Guard .NET LLM apps with typed checks

01 · Role in the system

What Jev does here

Pattern matching misses anything paraphrased; asking a second LLM doubles cost and returns text you have to parse. Kassad's third route asks many small, well-scoped questions of a model built for exactly that shape and thresholds the numbers in code - cheap enough to run on everything, with confidence as a first-class signal: a flat distribution means I don't know, and your policy says what that earns (usually review) instead of a guess. Every policy for a stage batches into one request, so ten checks cost one round trip. The 0.1.0 release ships the engine, policy format, TypeSafe client, OpenAI- and Anthropic-shaped middleware, and an eval harness over three inbound datasets and a grounding set - with the README insisting the published numbers came from public single-turn sets, not your traffic, before production use.

02 · Control boundary

Where Jev sits

Per-stage policy batching to one typed-check request; code thresholds probabilities into four verdicts and owns every outcome.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• First release at 0.1.0 with a public API that can still change before 1.0.
• The committed eval numbers are from public single-turn datasets with the sample policy, explicitly not a production-traffic claim.
• .NET only - the runtime story for other ecosystems needs other libraries.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• jacob-berendsohn ↗Community · github · public · checked 2026-09-24