Test anything
That was one model change.Platine tests 16 types of changes.
Prompts, models, tools, retrieval, memory, routing and more. Platine guides you through what to test, which business requirements to protect, and what evidence you need before shipping, with a signed record of the decision.
The process
One loop, from the first failureto the business result.
Seven steps, one record. Each step builds on the last, and what you learn in production starts the loop again.
StepWhat happensYou get
01Connect
Bring in your agent’s code, prompts, policies, traces and existing tests, along with the business outcomes you care about. Import from the tools you already use:
OpenAI Evals
LangSmith
Langfuse
Arize Phoenix
Braintrust
OpenTelemetry
Promptfoo
InspectCSV and JSONL
OpenAI EvalsLangSmithLangfuseArize PhoenixBraintrustOpenTelemetryPromptfooInspectCSV and JSONLA map of what’s known, what’s inferred and what’s missing
02Discover
Find failures in real tasks and workflows, including rare and costly ones. Platine helps you turn them into requirements, protected groups and acceptable limits that your team confirms.
Ranked failures and clear, agreed requirements
03Evaluate
Build evals and calibrate graders against your team’s judgments. Separate the cases used to improve your agent from fresh cases reserved for the release decision.
Evals grounded in your business, calibrated graders and a protected fresh set
04Propose
Develop alternatives across all 16 types of changes, with the right comparison and risks identified for each. Ideas can come from you, your team or Platine.
Prompts and instructionsModel selectionInference settingsContext constructionRetrieval and RAGTools and tool useWorkflows and agent logicMemory and stateRouting and escalationRetries and fallbacksCaching and servingRequirements and evalsComparison and decisionImplementation and releaseProduction feedbackIntegrations
Candidates, each labelled with who proposed it and what it changes
05Compare
Explore alternatives on practice cases, then use those results to plan how much fresh evidence you need. Lock the candidate and test plan before the final comparison. See the effect, uncertainty, cost and latency against your requirements and protected groups.
Ship, Don’t ship, or Not enough evidence, with reasons and next steps
06Release
Turn an approved candidate into a reviewed release, with its signed evidence record attached. Plan a staged rollout with explicit stop rules, under your team’s deployment controls.
GitHub pull requestsYour CI
GitHub pull requestsYour CIA reviewable pull request, a rollout plan and a signed record others can verify and recompute from scratch
07Monitor
Track whether the expected benefits appear in real business outcomes. When results deteriorate or a model, grader or policy changes, flag the decision for reassessment and feed new failures into the next improvement.
Live outcomes, emerging failures and decisions flagged for reassessment
What Monitor finds becomes the next thing Discover works on.
Why Platine
Your eval tool shows the scores.That’s not enough.
Already have evals? Keep them: Platine adds the decision they leave to you. Starting from scratch? Platine builds them from your agent’s real tasks and activity.
Eval and observability tools
Platine
What you get
Eval and observability toolsScores, traces and configurable checks
PlatineA statistically grounded release decision, with reasons and next steps
Requirements
Eval and observability toolsYou configure metrics, groups and thresholds
PlatineGuided requirements, with an acceptable limit for each protected group
Enough evidence?
Eval and observability toolsYou design the sampling and stopping procedure
PlatineStatistical guidance on what to test, how much evidence to collect, and when to decide
What you keep
Eval and observability toolsDashboards, reports and exported results
PlatineA signed record others can verify and use to recompute the verdict from scratch
Safeguards
An average can hidethe failures that matter.
Platine makes the call under a rule you lock in advance: every critical group within its own limit, confirmed on fresh cases, with the risk of a wrong ship stated up front.
2.5%False-ship risk per decision, under your stated assumptions
0Fresh test results used to tune the candidate
1Decision rule locked before fresh testing
Verify
Don’t take our word for it.Verify.
Every release can carry a signed evidence record. Anyone you share it with can check its integrity and recompute the decision from scratch, without relying on Platine’s dashboard.
- Who it’s for
- Leads approving a release, customers assessing a vendor, security and audit teams.
- Who sees it
- You choose who to share it with.
- Signature
- Check who signed the record and that its contents haven’t changed since signing.
- Recompute
- Recompute the statistical results and verdict from the recorded inputs, using the plan locked before fresh testing.
- How
- A standalone verifier that works without Platine.
Sample record · illustrative animation
$ platine verify decision-CH-14r.platine
Press Preview verification to see the checks the verifier runs.
Your evidence moves through procurement, customer reviews and internal approvals, and stays usable if you ever leave Platine.
A valid signature doesn’t make weak evidence strong. The record exposes the cases, graders, limits and assumptions, so recipients can judge the evidence as well as verify the calculation.
Pricing
Pay for the process,not for your data.
Every plan includes the same statistical methods, guided decisions and signed records. Test your own agents or compare vendors. Bring your own model keys and pay your providers directly, with no markup from Platine.
Swipe left or right to compare plans
| Choose your scale.The same statistical methods on every plan. Scroll to compare all features. | Free$0Build your first evidence-backed decision. | Pro$49per monthYour first agent in production. | Team$199per monthFor startupsSeveral agents. Shared release decisions. | Business$699per monthRelease controls across a growing organization. | EnterpriseCustomannual contractYour deployment, governance and support requirements. |
|---|---|---|---|---|---|
| Usage and capacity | |||||
| Included agents | 1 in development | 1 in production | 5 in production | 20 in production | Custom |
| Additional production agents | − | − | $39 / agent / month | $29 / agent / month | Custom |
| Users | 1 | Up to 3 | Unlimited | Unlimited | Unlimited |
| Hosted execution hours / month | 2 | 50 | 250 | 1,000 | Custom |
| Monitored production tasks / month | − | 20,000 | 100,000 | 500,000 | Custom |
| Stored evidence | 1 GB | 50 GB | 250 GB | 1 TB | Custom |
| Hosted retention | 30 days | 1 year | 2 years | 3 years | Custom |
| Evaluation and evidence | |||||
| Build evals, explore fixes and confirm on fresh cases | ✓ | ✓ | ✓ | ✓ | ✓ |
| Evaluate your agents or compare vendors | ✓ | ✓ | ✓ | ✓ | ✓ |
| Guided decisions and signed records | ✓ | ✓ | ✓ | ✓ | ✓ |
| Download records and evidence for recomputation | ✓ | ✓ | ✓ | ✓ | ✓ |
| Verify existing records in public pull requests, without another evaluation | ✓ | ✓ | ✓ | ✓ | ✓ |
| Production monitoring | |||||
| Monitor outcomes and recorded dependencies | − | ✓ | ✓ | ✓ | ✓ |
| Flag decisions that need reassessment | − | ✓ | ✓ | ✓ | ✓ |
| Suggestions for cheaper models to evaluate | − | ✓ | ✓ | ✓ | ✓ |
| Slack alerts | − | − | ✓ | ✓ | ✓ |
| Collaboration and release controls | |||||
| Require approval from the people you choose | − | − | ✓ | ✓ | ✓ |
| Share hosted records with named customers and revoke access | − | − | ✓ | ✓ | ✓ |
| Optional public page of verified releases | − | − | ✓ | ✓ | ✓ |
| Company-wide requirements, risk policies and approval controls | − | − | − | − | ✓ |
| Security, deployment and support | |||||
| SAML single sign-on | − | − | − | ✓ | ✓ |
| Exportable audit logs | − | − | − | ✓ | ✓ |
| Evidence packages for EU AI Act, ISO 42001 and model-risk reviews | − | − | − | − | ✓ |
| User provisioning and access roles | − | − | − | − | ✓ |
| Private cloud or on-premises deployment | − | − | − | − | ✓ |
| Support SLA | − | − | − | Business hours | 24/7 · dedicated support |
| Add capacity when you need it | − | No automatic overage charges | No automatic overage charges | No automatic overage charges | No automatic overage charges |
Choosing an AI vendor? Ask for its Platine records and verify them for free.
Know what to ship.Know why it holds up.
Coming soon
Talk to the founder →