# Sovereign Evolution 1: contribution from session -17

Scope: the `/verifygate` Z3 engine, execution routes that do not go through MCPGate, what host enforcement really does, and evidence-ledger checks. The machine-readable record is `findings.json`; session -46 merges it into the top-level Evolution 1 files.

Baseline revision `6cfa3b9e`. Fix commits `3537d76e` and `a098fa61`. Deployed 2026-10-12 (rolling, deploy/20261012T034805Z), verified on blue and green and on the public domains.

## What was fixed (reproduced first, regression tests fail on the old code)

| ID | Problem | Fix |
|---|---|---|
| D-S17-01 | `prove_monotonicity` treated an unknown ceiling name as `NONE`, so ADMIN → "SUPERUSER" came back as proved | Unknown names give `INVALID_INPUT` |
| D-S17-02 | `prove_memory_containment` priced an unknown operation as WRITE, so VERIFIED/ADMIN/"SHUTDOWN" was proved | Same |
| D-S17-03 | A solver UNKNOWN was reported as `SAT` with a made-up counterexample; a solver exception gave an HTTP 500 | `UNKNOWN` and `ERROR`; only `UNSAT` proves |
| D-S17-04 | Half-life was clamped to at least 1 s, giving a looser bound than the real decay; NaN crashed the solver | No clamp; non-finite or non-positive input gives `INVALID_INPUT` |
| D-S17-05 | Anonymous `POST /fabric/continuum/verify/run` made the gateway spawn the heavy harness worker | Operator credential required; the cached GETs stay public |

Tests: 30 new verifygate tests (26 failed on the old engine) and 5 new continuum tests (3 failed on the old router). With the fixes, the new tests plus every suite that imports the engine: 106 passed, 0 failed.

## What was found and left open

- **I9, offline fallback: FIXED** in 9672bdc1 (owner-approved). An unreachable gateway now fails closed in both `cain.guard` and `cain_sdk.guard` unless `CAIN_ALLOW_OFFLINE_LOCAL=true`. Installed SDKs get it with the next published wheel.
- **I3, unsigned gateway decisions.** `cain.guard` trusts the `decision` field of whatever answers at the gateway URL and caches that ALLOW.
- **Kernel enforcement overclaim: wording corrected in every IP recalc V33-V45** (9fc79332, 656dac9a; owner-approved). The signed 8.1.0/8.2.0 kits keep the old V45 text; the next kit carries the fix. The missing kernel enforcement itself is still open. `cain-ebpf/cain_enforce.bpf.c` attaches only to syscall tracepoints and always returns 0, so it observes and blocks nothing. Published evidence (IP recalc V38, "EBPF TRACEPOINT BOUNDARY INTERCEPTOR") says it rejects execution and egress in the kernel. The real deny mechanism in the repository is `cain/ebpf_enforcer.py` (`bpf_override_return`, `-EPERM` before the syscall runs). It was not re-run here because it attaches kprobes on the live host. The BPF LSM is not active on this host.
- **No durable MCPGate ledger.** MCPGate's `evidence_digest` is mostly a hash of the reason text, and the proxy writes no chain of its own. The substrate proof-run chain does not hash the counts it returns, but nothing in production calls it.

## What this does not show

- The verifygate models compare constants. UNSAT means only that the stated comparison cannot be violated with those numbers. It says nothing about an agent, a program or the host.
- Route coverage came from reading the source. The continuum route was POSTed in production only after the fix was live on both instances, to confirm the 403; no POST was sent to the vulnerable version.
- A green run is not production readiness. The SDK offline fallback, unsigned gateway decisions and the lack of kernel enforcement on the hosted path each block a release claim of complete mediation.
