- Decision
- ADR-022
- Date
- 2026-08-01
- Status
- AcceptedAccepted
- Mentions
- ADR-021
- Mentioned by
- ADR-023, ADR-024, ADR-029
1Decision
Record the confirmed (partial) root cause of the ADR-021 reproducibility failure: an unpinned, date-derived build number, not build-environment compromise. Baseline stays Rejected pending a re-run with it pinned.
2Why and evidence
Investigated on the existing build artifacts (no rebuild): repro's system/build.prop carries ro.build.version.incremental=2026080100 / ro.build.date=Sat Aug 1 04:01:57 UTC 2026; the primary build's recorded fingerprint carries 2026073000 — each matches its own build's actual calendar date. Neither build's log sets BUILD_NUMBER or SOURCE_DATE_EPOCH. build/make/core/version_util.mk in the synced source confirms HAS_BUILD_NUMBER defaults false and build-number generation "moves to soong_ui" — i.e. soong_ui stamps a fresh date-derived value into every partition's build.prop unless explicitly pinned.
3Consequences
Explains, with high confidence, why system/vendor/product/system_ext/system_dlkm/vbmeta (and their -qemu counterparts) differ: vbmeta.img signs a dm-verity root over that content, so any build.prop change cascades into it. Images without build.prop (dtb, encryptionkey, super_empty, the vendor_ramdisk*/vendor_boot-test-harness set) correctly match.
This is a partial root cause: cache.img, userdata.img, ramdisk.img, and ramdisk-qemu.img also differ but were not individually confirmed against this same mechanism — they may share it or stem from a related but distinct cause (unpinned filesystem-image creation timestamps/UUIDs, also not fixed via SOURCE_DATE_EPOCH in either build).
This is a known, fixable non-determinism source, not evidence of tampering or a compromised toolchain, but it does not by itself flip the ADR-021 Rejected baseline decision — no re-run with the build number pinned has been performed to confirm a clean diff.
4Review when
A follow-up comparison run pins BUILD_NUMBER (or the soong_ui-recognized override) identically between both workspaces and either reproduces a clean diff (supports re-accepting the baseline) or still shows a residual difference (narrows the cause further, e.g. to filesystem-image timestamps).
Read from docs/governance/decision-log.md, entry ADR-022.