Mensura

The instrument

Mensura is a private, single-user tracker for recording substance consumption hour by hour and reviewing patterns as trends, a heatmap and a forecast. Each row represents a substance; each cell records an hour and amount. Explicitly logged zeroes can mark that use stopped, while days with no information are not silently treated as confirmed abstinence.

How the forecast learns

There is no population model or cloud training job. Mensura builds a separate fit for each substance from the user’s own recent history, using up to 180 calendar days of hourly records. The training window includes other tracked substances too: their activity can help reveal that one substance tends to follow another. Unlogged gaps carry the timeline forward but are not counted as evidence of no use.

The forecast has two parts. First, a regularised logistic model estimates the probability of use in each hour. A challenger based on gradient-boosted trees is trained alongside it so the model can capture interactions, such as a time-of-day pattern that changes on weekends. Both compete on expanding, held-out time blocks; the challenger replaces the simpler model only when validation supports the change. Probabilities are calibrated out of sample. Second, ridge regression estimates the amount, conditional on use, by modelling log(1 + amount).

Features are built only from information available before the hour being predicted: clock and weekday patterns; several time scales of recency after the person’s own previous logs; whether use has already happened that day; amount accumulated so far; recent overall rate; active-use duration; and activity in other tracked substances. Recent days carry more weight, with the recency half-life selected from candidates using out-of-sample validation. A detected change in the pattern can shorten that memory so older behaviour does not overwhelm a new trend.

From a fit to an outlook

Mensura uses Monte Carlo simulation to project the remaining hours of a day 400 times. A simulated use feeds back into later hours in that same simulated day, preserving bursts instead of averaging them away. The paths produce an expected daily total, a next-hour probability and an hourly risk profile. If the log is too sparse to support the fitted model, the app falls back to an hour-of-day profile rather than presenting an unsupported level of precision.

Boundaries and privacy

Entries and model fits stay in the browser’s localStorage; the app does not send consumption logs to a server. An optional access key encrypts local data and automatic backups in the browser. Page-view analytics are separate and do not receive log contents. The forecast is a personal heuristic for reflection, not a diagnosis, treatment recommendation or clinical predictor.