Ashin Walpola
83ccf335bb
Section 11 described how the detector works and the test-design finding from
2026-08-25, but carried no threshold values, no trigger conditions and no
emission semantics. That was inconsistent with section 10.4, which tabulates
all nineteen UseCaseDetectionConfig parameters for the V2X side. Adds 11.3 and
11.4 to close the gap.
11.3 tabulates all twelve DetectionConfig parameters in three columns, because
three different configurations exist and they do not agree. DetectionConfig's
KDoc says its defaults match the Phase A specification; TripRecordingService
overrides nine of the twelve when it constructs the detector, every one of them
in the direction of lower sensitivity. The shipping detector is not the
specified detector, and that was recorded nowhere outside a constructor.
11.4 gives the input rates, the qualifying condition for each of the three
event types, when each emits, and how confidence is assigned. It also explains
why braking compares against a reference speed latched at onset rather than a
per-frame delta: GNSS updates at 1 Hz against a 50 Hz detector, so a per-frame
delta is non-zero on one frame in fifty and could never coincide with a
25-frame sustain requirement. That is the same sampling lag that made four
tests unsatisfiable, seen from the implementation side.
Two discrepancies found while writing this are recorded rather than fixed,
since fixing either changes behaviour and belongs in its own change:
- EventDetectorTest states it keeps production thresholds for all signal
values. The values it keeps are the DetectionConfig defaults, not the ones
TripRecordingService runs. All 18 tests validate a configuration that never
executes on a phone. The logic under test is shared, so they remain valid
logic tests; they are not evidence about the shipped system.
- brakingHighConfidenceRate is documented as a rate in m/s per GNSS update but
is compared against the peak cumulative drop from the onset reference, which
is not a rate and grows with episode length. HIGH confidence is therefore
assigned more readily than the name implies.
Also notes the emission asymmetry: turning and stopping emit once per episode,
braking re-arms and re-fires roughly every half second at the shipping values.
Edited in place through the existing package rather than regenerated, so Word's
own parts and the manual edits from 312f094 survive. All sixteen package parts
verified present afterwards, section order unchanged, all nine image
placeholders intact.
92 KiB
92 KiB