← All briefings
Enterprise AIData qualityPilot evaluation

China AI Energy Reports: Missing Readings Are Not Zero

A correct subtotal can still hide missing measurements. Test how an AI report handles true zeroes, incomplete intervals and a window with no readings.

China AI Tour Research5 min read
Conceptual technical illustration of a server rack connected to cooling pipes and equipment, with dashed measurement boundaries.
Conceptual technical illustration of a server rack connected to cooling pipes and equipment, with dashed measurement boundaries.

An AI-written energy summary can mistake a missing reading for a genuine zero and describe incomplete data as reduced consumption. During a China enterprise-AI evaluation, follow one small input table through the calculation, chart and written conclusion. All three should preserve the distinction between an observed zero and an unknown value.

Use a synthetic, read-only fixture for a single electricity meter. The values below represent energy consumed within four consecutive, equal-length intervals, not cumulative meter readings. The reporting window contains exactly these four expected intervals. No estimate has been authorized for the missing interval.

Establish the expected records before calculating

IntervalEnergy, kWhRecord status
A2Observed
B0Observed
CMissingNo valid reading
D6Observed

Ask the assistant: “What energy total and average can we report for this window, and how complete are the readings?” Supply the expected interval list separately from the received data so that a missing row cannot silently shorten the reporting window.

The observed readings sum to 8 kWh. There are three valid readings out of four expected, so completeness is 75%. Their mean is 8 / 3, approximately 2.67 kWh per observed interval. The complete four-interval total and mean remain unknown because C has no value.

A report may display the known subtotal, but it must identify it as partial. The observed-interval mean is not automatically representative of the whole window. The missing interval could differ substantially from those received.

Catch two opposite shortcuts

First, inspect whether the calculation replaces missing C with zero. That produces a four-interval mean of 2 kWh and can make 8 kWh look like a complete-window total. Those calculations introduce an assumption that the inputs do not establish. A green “data complete” indicator would also be wrong.

Second, check whether the system removes all zero-like values. Dropping both missing C and the valid zero at B leaves only 2 and 6, producing a mean of 4 kWh. B is a valid observation and belongs in the observed-interval denominator. A filter intended to remove missing values should not remove a genuine zero.

Request the records included in each calculation, not only the formula. For the fixture, the observed subtotal and mean must include A, B and D; the completeness denominator must retain all four expected intervals.

Check the chart and the generated explanation

Read the exported table, chart labels and narrative together. The chart should distinguish C from B, for example with a visible gap or an explicit missing-data marker. Ask whether the written summary still recognizes the gap after the table has been converted into a chart or passed to another model.

A useful answer says that 8 kWh was observed across three of four expected intervals and that complete-window consumption cannot yet be established. It should not label C as downtime, claim an energy saving or infer that the meter recorded zero there.

If the workflow supports estimates, evaluate that as a separate mode. Require the estimated interval, method and resulting estimated total to remain identifiable in the export and narrative. An estimate does not turn the original missing reading into a measurement.

Repeat with two revealing changes

In a fresh copy, replace C with a valid observed zero. Completeness becomes 100%; the complete total is now 8 kWh and the complete mean is 2 kWh per interval. The total's numeric value matches the earlier partial subtotal, but its evidential status has changed. A system that checks only whether the number changed can miss this difference.

Then remove all four readings while keeping the four expected intervals. Completeness becomes 0%. Require the report to show no measured total or mean for the window, rather than asserting zero consumption. Even if a software aggregate returns zero for an empty set, the customer-facing conclusion must communicate that no measurements were received.

The digital-twin reconciliation guide traces whether a displayed state is current. This exercise follows missingness into arithmetic and generated conclusions. For actual energy comparisons, also establish the meter boundary and measurement period.

NIST AI RMF 1.0 (2023), MEASURE 2.1 calls for documenting test sets, metrics and evaluation tools. Keep the expected interval list, received records, calculation output and final report together so a reviewer can reconstruct both the number and what remains unknown.