N.01 / EVIDENCE
What three months can prove
A long window can show that drift did not appear. It cannot prove bounded memory by itself. Notes on scope, counters and the limits of extrapolation.
A number on a home page is easily read as a conclusion. Three months and 68 MB of resident RSS look like “stable”, and “stable” is easily read as “bounded memory”. Neither step holds.
First, what was measured
In Sodama’s record, three months is a production health window and 68 MB is the resident RSS observed across it. They support a limited but useful judgment: with this build, this machine, this traffic mix and this configuration, no sustained memory growth or memory step was observed.
The same record contains a reload acknowledgement that fell from 19–34 seconds to 0.74 seconds. That number describes the panel confirmation on the validation → persistence → acknowledgement chain. It is not full request latency, and it is not an upper bound for every reload path.
Why the 25-hour window and 1,486 samples are gone
The earlier 25-hour window and sample count belong to an older set of observations. Keeping them next to the three-month window would imply that they came from the same test. Data can be updated, but the scope has to move with it; old numbers should not remain as decoration.
What a window cannot prove by itself
Three months without a memory step does not mean memory is bounded. It does not cover traffic mixes that did not occur, protocol paths that were not triggered, other kernel versions, extreme concurrency, or low-frequency events over a longer life. It says nothing about the next reload, the next upgrade or the next plugin.
“Stable” here is a description: within the recorded window, behavior did not show drift that needed explanation. It is not a permanent property, and it is not a label that can be attached to any version.
Evidence should arrive with its boundary
That is why the Sodama engineering note now lists its evidence scope separately: what was measured, which window it belongs to, which counters do not cover every failure mode, and which environment details are not public. This does not make the project look weaker. It tells the reader how far to trust it.
To turn this into a reproducible benchmark, at least the hardware, kernel version, commit SHA, full traffic profile and sample log are still needed. Until they exist, three months and 68 MB are the beginning of a checkable record, not the end of one.