Skip to content

Shadow Harness Architecture

Components

ComponentResponsibilityBoundary
Capture adapterConvert live traces into a stable envelopeRedacts before persistence
Case registryStore immutable task and dataset versionsContent-addressed identity
Replay plannerSelect cases and candidate configurationsNo execution privileges
ExecutorRun bounded candidate workloadsIsolated credentials and quotas
GraderProduce versioned outcomes and rationalesCannot mutate task records
ComparatorPair baseline and candidate resultsOperates on immutable runs
Review surfaceSlice results and inspect examplesRead-only release evidence

Data contracts

A case describes the task and expected evaluation method. A configuration fingerprints every behavior-affecting setting. A run binds cases to one configuration. A result contains trace references, output, operational measures, and grader outcomes.

Safety model

Replay is deny-by-default for side effects. Credentials are scoped by executor and environment. Sensitive fields are classified at capture. Retention applies independently to payloads and metadata. All access to raw content is auditable.

Release gate

A candidate is eligible only when required slices meet quality and reliability thresholds, no critical safety regression appears, cost remains inside budget, and failure examples have been reviewed. The gate creates evidence; it does not replace engineering judgment.

Built as a living AI engineering knowledge base.