# Evolution 80 — MCPGate Report (execution boundary)

## What was exercised

Evolution 80's enforcement boundary is `cain.cain2027.e80.core.Gate`. Every model inference and
every sandbox execution in the benchmark, memory evaluation, skill training, mission controller and
durable drill passes through it:

1. An identity requests a capability it is allowed to request.
2. CAIN's authority key signs a grant bound to agent, tenant, tool, scope hash, unit count,
   single use, expiry and the current fencing token.
3. `Gate.execute` re-verifies all of that at the boundary, burns the grant, decrements budget, and
   only then calls the tool.
4. A refusal is a typed `Denied` reason recorded in a hash-chained audit; the effect is not
   produced.
5. Model calls are additionally signed by the requesting agent and logged with model id,
   weight-file SHA-256, prompt hash, output, tokens, latency, temperature, seed and `cost_usd: 0.0`.

The boundary is reference code inside the E80 package. **The production MCP proxy
(`cain/mcp_proxy.py`) was not modified and was not re-tested in Evolution 80.** No claim is made
that a tool call outside this reference gate was governed.

## Bypass attempts and results

The adversarial lab (`security80.py`, 51 scenarios) includes direct bypass attempts:

| Attempt | Result |
|---|---|
| Call a tool without a grant | refused (`NO GRANT` path is not reachable in the reference API; the only call path is `decide` + `execute`) |
| Forge the authority signature | `GRANT_SIGNATURE_INVALID` |
| Reuse another agent's grant | `GRANT_AGENT_MISMATCH` |
| Reuse a grant after use (15 concurrent replays) | 1 effect, 15 × `GRANT_REPLAYED` |
| Change the tool or the scope | `GRANT_TOOL_MISMATCH` / `GRANT_SCOPE_MISMATCH` |
| Use an expired grant | `GRANT_EXPIRED` |
| Use a grant after the kill switch | `KILL_SWITCH_ENGAGED` |
| Use a grant after revocation | `IDENTITY_REVOKED` |
| Cross-tenant request | `CROSS_TENANT` |
| Exceed the capability manifest | `TOOL_NOT_IN_CAPABILITY_MANIFEST` |
| Exhaust the budget | `NO_BUDGET` |
| Delegate more than the parent | `DELEGATION_EXCEEDS_PARENT` |
| Commit after a newer fencing token | `STALE_FENCING` (queue) / `GRANT_...` (gate) |

51/51 scenarios blocked, 0 bypassed. These are same-project tests, not a third-party assessment.

## What is supported today

- Governed **local model inference** (three local CPU models, pinned by weight-file hash).
- Governed **Python expression execution** in bubblewrap, with syntax/size screening and no network.
- Governed **mission node execution** with a different-identity verifier.
- Governed **skill evaluation/training** calls.
- Single-use, expiring, tenant-bound, budget-bound grants with attenuation and revocation.
- Hash-chained audit and offline verification of the evidence that references it.

## What is not supported / not evaluated

| Item | Status |
|---|---|
| Production MCP tool proxy enforcement | **NOT EVALUATED in E80** (unchanged; not re-tested) |
| Remote APIs, databases, cloud infrastructure, workflow systems | **NOT IMPLEMENTED** as governed connectors |
| Customer-controlled self-hosted deployment | **NOT DONE** |
| eBPF / seccomp / VM-level enforcement | **NOT IMPLEMENTED** |
| Distributed kill switch across hosts | **NOT IMPLEMENTED** |
| mTLS between components | **NOT IMPLEMENTED** |
| Physical AI interfaces | **NOT IMPLEMENTED** |
| Third-party bypass audit | **NONE** |

## Honest summary

The reference gate demonstrates the CAIN rule at a real execution boundary: intelligence proposes,
authority is a separate signed object, and the effect does not happen without it. It is not a
production MCPGate release, it has not been penetration-tested by a third party, and it should not
be described as protecting traffic it did not carry.
