JevGuard

# JevGuard: Deterministic Decision Runtime for TypeSafe AI (Jev)

[](https://github.com/seb4ez/jevguard-mcp)
[](https://www.python.org/)
[](LICENSE)

> **Official Model Context Protocol (MCP) Server:** JevGuard includes an official, zero-dependency MCP server designed for autonomous AI agents in Google Antigravity, Cursor IDE, Claude Desktop, and Cline: **[seb4ez/jevguard-mcp](https://github.com/seb4ez/jevguard-mcp)**.

JevGuard is a deterministic evaluation, caching, and calibration runtime for TypeSafe AI's Jev model (System One). It wraps standard System One evaluations with closed-world escape injection, certainty calibration, state pruning, volatile key masking, zero-token SHA-256 caching, resilient retry backoff, and episodic session memory using only the Python standard library.

## Empirical Upstream Benchmark (50 Live Production Requests)

The following benchmark report reflects 50 complete requests executed against the official TypeSafe AI endpoint (`https://api.typesafe.ai/v1/systemone` using model `jev-latest`):



| Key Metric | Direct Upstream API | JevGuard Runtime | Empirical Advantage |
| :--- | :--- | :--- | :--- |
| **Production Requests** | 50 live calls verified | 50 live calls verified | 100% pass rate across 5 operational domains |
| **Average Network Latency** | 784.53 ms | 735.88 ms (pruned) | State pruning reduces wire payload size |
| **Deterministic Cache Hit** | 763.40 ms (144 tokens charged) | **0.099 ms** (**0 tokens**) | **7,711x latency speedup; 100% token savings** |
| **Out-of-Scope Protection** | Forced False Positive (`chargeback`) | `UNRESOLVED_OR_OTHER` | **Zero false positive** via neutral escape injection |
| **Uncertainty Calibration** | Raw unverified probabilities | `AMBIGUOUS_STATE` alert | Automatically flags bimodal ties an