ObliviOS Developer portal Project site

Home / Decisions / ADR-063

Close docs/superpowers/plans/2026-08-28-duress-decoy-session-indistinguishability.md.

Accepted, recorded 30 August 2026.

Decision
ADR-063
Date
2026-08-30
Status
AcceptedAccepted (qualified — several surfaces remain honestly unchecked, and Component A's original P0 bug remains open by design per ADR-061)
Part
Duress Unlock
Mentions
ADR-059, ADR-061, ADR-062
Mentioned by
ADR-064, ADR-066, ADR-067, ADR-068, ADR-095

1Decision

Close docs/superpowers/plans/2026-08-28-duress-decoy-session-indistinguishability.md: the owner flashed ADR-062's candidate to the physical lynx device, and Task 11 Steps 2-4 (rescoped Component B/C/D on-device validation, the latency-remeasurement scope decision, and this closing record) are complete. All 11 tasks of the plan are now done.

2Why and evidence

The controller independently verified the flash before trusting the owner's report: ro.build.version.incremental=20260829213857 matches the signed candidate exactly, ro.boot.verifiedbootstate=orange/vbmeta.device_state=unlocked unchanged as expected, and a full dumpsys dropbox scan for crash\|anr\|tombstone\|fatal returned zero matches both immediately post-flash and again at the end of the whole test session.

Real, direct evidence on physical hardware: Component B's QS user switcher and Settings ▸ Storage both show only the decoy's own identity, no leak; Component C's USER_SETUP_COMPLETE registration-time write was proven with a genuine controlled before/after (the controller reset the flag to a known 0 baseline, the owner ran the real set-decoy-profile re-registration themselves, the flag flipped back to 1 as a direct, timed effect — not a stale-state coincidence), and its screenshot-overlay and QS-settings-gear consequences both fire correctly; the non-decoy regression check passed with stronger evidence than the emulator run (am switch-user to an ordinary locked user is refused outright by AOSP itself — SWITCHABILITY_STATUS_SYSTEM_USER_LOCKED — while the device sits locked, and a legitimate switch away from an unlocked non-owner session correctly re-shows keyguard for a different locked target, with no bypass in either case).

A genuine device constraint was found along the way and worked around, not ignored: org.aegisos.usertype.full.DECOY enforces a real max-count-of-one, so Component C's registration-time-write proof had to reuse the device's single existing decoy user (reset-and-re-register) rather than spinning up a clean second one.

The credential-handling hard rule was upheld throughout by construction: the controller never entered, requested, or extracted the owner's real PIN at any point — the one command that needed it (set-decoy-profile --owner-credential) was handed to the owner to run in their own terminal, exactly as this project's standing rule already requires for fastboot flash.

A related, unplanned finding: simulated adb shell input tap sequences consistently failed to dismiss the decoy's own on-screen PIN bouncer even with verified-correct button coordinates and a controller-confirmed-correct PIN — consistent with GrapheneOS hardening against synthetic touch input on FLAG_SECURE keyguard surfaces, a real positive security property noted for the record but not investigated further as out of this plan's scope; the owner dismissed it via their own real physical touch, matching ADR-059's precedent.

Step 3 (latency remeasurement) was explicitly skipped, not silently dropped: Component A never shipped (ADR-061), so there is no double-PIN removal on this candidate to measure against ADR-059's baseline — re-running the old instrumentation would only reproduce the same already-on-record figures.

Several surfaces remain honestly UNCHECKED, consistent with Task 10's own reporting standard rather than forced into a pass: the lock-screen avatar switcher, the com.android.multiuser widget (impractical via adb alone), the clipboard overlay (impractical UI automation), the alarm icon (environment-wide limitation, not decoy-specific), and Material You theming (now confirmed non-emulator-specific — the same non-correlating data pattern appears on real hardware too).

Component D's individual P3 surfaces were not independently re-tested; judged sufficiently corroborated by Component B's now-confirmed-working shared filter mechanism.

3Consequences

Full evidence: .superpowers/sdd/task-11-phaseB-report.md. Disposable test artifacts created during validation (aegisos-hw-regression-check, id 11) were removed; the pre-existing decoy user (id 10) and its now-correct user_setup_complete=1 state were left in place, matching the device's state as handed back by the owner. screen_off_timeout was changed to 600000ms during testing and not reverted, since its original value was never recorded — flagged for the owner to reset via Settings if it mattered to them.

4Review when

This closes the Stage 3 plan as a whole.

Two things remain explicitly open for future work, not forgotten: a safe redesign of Component A's double-PIN fix (per ADR-061's own follow-up note — scope the keyguard skip to a decoy-specific signal, not AOSP's generic and too-permissive trust conditions), and the handful of UNCHECKED surfaces above, none of which have any evidence of failing, but none of which have positive evidence either — revisit if a lower-effort real-hardware technique for either the multiuser widget or the secure-bouncer synthetic-input limitation becomes available.

Task 12 (final whole-branch review + superpowers:finishing-a-development-branch) is the next and last step of this plan's own execution sequence.

Read from docs/governance/decision-log.md, entry ADR-063.