AI proposes
A fast Featherless model generates bounded adversarial inputs. Every case is validated and labelled.
Featherless-nativeDeterministic release gate
Break your AIbefore users do.
Stress-test one structured AI feature across open models. Get exact failures, comparable evidence, and a reproducible SHIP / FIX / BLOCK decision.
Release gate previewSeeded proof
01Load structured task contractReady
02Run 4 cases × 3 open models12 traces
03Critical holdout boundaryFailed
Baseline verdictBLOCK
Exact failing input and assertion preserved.
One APIThree open models, identical tests
Exact evidenceInput, output, assertion, latency
Honest boundaryAI proposes; deterministic code decides
Judge-safeLive inference plus labelled fallback
Interactive workbench
One contract in. One release decision out.
1 Define2 Attack3 Decide
Unknown failures are still unknown.
Run the same contract across open models. ModelGauntlet will turn every output into exact checks, traces, and one release decision.
20 deterministic checks/model2 untouched holdouts0 LLM judge scores
Trust boundary
Probabilistic discovery. Deterministic release gate.
Generated cases expand coverage; they do not prove universal safety. Human holdouts remain untouched and own the critical release boundary.
Code decides
JSON Schema, exact paths, forbidden content, latency, and thresholds produce the verdict.
Evidence survives
Exact inputs, outputs, normalization, usage, and failures remain open for inspection.
Honest limitationModelGauntlet evaluates structured text tasks and cannot certify universal model safety.
Architecture at a glance
Small system. Hard boundary.
Task contract→Featherless models→TypeScript assertions→SHIP / FIX / BLOCK