skills/ce-code-review/references/cross-model-eval.md

# Cross-Model Adversarial Pass — Skill-Creator Eval Spec

This is the load-bearing behavioral eval for ce-code-review's cross-model
adversarial pass. Deterministic route tests cover the worker; these cases cover
the SKILL.md/reference orchestration that only a fresh agent can execute. Inject
the current `SKILL.md`, `references/cross-model-review.md`, and the relevant
Stage 5 synthesis prose through the `skill-creator` workflow. Run on Claude Code
and Codex with fake peer CLIs first on PATH.

## Eval cases

1. **Activation fires only on the existing gate.** A local-aligned or standalone
   diff that selects `adversarial-reviewer` launches one detached cross-model
   adversarial job in the Stage 4 wave. A trivial diff that does not select the
   persona launches none. A `pr-remote` or `branch-remote` review launches none
   even when adversarial analysis is warranted.

2. **Host identity and fixed route precede egress.** The orchestrator keeps host
   harness and serving family separate, excludes an attestably same-family
   target, resolves exactly one target and concrete route, verifies every
   recipient against the allowlist, discloses the reviewed-code egress, and only
   then dispatches with `CROSS_MODEL_HOST_HARNESS` and
   `CROSS_MODEL_FIXED_ROUTE`. An unattested host serving family skips automatic
   dispatch.

3. **Cursor and Composer remain distinct.** A stated `cursor` preference selects
   Cursor default/Auto with no `--model`; `composer` selects an explicit
   Composer-family model through Cursor. A receiptless Cursor or Composer return
   records `model_actual: unverified` and `independence_verified: false`.

4. **Fold-in promotion requires verified independence.** A stubbed
   `adversarial-<provider>.json` finding matching the in-process adversarial
   finding promotes one anchor step only with `independence_verified: true`.
   With `independence_verified: false`, it remains attributed evidence and does
   not count as different-model corroboration. Peer findings never gain silent
   apply authority.

5. **Failures are additive and non-blocking.** A missing CLI means no job starts;
   a human-facing markdown review reports the pass as not run. A started timeout, failure,
   or unusable return is named in Coverage and the in-process review completes.
   The worker never changes recipients internally; any recipient-changing retry
   requires a new disclosure and sanction.

6. **Preferred-first bounded adaptation.** The declared mapping is attempted
   first. Only an observed unavailable, obsolete, or incompatible model permits
   a same-target/same-family override bound by
   `CROSS_MODEL_MODEL_OVERRIDE_TARGET` and `CROSS_MODEL_MODEL_OVERRIDE`.
   Cross-family substitution, override leakage, silent explicit-model changes,
   and new recipients are rejected.

7. **Detached lifecycle remains bounded.** The orchestrator starts the peer job
   in a short runner call, polls in bounded slices while other review work
   continues, reaps at the aggregate deadline, reads owned results and skip logs,
   names non-`done` terminal states, and removes private scratch.

8. **Mode-specific disclosure is honest.** Human-facing default mode announces
   the fixed route and egress before dispatch and calls it independent only when
   serving families differ attestably. Receiptless routes say "requested
   <model>; serving model unverified on this route." `mode:agent` emits no
   user-facing prose but retains the worker's stderr audit record.

9. **Oversized diffs recover without one giant prompt.** A fixture above the
   inline token or file-count limit sends the peer the orchestrator's compact
   semantic review map plus a private exact-diff path, never the whole diff.
   The worker does not cut semantic shards or invent risk divisions; the
   orchestrator does, and the adversarial agent reads bounded ranges and narrows
   them further rather than returning a progress note or silently omitting the
   pass. A normal-sized fixture keeps the direct diff path.

## Pass criteria

All nine cases pass on the current on-disk source on Claude Code and Codex. The
negative activation cases launch no peer, the fixed-route cases perform no
worker-internal recipient fallback, and only `independence_verified: true`
artifacts can promote agreement.