Log Inactivation from Counts

This is the measurement the whole framework is trying to predict: count the organisms going in, count them coming out, and the base-10 logarithm of the ratio is the log reduction. Two hundred and fifty thousand per 100 mL entering, twenty-five leaving, and the ratio is ten thousand — four logs, flat.

Logs rather than percentages, and the reason is that percentages crowd together where the interesting behaviour is. Ninety-nine percent and 99.99% differ by two characters on the page and by a factor of a hundred in survivors, and no one's intuition handles that gap. The log scale spreads it out, and — more usefully — logs add across barriers. Two-log removal in the filters followed by two-log inactivation in the contact chamber is four logs overall, and that additivity is the entire architecture of multi-barrier treatment. Percentages do not add: the same pair is 99% then 99%, which is 99.99% overall, not 198% and not 99%.

A non-detect is not a zero. This is the mistake that inflates more log-reduction claims than any other, and it is easy to make in good faith. A plate that grows nothing does not mean the count was zero; it means the count was below the detection limit. Divide by zero and the log reduction is infinite, which is a number no disinfection process has ever achieved. The honest substitution is the detection limit itself, and what you get is the minimum log reduction your data can support — say "at least 4.4 logs", not "complete inactivation". If you need to demonstrate more, you need a bigger sample volume or a more sensitive method, not a different arithmetic.

The counts themselves are noisier than they look. Microbial counts are Poisson-distributed, which means a plate reading of 4 carries a standard deviation of 2 — the uncertainty is the square root of the count. Any log reduction built on small numbers inherits that scatter in full, and the difference between 3.6 and 4.1 logs on a single pair of plates is frequently nothing at all. This is why performance is demonstrated over many samples and many days rather than announced from one good result, and why a seeded challenge study uses enormous spikes: a starting count of $10^7$ leaves room to measure six logs of removal with counts still large enough to be stable at both ends.

And the reduction you measured is not the reduction you are credited. Regulatory credit for inactivation comes from a demonstrated CT under approved conditions, and credit for removal comes from an approved and properly operated process — not from a favourable pair of samples. The measurement is how you find out whether your process is doing what it is credited for. It is monitoring, and it is valuable precisely as monitoring: a plant whose measured reduction drifts down over a season has a real problem, whatever its paperwork says.

One last framing that surprises people. When the arithmetic returns a survivor count below one — say 0.02 organisms per 100 mL — that is not an error and it is not zero. Read it as a probability: roughly a 2% chance that any given 100 mL sample contains an organism. A utility producing many millions of such volumes a day is still delivering a real number of them, which is why drinking-water targets are framed as an acceptable annual risk of infection rather than as an absence, and why the log framework does not stop at "none detected".