Dimensions can contradict: location right, activity wrong

What it is

Context dimensions can contradict each other: the location judgement "at home" is right while the activity judgement "sleeping" is wrong (the person is exercising; a rising heart rate was misread). High confidence in one dimension says nothing about the composite. Cross-dimension consistency is therefore an independent test of context quality: a combination like "at home + sleeping + every bedroom light on" should look suspicious to the system even if each dimension's individual confidence is respectable.

Here contradiction is a resource, not a nuisance: it is the earliest failure signal. Sensor drift, over-fitted models, stale calibration — the silent faults that raise no error flag — tend to surface first as cross-dimension inconsistency.

Why it happens

Context systems infer each dimension independently, each with its own confidence; combination happens downstream. Two paths make contradiction inevitable:

• Errors across dimensions are independent. Location rides on GPS/Wi-Fi, activity on inertial sensors and heart rate, and their error sources share nothing — one can be accurate precisely when the other is wrong. Multiplying per-dimension confidences into a "composite confidence" assumes uncorrelated errors and fails worst exactly under systematic faults (a broken sensor).
• Reality itself "contradicts". Falling asleep on the sofa, walking while in a meeting — real activity combinations violate ordinary templates. A contradiction therefore admits two readings: the inference is wrong, or the reality is atypical. The system cannot tell them apart a priori; the sound move is treating conflict as a signal to lower confidence, not adjudicating which side wins.

People run this logic natively: seeing lights off but hearing voices, the first reaction is not "which observation is fake" but "odd — check again". Consistency checking automates that instinct.

Studying it

• The sensor-fusion literature offers ready conflict detection frames: compare sources' estimates of the same quantity, flag disagreement beyond a threshold. Extended to heterogeneous dimensions, "conflict" loosens from "numeric difference" to "combination falls in a low-probability region", which the training data's joint distribution models directly.
• In activity recognition, multimodal agreement serves as an unsupervised quality probe: segments where audio-based and video-based inference persistently disagree are rich in labelling noise and model blind spots — the same idea transfers to production systems as drift early-warning.
• Evaluation practice: inject single-dimension errors (replace one dimension's reading with a stale or random value), measure the detector's hit and false-alarm rates; then inject real-but-atypical combinations, measure how often they are wrongly flagged. Both error rates must be reported together.

One methodological caution: the baseline for "atypical" drifts across households — a singleton's and a family-of-five's "normal combinations" differ wildly. Consistency models need per-household calibration before the baseline is usable; a global model condemns vast amounts of ordinary life as contradiction.

Where it stops holding

• Suppressing contradictions hurts real lives. Filtering out combinations like "asleep in the living room" produces the absurdity of a system refusing to believe the user just dozed off in front of the TV. The ceiling on contradiction handling is demotion and re-confirmation, never "correction" toward plausibility.
• Consistency checking adds latency. Cross-checking before every composite judgement taxes the real-time performance of fast dimensions; not worth it for low-consequence actions (turning on a light), indispensable for high-consequence ones (arming away-mode).
• The check presumes errors are genuinely independent — if two dimensions share a source (activity and presence both inferred from the same camera), their "agreement" is one fault vouching for itself and the check is void. Audit the lineage of data sources first.

Applying it

• Run a cross-dimension consistency check before composite judgement: a combination in a low-probability region lowers confidence, and high-consequence automations demote to confirm-or-defer.
• Log every conflict event: which combination, when, how long. Where conflicts cluster is where drift or over-fitting lives — a free health monitor.
• Annotate each dimension's data lineage (which physical sensor, which inference chain); dimensions sharing a source are exempt from mutual consistency checking.
• How to check: run single-dimension-error injection drills and read the detection rate; at the same time, track how often genuine atypical combinations get demoted in daily use — the first number high, the second low, both together count as passing.

Related

• Same group: Z2.06.1 Context splits into physical and social dimensions · Z2.06.2 Static context like location changes slowly; dynamic context like activity changes fast · Z2.06.3 Personal and shared context have different capture boundaries
• Nearby: Z2.01 Limits of sensors · Z2.02 Uncertainty of inference
• Search terms: context consistency · sensor fusion conflict detection · multimodal agreement · anomaly detection