01
Risk classification
Classify model inputs and outputs against policy domains such as instruction abuse, restricted content, unsafe automation, and operational risk.
ModelShield
A classifier for the AI safety layer.
ModelShield is an advanced guardrail classifier for organizations building with LLMs, agents, and automated workflows. It is designed to classify risky behavior, route uncertain cases, and leave a reviewable trail for teams responsible for deployment.
- Policy-aware classifications for allow, block, warn, and review workflows.
- Decision metadata that helps teams audit behavior after release.
- Designed for practical deployment boundaries rather than unverifiable safety theater.
In development
Risk routing preview
The product direction is classifier-first: normalize policy categories, score risk, decide a route, and preserve the evidence needed for humans to review edge cases.
Instruction hierarchy conflictEscalate
Restricted content requestBlock
Allowed business queryAllow
Route decisions are intended to be inspectable, not mystical.
Capabilities
Built around classification, routing, and evidence.
ModelShield is in development. The site intentionally describes the product direction without claiming production availability or unsupported benchmark results.
02
Guardrail routing
Route decisions into allow, block, warn, transform, or human-review paths so application teams can choose the right level of friction.
03
Review metadata
Attach policy class, confidence, reason codes, and trace metadata so reviewers can understand why a request was handled a certain way.
04
Deployment controls
Support configurable thresholds, environment-specific policies, and conservative fallbacks when the classifier is uncertain.
Why it matters
AI applications need control surfaces, not just prompts.
Prompt-only guardrails are brittle
Production teams need classifier-backed routing that sits beside the model and can be audited independently of any one prompt template.
Human review needs context
Escalation queues are only useful when each decision carries policy context, confidence, and the reason it was sent for review.
Safety claims need calibration
ModelShield is being shaped around measured behavior, known failure modes, and explicit limits rather than broad promises.
Also from Dwarrow
Console and applied safety engines.
Core infrastructure being built around ModelShield.
Access
Dwarrow Console
Unified login, app launcher, and admin controls for invited roles.
Console loginResearch and engineering
Applied safety infrastructure
Research-grade AI systems that emphasize constraints, observability, and long-term maintainability.
Why Dwarrow