- Decision
- ADR-204
- Date
- 2026-10-30
- Status
- ProposedProposed (owner's instruction to use the recordings and not to measure again; module-built, in
lynximage2026103001,GRB-024not run on it) - Part
- Grab detection
- Mentions
- ADR-195
- Mentioned by
- ADR-236, ADR-260
1Decision
Amend ADR-195 (owner, 2026-10-29: "I pick it up and it already locks"): an impact with no free fall locks only after a pulse. The impact route of grab detection decides at the first sample whose deviation from gravity is at least impactPeak (30 m/s^2). An impact after a free fall of at least 250 ms (a real fall) locks at once and one after a shorter fall (a set-down) does not, as before.
An impact with no free fall now waits until the magnitude has been at or above 15 m/s^2 for 120 ms (minPulseMs): the rise before the first sample over impactPeak counts, the time after it is summed within 400 ms, and the lock comes as soon as the sum reaches 120 ms; an impact that does not show that pulse is refused like a set-down (decided once, no cool-down, the rest of its pulse is not judged again). The impact line of the log gets pulseMs.
Patch frameworks-base/0011; no new setting, the 120 ms is a constant of GrabDetectionService.
2Why and evidence
[measured] the recordings of tools/grab-recorder (200 Hz, lynx, 2026-10-18 and 2026-10-25) with tools/grab-recorder/events.py: the 12 recorded grabs are 171 to 310 ms above 15 m/s^2, the 8 pickups and bumps with no free fall 4 to 13 ms (a pickup that was lifted slowly, 79 and 84 ms, did not reach 30 m/s^2); [measured] the real GrabDetector replayed in 5 phases of a 50 Hz grid: with the rule 12 of 12 grabs lock and no pickup, bump or set-down does, without it the pickup spike (46 m/s^2, 4 ms) locks in 2 of 5 phases and a spike of 13 ms in 1 of 5; [measured] on the phone the pickup of 2026-10-29 locked (peak=30.4 fallMs=0) and the 12 grabs of the recordings give 30 to 50 at their first sample, so the peak does not separate them (GRB-024); [verified-in-source] the same detector runs in shadow and active mode; 89 of 89 mutants of the detector are killed.
3Consequences
A pickup, a bump and a hand that taps the phone no longer lock it; a grab, a real fall and (as before) a set-down behave as before; the lock of a grab comes up to 120 ms later than at the first sample (the lock screen, the restart timer and the vibration follow the decision).
4Review when
Rejected: a higher impactPeak (45 m/s^2) misses the owner's gentler grabs, whose first sample is 31 to 41, and does not separate a pickup that peaks high; a rule on rest before the impact (a phone that lay flat), because a phone held still in a hand also lies still and a snatch from a table is not the threat model; a minimum pulse for every impact, because a real fall lands in 40 to 140 ms and must lock.
The risk that remains is a grab shorter than 120 ms, none of which is recorded (the shortest of 12 is 171 ms, those of 2026-10-29 were not recorded): the log shows pulseMs of every impact from the next image on, and minPulseMs is one constant. The first 4 ms spike of a pickup that locked in 2 of 5 phases in GRB-019 is the case this closes; ADR-195's remark that a minimum pulse width was left out is superseded.
Read from docs/governance/decision-log.md, entry ADR-204.