Automate the clear cases
Let more routine work proceed under policies you define.
Capsa
Risk signals for models you control.
Capsa adds model-level and output-level risk estimates to supported models you run, giving your application more information about each prediction.
Capsa provides a proprietary risk-estimation capability within Themis’s broader decision-layer approach. For supported models, these signals inform policies governing what proceeds, what needs more evidence and what should reach an expert.
Spend review where it counts
Focus experts and additional checks on the decisions that need them.
Put evidence where it matters
Choose when another document, simulation or experiment is worth its cost.
[ 01 ] How to add it
Add risk estimation to a supported model.
Load a supported model, wrap it with Capsa, and use the returned risk estimate alongside each prediction.
01Step 1
Load your model
Start with a supported model in your existing framework.
Original model: a single prediction, ŷ.
02Step 2
Add risk estimation
Capsa wrappers apply the model modifications needed for the selected uncertainty method.
Capsa applied to the hidden layers: each prediction ŷ now comes paired with σ, Capsa’s uncertainty estimate.
03Step 3
Use the additional signal
Use the resulting risk estimates alongside predictions to inform your application’s next step.
The pair sets the next step: proceed, or verify when σ is above the threshold τ.
[ 02 ] The signal
Not every prediction deserves the same trust.
On familiar inputs, a model is usually reliable. On unfamiliar ones, σ, Capsa’s uncertainty estimate, rises — and where it crosses the review threshold τ, the prediction is checked before anyone acts on it.
Reading · hover, tap or use the arrow keys
- Input
- #13 · unfamiliar
- Prediction ŷ
- 0.45
- Uncertainty σ
- 0.142
- Threshold τ
- 0.067
- Next step
- Verify
[ 03 ] Capabilities
See where the model is unsure.
Choose a Capsa capability to preview.
Prioritize review
Identify ambiguous inputs and potentially mislabeled examples for closer inspection.
Capsa demo · prioritize review
Report uncertainty during inference
Report uncertainty with every output at runtime and automate quality assurance by filtering low certainty output or requesting human intervention.
Capsa demo · report uncertainty
Detect failures
Capsa automatically detects prediction risk. In this example, Capsa correctly highlights the passing bus, which sits outside the training set.
Note: Bright colours represent areas of high uncertainty flagged by Capsa that were missed by the initial model.
Capsa demo · detect failures
[ 04 ] Uncertainty types
Quantify multiple types of uncertainty.
Aleatoric uncertainty
Aleatoric uncertainty
What the model can’t understand from data due to ambiguity and irreducible noise like errors.
Epistemic uncertainty
Epistemic uncertainty
What the model doesn’t know due to limited knowledge or data.
Vacuitic uncertainty
Vacuitic uncertainty
What the model can’t understand due to latent feature imbalance.
[ 05 ] ML workflow
Solutions for every stage of the ML workflow.
01
Data cleaning
Optimize data labeling costs by finding incorrect or noisy labels.
02
Data curation
Identify missing data to intelligently curate datasets.
03
Model conversion
Add a risk score to a model you already run.
04
Risk-aware training
Optimize models to overcome their deficiencies.
05
Quality control
Targeted validation on high-risk scenarios.
06
Anomaly detection
Quickly detect out-of-distribution events.
[ 06 ] Related research
Related research.
Uncertainty-Aware Language Modeling for Selective Question Answering
Converting a language model with the Capsa framework so it estimates uncertainty with every prediction.
Spotting the Unfamiliar: Out-of-Distribution Detection with Uncertainty
How epistemic uncertainty flags inputs a model has never seen, so the system can escalate or abstain.
Quantifying Epistemic Uncertainty with the Laplace Approximation
Recovering epistemic uncertainty from an already-trained model, post hoc, with no retraining.
From risk signals · to operating decisions
From risk signals to operating decisions.
Compare your current workflow with a policy that determines what can proceed, what needs another check, and where to focus expert review. A policy can deny an action that is not permitted, or abstain when the evidence is insufficient.
Working with frontier APIs? Themis brings decision-policy, model and production-engineering expertise to your workflow.