Codex Resonance · Research Note · v0.2
“The average restoration time fell from ten hours to eight.”
That is useful information. It does not tell us whether every customer waited less.
Consider two entirely synthetic groups of four service incidents. Each incident took ten hours to resolve in the first period. In the second period, the groups look like this:
Incident | First period, both groups | Second period, Group A | Second period, Group B |
1 | 10 hours | 8 hours | 2 hours |
2 | 10 hours | 8 hours | 2 hours |
3 | 10 hours | 8 hours | 14 hours |
4 | 10 hours | 8 hours | 14 hours |
Average | 10 hours | 8 hours | 8 hours |
Both averages fall by two hours. In Group A, every incident takes two hours less. In Group B, two take eight hours less and two take four hours more.
The average is correct in both cases. The underlying experience is different.
What the average leaves out
The mean retains the total divided by the number of incidents. In this example, both second-period totals are 32 hours across four incidents.
It does not retain how those hours are distributed. The same average is compatible with different answers to “Did every incident become faster?” In Group A the answer is yes. In Group B it is no.
This is a familiar property of aggregation. The example demonstrates it directly; it is not a new statistical theorem or an analysis of real service records.
More detail can help distinguish the groups. Their ranges and proportions above ten hours differ. A median alone would not resolve this example: both second-period medians are also eight hours.
The additional evidence should follow the question. If we want to know how many customers exceeded a service target, we need information about those incidents. If we want to understand repeat problems, we need identity and history. If we want to explain the change, we need evidence about possible causes.
Even individual records do not establish causality on their own.
Keep the useful summary
For groups of the same size, the mean supports comparison of average duration and, here, total restoration hours. It remains useful for those questions.
If the question concerns variation, individual experience or the longest waits, the mean alone leaves relevant distinctions unresolved. An improved average can coexist with some longer waits. It does not prove that this happened in any real operation.
This matters when a summary travels. A dashboard may provide the average to a manager. An AI assistant may reuse it in a customer report. A later reader may interpret “average restoration time improved” as “customers are now restored faster”.
That last statement needs care. Does it mean a lower average across the group, a change for most customers or a prediction about the next customer? The evidence needed differs.
Ask the next question explicitly
Before reusing a summary, ask:
- Which group and period does it describe?
- What question does the summary answer?
- Which distinctions were removed?
- What additional evidence would support the conclusion now being proposed?
Codex Resonance examines these boundaries between evidence and inference. In this case, two different underlying patterns share the same summary. A system receiving only that summary cannot tell which pattern it came from.
Recognising the limit gives us a practical next step: retain the conclusion the evidence supports, and request the information needed for the question that remains.
Synthetic illustration only. The incidents are deliberately constructed and do not describe customer results. This note demonstrates an information limit; it does not establish service performance, treatment effects or a new statistical result.
© 2026 Codex Resonance. Operated by Arqua Pty Ltd.