# Pre-registration: vrancea-nu-are-ceas

Written 9 Nov 2026, ~07:25 local, before any statistic was computed on ROMPLUS.

**What was already seen:**
- The M ≥ 5 USGS numbers (t14_ceas) and the INFP M ≥ 3.5 numbers were opened by the 20-brief agent earlier today. They are reported as **seen before pre-registration**.
- The ROMPLUS list of Vrancea Mw ≥ 7 dates was printed once while checking the parser. No gap, CV, test or fit was computed.
- On the people side: the yearly top days of the ro.wiki *Cutremurul din 1977* page were printed once. 4 March is the top day in most years. T5a below is therefore **seen**. T5b is not.

The claim tested: *"the big Vrancea earthquake is overdue / comes every ~40 years"*. The big one means a quake like 1940, 1977 or 1986, so Mw ≥ 7.

## Data
- **ROMPLUS** (INFP, `https://www.infp.ro/data/romplus.txt`, downloaded 9 Nov 2026, AD 984 – Feb 2026).
  - Vrancea box: lat 45.2–46.1, lon 25.9–27.0, depth ≥ 50 km. Historical rows have a nominal depth of 150 and are kept.
- **USGS ComCat**, same box, M ≥ 5, from 1973. INFP FDSN is used for anything after Feb 2026.
- **A second historical catalogue** (SHEEC 1000–1899 / EMEC 1000–2006) is used to check the list of Mw ≥ 7 dates. If it drops or adds an event, the tests are rerun on its list and both results are printed.
- **ro.wikipedia pageviews** (user agent, daily, Jul 2015 – Nov 2026): *Cutremurul din 1977* and *Cutremur*.
  - For the normalising series: the total of all ro.wiki user views (project aggregate), daily.

## Fixed rules
- **Merge:** events closer than 30 days count as one episode, dated by the strongest.
- **Primary threshold:** ROMPLUS Mw ≥ 7.0. Sensitivity runs at Mw ≥ 6.9 (adds 30 May 1990) and Mw ≥ 7.2.
- **Primary window:** 1700 → 9 Nov 2026. Sensitivity windows start in 1600 and 1800.
  - Reason: before ~1700 ROMPLUS has placeholder 1 January dates and gaps of over 100 years, which means missing records, not quiet.
  - If the second catalogue or the ROMPLUS paper states a different completeness year for Mw ≥ 7, that year becomes a further sensitivity run. It does not replace 1700.
- **Open gap:** the open gap from the last Mw ≥ 7 episode in the list to 9 Nov 2026 is reported as censored. It is never treated as a closed gap.

## Tests

**T1. M ≥ 5, 1973–2026 (USGS)** (*seen*). Reported:
- coefficient of variation (CV) of the gaps;
- Kolmogorov–Smirnov (KS) test against the exponential;
- the share of past gaps longer than the gap open today.

**T1-power.** Simulate 10,000 renewal sequences with n = observed, with gap CV 0.3 and 0.5 (gamma gaps).
- Report the share of sequences in which the test rejects randomness, using a one-sided rule: observed CV below the 5th percentile of CV under Poisson with the same n.
- If power at CV 0.5 is ≥ 80%, the page may say *"a clock this regular would have shown"*.

**T2. Mw ≥ 7 since 1700 (ROMPLUS)** (*primary, not yet computed*).
- (a) CV of the closed gaps, with a p-value from 100,000 Poisson simulations with the same n. The p-value is the share of simulations with CV ≤ observed.
- (b) Power: the same power check as T1-power with this n.
- (c) The rank of today's open gap among the closed gaps: the share of closed gaps that are longer.
- **Reading:**
  - "Clock" is supported if p < 0.05.
  - "No sign of a clock" if p ≥ 0.05.
  - If power at CV 0.3 is < 80% as well, the page must say a mild clock cannot be ruled out with this many quakes.
  - **The page prints whatever comes out.**

**T3. Held-out forecast (the page's own error).**
- Fit both models on the Mw ≥ 7 episodes 1700–1899 only:
  - **No clock:** a Poisson model with a constant yearly probability of 1 − exp(−1/mean gap).
  - **Clock:** a Brownian Passage Time renewal model (mean and aperiodicity fitted by maximum likelihood on the 1700–1899 gaps).
- For every year 1900–2025, each model gives the probability of an Mw ≥ 7 episode that year, given the time elapsed since the last one.
- Score both by Brier score and log score against what happened.
- **Reading:**
  - "The clock forecast did better" only if its log score beats Poisson's.
  - Both scores are printed as the method's error on years it did not see.
- Sensitivity: refit with Mw ≥ 6.9.

**T4. Earth calendar.** Day-of-year test on Vrancea quakes, ROMPLUS 1940–2026, Mw ≥ 4.0.
- Declustered with Gardner–Knopoff (the shared `decluster` function).
- Rayleigh test on the year-angle, plus a χ² test on the 12 months.
- Expectation: no calendar (p > 0.05). Either result is printed.

**T5. People's calendar and the "overdue" feeling.**
- **(a)** *(seen)* The share of years 2016–2026 in which 4 March is the top day of the year for *Cutremurul din 1977*. Also reported: the same with years that had a Vrancea M ≥ 5 excluded.
- **(b)** *(not seen)* Does attention grow as the quiet stretch lengthens?
  - **Series:** weekly *Cutremur* views ÷ weekly total ro.wiki views.
  - **Excluded:**
    - weeks overlapping 1–7 March;
    - weeks within 30 days after any Vrancea M ≥ 4.5 (INFP) or any M ≥ 6 within 1,500 km felt news (6 Feb – 31 Mar 2023, the Turkey quake);
    - weeks overlapping 10–31 January (school-term spike seen in the raw series).
  - **Test:** within each gap between consecutive Vrancea M ≥ 5 (INFP/USGS) that lasts ≥ 180 days, the Spearman ρ of normalised attention against days since the last M ≥ 5. Also the pooled partial ρ with year dummies.
  - **"Overdue feeling" in the readership** if the median per-gap ρ > 0.2 and the pooled ρ > 0.2 with p < 0.05.
  - Otherwise: "attention fades after each quake and does not rise as the quiet lengthens".

## Third body of data (press)
- GDELT DOC 2.0 timeline (English-indexed: `earthquake` with `sourcecountry:RO`), 2017–2026, weekly.
- Same T5b test on the press series.
- If GDELT cannot be read, the alternative is chosen and named here before it is run.

## Not claimed whatever happens
- No probability for next year at any place.
- Nothing about magnitudes beyond the record.
- Small quakes do not "release pressure".