- Decision
- ADR-079
- Date
- 2026-09-04
- Status
- AcceptedAccepted, corrects ADR-078
- Part
- Duress Unlock
- Mentions
- ADR-041, ADR-046, ADR-062, ADR-064, ADR-072, ADR-073, ADR-078
- Mentioned by
- ADR-086
1Decision
Correct ADR-078's signing-key attribution and re-flash the decoy primary-profile visual parity candidate with the real alpha key; independently confirm all four gated apps' real on-device presence.
2Why and evidence
A closing whole-branch review (most capable model, dispatched after ADR-078 was already written) found ADR-078's own signing-key claim contradicted ADR-072 and ADR-073 — both one day earlier, on the same device, explicitly citing "the alpha key (265ade399a0f3ee65003e56daef26a8f4d70e6a4)" — while ADR-078 claimed the signing host's vault key (<vault>/keys/lynx/avb.pem) was correct "since ADR-046." Investigated directly rather than guessed: avbtool extract_public_key against the real alpha_key.pem on the build server (<build-root>/keys/alpha/alpha_key.pem) confirmed its real fingerprint is exactly 265ade399a0f3ee65003e56daef26a8f4d70e6a4, matching ADR-072/073 precisely — the controller's own earlier openssl-based fingerprint attempt had used the wrong hashing method (raw DER SubjectPublicKeyInfo hash, not AVB's own public-key-blob encoding) and produced a spurious mismatch that was never resolved before ADR-078 was written.
The router's own much-earlier note that "the signing host's ADR-041-generated vault key, not the alpha key, now signs this device's vbmeta" (from the OTA-signing-infrastructure era, ADR-046-049) was real but stale by the time of this candidate — the project's actual, current, repeated practice for physical duress-unlock hardware-validation candidates specifically (ADR-062, ADR-064, ADR-072, ADR-073, all citing the same fingerprint) had already reverted to the alpha key well before this session, and the controller cited the older, superseded router paragraph instead of checking the two most recent, directly-on-point ADRs first.
Corrected: all 12 partition images (the same set ADR-078 already assembled, system_ext.img included) re-signed directly on the build server with <build-root>/keys/alpha/alpha_key.pem, matching ADR-072/073's exact parameters (SHA256_RSA4096, rollback index 1783209600, --padding_size 65536), 13/13 independently verified on two machines (the build host's AOSP-tree avbtool.py and the controller's own workstation /usr/bin/avbtool, a genuinely separate install).
The owner reflashed the corrected candidate (fastboot flashall, no -w, same as ADR-078).
Post-flash, independently confirmed: ro.boot.verifiedbootstate=orange, ro.boot.vbmeta.device_state=unlocked, zero crash/ANR/tombstone, both patch sets still present by direct dex/jar string extraction (unaffected by the re-sign, since only the vbmeta signature changed, not any partition's content) — and, closing the gap ADR-078 itself was already missing, all four gated apps' real installed package names confirmed present via adb shell pm list packages directly on the flashed device (app.seamlessupdate.client, app.grapheneos.pdfviewer, plus Seedvault and Auditor), not merely installed-files.txt's build-artifact listing.
3Consequences
The device's actual, currently-trusted AVB root of trust for physical duress-unlock validation candidates is the alpha key (265ade399a0f3ee65003e56daef26a8f4d70e6a4) — this is now internally consistent across ADR-062, ADR-064, ADR-072, ADR-073, and this entry; ADR-078's own text is left unedited (this project's decision-log discipline is append-only correction, not silent rewrite) but should be read as superseded on this one point.
Nothing else in ADR-078 (the build-defect root cause, the two signing-ceremony procedural incidents, the disk-staging-location lesson, the physical-hardware dynamic-validation results for the "End session" and "Owner" label fixes) is affected — those findings are independent of which key signed the candidate and stand as recorded.
4Review when
Before citing "the vault key" or "the alpha key" for any future physical lynx candidate, check the fingerprint directly (avbtool extract_public_key) against both <build-root>/keys/alpha/alpha_key.pem and <vault>/keys/lynx/avb.pem on the signing host, rather than reasoning from a router paragraph's prose alone — the router narrates history in the order it happened and is not guaranteed to reflect which key is current practice at any later point without checking the most recent same-class ADRs first.
This project's own decision-log format-compliance script or future review should flag a same-day/next-day contradiction between two entries' factual claims automatically if one becomes available.
Read from docs/governance/decision-log.md, entry ADR-079.