What the review changed

This dossier preserves the first edition's synthetic scientific review. The figures and verification counts below refer to the original Băneasa example, April–September, after 14 days with at most 5 mm per day. The revised edition opens with Constanța, June–August, after 30 days. The current method explains the new example and its limits.

Verification of the revised edition

June–August was calculated as a separate selection with unchanged observation and display rules. The Constanța example after 30 days with at most 5 mm per day contains 64 episodes across 51 years: 12 threshold exceedances within a week and 28 within two weeks. Băneasa's 28 episodes at the same setting fall below the display threshold.

Independent verification now covers all 1,296 outputs for each of the two stations, 2,592 in total. It also reproduced the new and former examples' intervals and all 64 records in Constanța's dated history. All 10 production-function tests and general checks passed across 31,104 outputs in 864 configurations. The 648 former configurations remain identical.

At Constanța, excluding the 30 source-S days in June 2024 changes the seven-day frequency from 12/64 = 18.75% to 11/62 = 17.74%. That sensitivity reuses production functions. The verification evidence and source-S sensitivity describe the revised edition. The following sections preserve the first edition's objections and figures.

First-edition synthetic scientific review

Synthetic review completed on 4 September 2026. Two AI agents separately examined statistical methods and anthropological interpretation. The statistical reviewer subsequently checked the calculations with a separate implementation. No scientists participated; the cited study's authors have not reviewed this project.

Verdict: the calculator is suitable for publication as a descriptive analysis of historical precipitation. Available evidence cannot estimate ritual effects or confirm the social mechanism in Romania.

1. One denominator for every deadline

The initial objection was that separate completeness rules could compare different samples. Every episode now requires 28 fully observed subsequent days, even for the next-day estimate. At Băneasa, all four deadlines use the same 277 episodes. Checks across the published output confirm identical denominators and success counts that never decrease as deadlines grow.

2. The reference changes the interpretation

After 14 qualifying days, the seven-day observed frequency is 50.2%. The same-calendar-month reference is 49.8%; the +0.4 percentage-point difference has an interval of −3.9 to +4.6. The same-month-and-year reference is 42.2%; the +7.9-point difference has an interval of +3.3 to +12.2. Both are descriptive comparisons. Neither isolates a causal effect of waiting or prayer.

The daily curve is separate from cumulative probabilities. The raw next-day difference between durations of 30 and seven days is +3.0 points, with an interval of −4.0 to +10.7. It does not establish an increasing curve throughout. Comparisons across seasons, periods and thresholds are exploratory, with pointwise intervals.

3. Numerical verification

A second implementation directly parsed Băneasa's daily source file without importing production functions. It reproduced all 972 outputs: denominators, successes, anchor years, excluded windows, probabilities, both references and both differences. It also reproduced the initial example's intervals using 1,000 annual resampling draws.

Nine production-function tests cover threshold equality, missing observations, unknown history, incomplete windows, year crossings, leap years, period boundaries and other edge cases. General checks cover 23,328 outputs in 648 configurations from 30 stations. These general checks test numerical invariants; the full independent recount applies to Băneasa.

4. Observation provenance matters

The audit retains measurement and source flags by station and year. A station identifier does not guarantee one source on every date. Băneasa's 30 days in June 2024 have source flag S, identifying Global Summary of the Day. NOAA advises caution with precipitation from this source.

We repeated the initial example with those days treated as missing, without replacement. The sample falls from 277 to 275 episodes, and the seven-day result changes from 139/277 = 50.18% to 139/275 = 50.55%. Its interval becomes 45.62%–55.44%. The largest change across the four deadlines is 0.52 percentage points. The interpretation is unchanged. This sensitivity uses production functions and is separate from the independent recount.

The public RoClimHom file was excluded: the downloaded structure contains months, although the article describes daily source observations. Turning monthly totals into daily spells would invent information.

5. Anthropological limits

The review separately checked the article, supplement and Romanian documentation. The 44% and 30% figures are raw means from supplement Table D11. The global association, classification at the 99th percentile and cultural-group comparison have different units from a station-based calculator.

Caloian and Paparuda have ethnographic attestations; the repertoire's 716 texts are not 716 communities. Contemporary processions are a different kind of documentation. The map locates weather stations without measuring ritual distribution. Event sequence does not establish beliefs, intentions, forecast use or deception.

What remains open

The series have not undergone additional homogenisation, and instrument or station-move histories were not checked. Missing observations may be selective. Annual blocks leave some dependence between years untreated. Eligible stations are not a representative sample of Romania's population. This analysis contains no systematic sample of dated and located rituals; the published protocol describes future research and is not preregistered.

Evidence and reproducibility

The original reports are archived separately: 972 independently recounted outputs, first-edition tests, Băneasa sensitivity excluding source S, and the flag audit. They document the first-edition figures above.

A synthetic review can identify errors and demand useful checks. It cannot replace external peer review or guarantee that every error has been found.