What an average won’t tell you.

Comparing 2026-08-03 with 2026-08-10 across 2 matched workflow/market groups. Task-weighted accuracy changes +11.0 pp, but the fixed-mix change is -5.0 pp.

Accuracy must apply to the Tasks denominator. Rounded rates produce approximate error counts. These aggregated rows do not support individual-level significance tests or causal claims. No unmatched groups were silently excluded.

Explicitly “Not helpful” outcomes represent 16.7% of records but 38.6% of recorded handling time. Prioritize measured friction rather than the most common words.

Channel mix, repeat contacts, and who chooses to submit feedback may bias results. “Helpful” is not a verified resolution. No causal or population-level conclusion is claimed.

The supplied probabilities imply 158,700 in weighted value from 271,000 in face value. Lumen alone represents 66.5% of that weighted value. Values retain the source currency; no currency is inferred.

These are user-supplied probabilities, not validated win rates. There is no quota, close-date horizon, or win/loss history here, so no attainment forecast or confidence interval is justified.

37.1% of source cells are blank, but 24 blanks occur in fields that are always empty within their event type. Review the schema before treating them as data loss.

0 dates inside the observed span have no records. Even active dates may be partial. No daily baseline, health interpretation, or feeding/sleep relationship is claimed from incomplete logging.