# CAIN-A75-033 Clinical Tool Boundary Review

**Verdict:** VERIFIED

> Recomputes a clinical MCP tool surface: human owner and pinned version, unique ASCII names, no hidden instruction or smuggling, typed and bounded patient/consent parameters, an exact allowlist, and scanner findings, risk score and verdict equal to the deterministic recomputation.

| # | Verdict | Tools | High | Medium | Risk score | Scan verdict | Server | Failed checks |
|---|---|---|---|---|---|---|---|---|
| 0 | SOUND | 3 | 0 | 0 | 0 | ALLOW | pathology-mcp |  |

## What this proves and what it does not
- Proves: the surface names a human owner, a transport and a pinned version; every tool has a unique lower-case ASCII name, no hidden instruction, no zero-width or bidirectional smuggling and no destination URL under a network-forbidden policy; every parameter is typed and every sensitive parameter is bounded; the tools are inside the declared allowlist and size limit; and the recorded scanner findings, risk score and verdict equal the deterministic recomputation.
- Does not prove: that a clean definition is safe to call, that the scanner finds every hidden instruction, or that the downstream tool enforces its schema. It is a static, heuristic review of the definitions; nothing is called.
- Verify offline: `python3 verify_offline.py run.json --pinned-key WSO9t4PhJswR5ilC9PKcSyVz8xLLjcE4XkJH7tKCgD4=`
