SOURCE DATA
Two valid activities.
Garmin/freddy returned a treadmill session and a Push B strength session from the same day.
FAILURE CASE · PROVENANCE & RELATIONS
Correct facts can still produce a wrong story.
Two same-day activity records were individually correct. The AI nevertheless produced a false chronology because it treated connector return order as time order.
This is not a Garmin or freddy measurement error. The failure occurred in the interpretation layer that connected two valid records with an unsupported temporal relation.
WHAT HAPPENED
A relation was invented between correct records.
AI INFERENCE
Return order became chronology.
The conversational layer interpreted the ordering of returned records as if it proved sequence and verbalized the second item as happening “afterwards”.
REALITY
The sequence was reversed.
The treadmill session actually happened before the strength session. The user corrected the narrative.
SYSTEMS LESSON
Provenance applies to relationships too.
Most data-quality checks focus on whether individual values are correct. This case shows a second failure mode: the values can all be correct while the relationship between them is unsupported.
A fluent narrative can be wrong without containing a single fabricated measurement.
Chronology, causality and “before/after” statements therefore need their own evidence. Connector result order is not a timestamp.
GUARDRAIL
Do not infer temporal order from list order.
When reliable start timestamps are unavailable, the system should either retrieve a source that can establish order, ask the user when the relation matters, or explicitly say that the sequence is uncertain.
LIMITATIONS
One observed failure, not an error rate.
This is a single documented episode. It does not estimate how often this class of error occurs, and it should not be generalized into a claim about Garmin or freddy data quality. Its value is methodological: it exposes a subtle failure mode in longitudinal AI interpretation.