# Synthetic methods review

Reviewer role: independent synthetic statistical and measurement reviewer. This is an AI-generated review, not external peer review or endorsement by a named scientist. Reviewed `BRIEF.md`, manifest metadata and the proposed estimator on 4 September 2026. No daily records, implementation or empirical results were inspected.

The design is suitable for a descriptive historical calculator after the issues below are resolved. Exact-age episode counts, missing-history resets, station-specific results, fixed thresholds, sample suppression and resampling years are sensible choices. The review cannot certify data quality or numerical correctness without the resulting audits and checks.

## P1: resolve before public interpretation

1. **Compare deadlines using the same eligible episodes.** Requiring all future days to be observed prevents accepting early successes while discarding incomplete failures. But separate completeness rules for 1, 7, 14 and 28 days create different denominators. A displayed success rate can then fall when the deadline grows, even though the underlying events are nested. For the interactive deadline comparison, require the full 28-day window for every horizon and retain identical anchors. Horizon-specific estimates remain defensible if shown separately with an explicit sample-change explanation. Check that counts of successes are nondecreasing across horizons and denominators are identical in the main comparison.

2. **A monthly benchmark does not fully control calendar composition.** Long dry spells can cluster in particular years and within-month dates. Weighting a 65-year monthly baseline by anchor months still compares, for example, drought-year anchors with wetter years. Add a same-station, same-year-and-month benchmark as a sensitivity check, using the same future completeness rule. A cheap day-of-year sensitivity, such as nearby calendar dates across years within the selected period, can assess remaining within-month seasonality if the central claim depends on a small delta. Describe the original baseline as month-adjusted, not fully season-adjusted. Changing the duration also changes which episodes survived and their calendar distribution; a higher plotted point is not an intervention effect of waiting.

3. **Precipitation is not necessarily rain, and 5 mm is not a rainless-day definition.** NOAA identifies PRCP as daily precipitation in tenths of a millimetre. The manifest includes mountain stations at 1,897 and 2,504 m. Winter precipitation outcomes must not be described exclusively as rainfall. Use precipitation in numerical labels, with the rain shorthand explained where needed. A day at or below 5 mm can have rain, so use "days without precipitation above 5 mm" or an equally precise definition. A 5 mm threshold also does not diagnose agricultural drought. Keep the exact greater-than rule visible. See the [NOAA format specification](https://www.ncei.noaa.gov/pub/data/ghcn/daily/readme.txt).

4. **State the question this analysis actually addresses.** The paper's abstract concerns increasing daily rainfall hazard, historical prayer timing in Murcia and cross-cultural ritual prevalence. A growing cumulative probability with a longer deadline follows from nested outcomes even under a constant or declining daily hazard. It therefore provides no evidence, by itself, for an increasing daily hazard. Keep the daily `w=1` curve central, show uncertainty and avoid declaring its slope from two selected points. The proposed grid is an exploratory Romanian meteorological analogue. It neither tests Romanian prayer timing nor reproduces the paper's full estimator or ritual claims. The supplied post image was unavailable to this reviewer, so fidelity to its wording remains unverified. See the [authors' university abstract](https://economics.yale.edu/research/praying-rain).

## P2: resolve in extraction and uncertainty checks

5. **Define the calendar and episode boundary exactly.** For anchor date `t` and age `d`, require an observed qualifying wet day at `t-d`, followed by exactly `d` observed days at or below the threshold through `t`. Outcomes occupy `t+1` through `t+w`. At `d=0`, the anchor itself is a wet day; label this accordingly. There is one anchor per episode for each selected age, not one anchor shared across all ages. Consecutive wet days require an explicit zero-age convention. Missing or rejected observations make the age unknown until the next observed qualifying wet day. Do not restart spells at New Year, join separated seasons, interpolate absent dates or treat invalid day-31 slots as missing real days. Filter seasons by anchor month only and say that the outcome can enter another season.

6. **Fix period-boundary policy before counting.** Decide whether the named period limits anchor dates alone or the full observation window. Both policies can be coherent; mixing them cannot. Apply the same rule to the benchmark, preserve continuous dates across ordinary year boundaries and document excluded edge anchors. A retrospective calculation that uses later observations is not a forecast validation. Never obtain a spell's age from its eventual total length, and never discard an episode solely because its eventual terminating rain falls beyond the analysis boundary when the requested outcome window is already observed.

7. **Retain source flags and audit measurement changes.** Excluding nonblank QFLAG, missing values and MFLAG `P` is appropriate. Retain source and measurement-flag frequencies by station and year. NOAA warns specifically about precipitation from source `S`; inspect whether it enters eligible stations or replaces another source. `B` and `D` identify subdaily totals assembled into a day, while `T` is a trace; these are not automatically invalid. Do not fill daily gaps from multiday totals. A blank quality flag is not proof that instruments, station location or observation time stayed constant. Use source-stable or split-period sensitivity where a transition affects a headline. See [NOAA source descriptions](https://www.ncei.noaa.gov/products/land-based-station/global-historical-climatology-network-daily).

8. **Report selection after filtering.** Ninety percent overall coverage can conceal missing wet seasons or gaps that erase long dry spells. Publish usable-day coverage by year, the number of known-age candidates, losses to incomplete future windows and accepted anchors by age and horizon. The complete-window rule removes a specific outcome-dependent inclusion error; it does not establish that missing observations are unrelated to precipitation. Do not equate manifest first/last years with uninterrupted daily coverage. The manifest distinguishes București Filaret from Băneasa; preserve full station names and IDs. Check shared identifiers and nearby coordinates before any future pooling, without merging separate stations merely because they serve the same city.

9. **Resample intact annual contributions, not artificial daily sequences.** Extract real chronological anchors and outcomes first. Resample their anchor-year contributions with common year multiplicities for probabilities, benchmarks, deltas and every compared duration/horizon. Do not concatenate sampled years and detect new spells at the joins. Keep original outcomes when they cross New Year. Ordinary year blocks retain within-year dependence but miss dependence across adjacent years; this limitation matters most for rare long spells. The 30-anchor/10-year rule is a minimum display rule, not assurance of stable tails. Show anchor years and counts. If every resampled estimate is 0 or 1, a collapsed percentile range does not establish certainty; label the uncertainty as unresolved by resampling. State that the 1,000-draw intervals are approximate pointwise intervals, not a simultaneous band supporting a curve-wide claim.

## Minimum release evidence

Add focused fixtures for threshold equality, a trace, a missing/presumed-zero day within history, an early success followed by a missing future day, a year-crossing spell, a leap day and a right-censored episode. Check common-cohort nesting, benchmark weighting, deterministic resampling and low-sample suppression. Publish an issue-resolution note linking each material change or limitation to its final implementation or wording. No extra hazard model is necessary for the scoped historical calculator.