Skip to content

The AI-change reference case

An AI coding assistant can propose a network change, open a pull request, and have it merged — sometimes faster than a human reviewer can read it. Cassian Gate's answer to that is deliberately unremarkable: an AI-authored change is validated by exactly the same gate as a human-authored one. There is no separate path for "AI changes," no provenance score, no trust dial. The verdict is reached the only way the gate ever reaches a verdict — by executing the declared scenario on clean-state runtime and binding the result to a declared expectation. (See Scope and scale discipline: in Cassian Gate, AI never holds verdict authority.)

This page walks the canonical reference case that demonstrates that property. The example ships in the repository under contrib/topologies/ai-change-reference-case/.

The shape

The reference case is two small BGP topologies, each in a passing/ and a failing/ form, with the gate's own as-shipped evidence committed alongside them:

  • core-builtin-bgp/ asserts a built-in invariant — bgp_session_up — that the session between two routers establishes.
  • udi-bgp-variant/ asserts a user-defined exec invariant — that a specific route is present in a peer's table.

In both variants the passing form is a correct change and the failing form is a realistic mistake of the kind an assistant might introduce. Nothing about the topology tells the gate who wrote it.

Two ways a change fails — and neither passes silently

The two failing forms are chosen to show the two shapes a gate failure takes, because a gate is only as trustworthy as its weakest "looks fine to me."

The session that never comes up. In core-builtin-bgp/failing/, the change carries a mismatched autonomous-system number, so the BGP session never converges. The gate's control-plane precheck catches this before the invariant runs: the invariant is blocked, and the run is an authoritative FAIL that says so — observed: blocked, with an explicit note that structured detail is unavailable for a check that never got to execute. The point is what does not happen: the gate does not shrug and report success because nothing ran. Silence is not a pass.

The session that comes up missing a route. In udi-bgp-variant/failing/, the change is subtler — the session is healthy, but the router never advertises the prefix it was supposed to. Here the exec invariant does execute: it runs a read-only show bgp ipv4 unicast 1.1.1.1/32 json on the peer, observes an empty table, and FAILs with the observed state rendered — the command, the return code, and the empty output that proves the route is absent.

One failure renders an explicit "couldn't observe"; the other renders exactly what it observed. Both are bound to the same declared expectation (expect: pass), and both fail the gate.

What the reference case does not claim

It is worth being precise about the boundary this case draws, because "AI safety" invites overclaiming:

  • The gate does not detect that a change was authored by an AI. There is no provenance field, no origin flag, no special handling — and the included proof harness checks the engine source to keep it that way.
  • The gate does not score, rank, or rate AI changes. A change is correct or it is not, and that is decided by execution.
  • AI holds no verdict authority. It can propose the change and even draft the invariants; the PASS/FAIL is still produced by running the declared scenario. (This is the same boundary described for autonomous operations and agentic platforms in Scope and scale discipline.)

In other words, the reference case is not a new capability bolted on for AI. It is a demonstration that the gate's existing commitment — execute the change, observe the result, bind it to a declared expectation — already does the right thing when the author happens to be a machine.

Try it

From a checkout, run either variant the same way you would run any topology:

1
2
cassian test contrib/topologies/ai-change-reference-case/udi-bgp-variant/passing/topology.yaml   # PASS
cassian test contrib/topologies/ai-change-reference-case/udi-bgp-variant/failing/topology.yaml   # FAIL (route absent)

The committed evidence/ directory next to each topology holds the gate's own results.json and human-readable summary from a real run, so you can read the exact verdict surface without standing up a lab. The behaviour above is pinned by a lab-free proof harness, tests/ai_change_reference_case_proof.py, which runs in CI on every change.

See also