faultgate 0.3.0: a refund over the limit fails, even inside the noise
Business rules on tool arguments, an honest INCONCLUSIVE verdict with sample size math, vendor trace import, and zero cloud telemetry. Release gates for agents.
BLOG
Notes from the workbench.
Field notes, empirical failure analysis, and design records from building deterministic agent reliability tooling.
Regression or noise? A release gate that measures first
You run the same model five times on the same thirty inputs and pass rates swing from 0.60 to 0.87. A fixed threshold is a coin flip. faultgate measures your spread instead.
A local firewall for MCP tool calls
mcp-guard sits locally on stdio between client and tools. It logs calls with argument names only, flags poisoned descriptions, and blocks rule violations in enforce mode.
Why a single successful run tells you almost nothing
Why pass^k across repeated trials reveals agent reliability failure modes that single-shot benchmarks systematically miss.
