- Decision
- ADR-105
- Date
- 2026-09-27
- Status
- AcceptedAccepted.
2026092501is now independently verified 13/13 on two separate machines, closing ADR-103's one open item. - Mentions
- ADR-097, ADR-099, ADR-103, ADR-104
- Mentioned by
- ADR-103
1Decision
Complete the second-machine AVB re-verification this row's own review trigger asked for: derive the vault's AVB public key as a PEM (openssl pkey -in <vault>/keys/lynx/avb.pem -pubout) on the signing host and use it on the build host to independently verify candidate 2026092501. Supersedes ADR-103's "partial" status with a full pass.
2Why and evidence
The derived PEM matched a pre-existing <home>/avb_pub.pem on the signing host from 2026-08-24, confirming the vault key has not rotated since.
The shared mount's real path differs by host: it is <build-root>/signed on the signing host (virtiofs source build-signed) but <build-root>/signed on the build host — the ADR-097 aegisos→sovereignos rename never reached the signing host's live mount, a stale-record-vs-live-state gap the first attempt hit directly (mkdir/openssl silently failed against the non-existent <build-root>/... path on the signing host).
The isolated verify directory needed the complete 13-partition set the vbmeta actually describes (boot, dtbo, init_boot, pvmfw, vendor_boot, vendor_kernel_boot, product, system, system_dlkm, system_ext, vendor, vendor_dlkm, plus the vbmeta struct itself) extracted fresh from the published bundle into <build-root>/verify/avb-check-2026092501/ on the build host — avbtool verify_image reported all 13 Successfully verified, exit 0.
Scratch extraction (~3.5GB) was removed afterward.
3Consequences
This session's local commits (ADR-099 through ADR-104) plus this row are still not pushed to origin/master. The the signing host-side mount naming (aegisos) should be corrected to sovereignos (or documented as an intentional exception) so this doesn't cost another failed attempt next time.
4Review when
The the signing host mount is renamed to match, or a future ad-hoc the signing host path assumption should explicitly check the live mount table first rather than inferring the path from the build host's naming.
Read from docs/governance/decision-log.md, entry ADR-105.