EyeTrustAIPrediction sets
A set of plausible labels, calibrated to a marginal coverage target.
Open full walkthroughClassification · ForgeSight
Synthetic examples · Not customer deployments
Platform · For model builders
Add statistical guarantees
to your model’s outputs.
Bring your model and evidence. Calibrate prediction sets, intervals or regions, then inspect what the guarantee supports.
Platform in development · Interactive walkthroughs available
EyeTrustAI / Model workspacePlatform preview · Classification example
From a label to a prediction set.
Existing model: visual inspection · Version 1
CHOSEN OBJECTIVE95% marginal coverage
CALIBRATION RECORDS199
EVALUATION RECORDS400 · separate
One evaluation example
Original answer
ScratchCalibrated set
Scratch DentReference: Dent · included in the set
Whole evaluation
Observed coverage92.5%
Mean labels returned1.16
Reference misses30 / 400
The finite evaluation falls below the 95% target. The target is a population-level statistical statement, not certainty for each result.
Model version + evidence + method + target → versioned output artifactSynthetic data · Real calculations in the linked walkthrough
A preview of one platform experience. Explore the complete classification journey below.
How it works
One lifecycle. Different model outputs.
01 / Connect
Start with your model.
Map scores, predictions and references. Keep the original output available.
02 / Prepare
Separate the evidence.
Keep model development, calibration and evaluation distinct. Define the population and error.
03 / Calibrate
Choose the guarantee.
Set coverage or expected-loss targets. Calculate the output rule from reviewed evidence.
04 / Inspect & maintain
See what the model can support.
Inspect errors and output size. Export a versioned artifact and reassess when conditions change.
Explore the platform
Follow a complete builder journey.
Three fictional scenarios using the same lifecycle. Calculations run locally; deployment is simulated.
Prediction intervals
A range around a numerical prediction, in the units your model uses.
Open full walkthroughRegression · Voltara
Synthetic examples · Not customer deployments
Controlled regions
A spatial region calibrated to a defined omission loss.
Open full walkthroughSegmentation · ContourWorks
Synthetic examples · Not customer deployments
The guarantee has a defined scope
Choose the right
statistical statement.
Coverage, average omission and error among selected outputs answer different questions. The method, population and reference define what a result means.
See the validation studiesWhat does a statistical guarantee cover?
A defined risk or coverage property for a specified population, under the method’s assumptions. Coverage concerns whether the reference falls inside a set or interval. Expected-loss control concerns an average loss, such as omitted anatomy. Error among selected outputs requires a method that controls that particular subset. None means every individual output is correct.
What do the walkthroughs calculate?
ForgeSight and Voltara calculate split-conformal sets and intervals. ContourWorks calculates a region using conformal risk control. Each uses a disclosed synthetic dataset and shows separate evaluation results, including misses.
What changes when the model changes?
The output artifact is tied to a model version, evidence and method. The walkthroughs show version checks and new evidence rounds. They do not establish guarantees under arbitrary distribution shift.
What is available today?
Complete interactive demonstrations and public research artifacts. The shared platform is in development. Our focused solutions serve teams operating AI in their products and workflows.
Explore the solutionsWhat are you building?
Tell us about your model and the uncertainty you need to control.