**Review:** `docs/superpowers/plans/2026-10-10-instance-key-provision.md`, rev 4 (3e42e8e7). It answers your round 4 (`CODEX-PLAN-INSTANCE-KEY-R4.md`). **Ruling 26 is Paul's decision:** a replica gets its home's key. It replaces rulings 17 and 25. The rev 3.1 strip is reverted.
4
5
**Mechanism.** Before any replica process starts, the agent allows it to run in only two cases:
6
- with exactly its home's key, verified against the home key's digest;
7
- with no key but with a control socket, so that the helper refuses instead of minting.
8
9
With no key and no socket (seeds, claim launches before O6, a failed socket bind), the agent refuses the launch itself.
10
11
A pre-roll check records verified home keys in `CARLOS_REPLICA_HOME_KEYS` before any box is upgraded. Replicas that already share their home's configured key therefore keep working.
12
13
**New obligation O8 (replica placements, #162).** The replica record names its home. The replica reads the home's published `<host>.env` item, verifies it by digest, and writes it create-only. Until O8 lands, a keyless replica refuses to start.
14
15
**Judge:**
16
1. Are both round 4 findings closed?
17
2. Is the replica refusal reliable on every launch path (seed, spawn, nested, claim, socket-bind failure)?
18
3. Can a replica ever run with a key other than its home's?
19
4. Is the pre-roll check sound? Is `CARLOS_REPLICA_HOME_KEYS` safe: is it exposed to the tenant, is it in logs?
20
5. Is O8 well-defined, and is its fallback safe?
21
6. Anything new?
22
23
Give findings P1/P2/P3 with file:line, then a verdict: ready to build, or not.
amadan