sentinel guardrail scores, your team's spot-checks, and your name, email address and sign-in times. API keys and tokens you give Litmus for your system are stored encrypted.

To generate tests and grade answers, Litmus sends your system description, document text, test inputs and your system's answers to large language models. The AI judges use Google Gemini by default; OpenAI and Anthropic models can also be chosen. During runs, Litmus also checks test inputs and answers with Sentinel, which stores and processes them like any other check. Read how we handle your data.

What does Sentinel keep, and where does it send our data?

Sentinel keeps a record of every check, in staging and in production: the text, the policy or guardrails that ran and their settings, any context you send to /validate (such as your system prompt or earlier messages), any tags or metadata, and the results, including any personal data a guardrail finds. Your team can see its production checks on the Sentinel dashboard.

Depending on the guardrails you choose, Sentinel sends the text, and any context you include, to models GovTech hosts or to commercial AI services, some of which may process it outside Singapore. LionGuard 2 and system prompt leakage use Microsoft Azure OpenAI. LionGuard 2.1 uses a Google Gemini embedding model through GovText's model gateway, and prompt attack and refusal use a Google Gemini model as a judge. The AWS guardrails use Amazon Bedrock Guardrails in the AWS Singapore region. LionGuard 2 Lite and off-topic use models GovTech hosts.

Each day, an automated job sends about one in ten of the previous day's LionGuard 2 checks, from every team (including checks Litmus makes during test runs), to an AI model, OpenAI's GPT-5.1 by default, to measure how accurate the guardrail is. Read how we handle your data.

Is our data used to train AI models?

To be confirmed: whether data you submit may be used to improve GovTech's own guardrail models, such as LionGuard.

To be confirmed: where each outside AI service processes text (including Azure OpenAI, Google Gemini, the OpenAI and Anthropic models Litmus's judges can use, the model behind Sentinel's daily quality check, and the PlatformAI and GovText gateways), and whether the providers keep it or use it to train their models.

See which AI models Litmus and Sentinel send text to.

How long is our data kept?

To be confirmed: how long Litmus and Sentinel keep your data, and how to ask for it to be deleted.

In Litmus, you can delete documents you've uploaded, and test cases that aren't yet part of a saved dataset version. Test cases in a saved version (every run uses one) stay, with your system's answers to them. You can't yet delete a whole system yourself.

What data must we not submit?

Litmus and Sentinel are not yet cleared for CCE systems or data. Check with us before onboarding such a system.

To be confirmed: what CCE covers, which other data classifications each product supports, and what you must not submit.

Have Litmus and Sentinel been security tested?

To be confirmed: the status of independent security testing (VAPT) for the new Litmus and for Sentinel.

If you find a security issue, report it through GovTech's vulnerability disclosure programme⁠ (opens in a new tab).

Litmus classic

For teams already using Litmus classic.

What is Litmus classic?

The first version of Litmus. It sends test prompts to your application and checks the responses with Sentinel guardrails. Teams already using it sign in to Litmus classic⁠ (opens in a new tab) with TechPass.

How long will Litmus classic be available?

Until March 2027. To be confirmed: whether teams that don't use Litmus classic yet can still start on it.

Where are the Litmus classic guides?

In our documentation: the Litmus classic overview, getting-started guide and troubleshooting guide. These guides describe Litmus classic and may point to an older sign-up form. To ask for access, use the onboarding form⁠ (opens in a new tab).

How do we move from Litmus classic to the new Litmus?

Tell us through the onboarding form⁠ (opens in a new tab). Check first that the new Litmus does what you need: it doesn't yet run automated red-teaming or offer CI/CD integration, which Litmus classic does.

Ready to start?

Sign in to Litmus with TechPass and add your system. For Sentinel, fill in the onboarding form and our team will contact you about next steps.

Start with Litmus⁠ (opens in a new tab)Start with Sentinel⁠ (opens in a new tab)