Method and limits

Sources

National Institute of Korean History (Ministry of Education), 'Veritable Records of the Joseon Dynasty — original text' (data.go.kr dataset 15053647, KOGL Type 1): 25 reigns Taejo–Cheoljong, 25 principal editions plus 2 revised, 1 supplemented, 1 supplement-and-corrections edition and the two drafts of the Daily Records of Gwanghaegun; 676 XML files. The Gojong and Sunjong records come from 'Gojong–Sunjong original text' (15053646, KOGL Type 4: no derivatives, non-commercial) and are used only for counts that neither alter nor quote the text (articles, characters, classification, keyword hits). Counted at the article level (biblioData type="T"). The range is therefore Taejo to Sunjong, 1392–1910; the Gojong and Sunjong records appear in the article table without titles (ID, date, character count, classification and link only) and without context excerpts in the keyword table. The sky series (C) covers Taejo–Cheoljong only. Sample coding does include articles from the two records, but in the published coding tables their title, quote and summary are blanked. Reign length is measured as the JDN span between the first and last dated articles, the same rule before and after the 1896 calendar switch.

NIKH, 'Veritable Records — supplementary person data' (15053645): 115,859 persons, 336,255 office-holding rows. Academy of Korean Studies, 'Successful candidates of the civil service examination' (15052752): 15,151 examination passers with name, clan seat and year.

Counting unit

Everything is counted per article: one biblioData type="T" element in the NIKH XML. Of 414,501 articles, 412,769 carry a date. Time series use the principal editions only (the final draft for Gwanghaegun); revised and supplemented editions appear only in the editions series. Characters are the Chinese characters of the article text.

Dates

Sillok dates are lunar. The CE year and lunar month/day are kept as given; solar dates come from korean_lunar_calendar, shown in the Julian calendar before 15 October 1582. From 1 January 1896 (Geonyang 1) the court used the Gregorian calendar and the NIKH XML marks those days 陽曆, so their month and day are taken as given rather than converted (`date_system` column). The conversion was checked astronomically: the 6,642 lunar month-starts of 1392–1928 were compared with DE406 new moons in Seoul local mean time (UT1 + 126.978° E; ΔT from the Stephenson–Morrison–Hohenkerk 2016 model) — 6,463 (97.3%) fall on the same day, 179 one day off, none two or more. Eclipse matching therefore uses a ±1-day window, with the same-day value shown alongside. Of the dated articles, 33 (26 for Taejo–Cheoljong) have no solar_date because the conversion package rejected the day: the 30th of a month to which the package assigns 29 days, 2 on a 31st, and 1 with an unknown day (99). A 1.2.4 check against Ahn et al. 2000 (JKAS 33: 177–183, Tables 1–2) and the KASI lunar–solar converter showed that the package's month lengths agree with the KASI calendar database; it is the Sillok that begins those months a day apart. Because every article carries its ganji (sexagenary day), the table now has a ganji-anchored column, solar_date_ganji: in 23 months where the article ganji is one day off the converted day throughout the month (1,139 articles; 14 of these months appear in Ahn et al. 2000, Tables 1–2) the ganji is followed (date_flag = ganji_month_shift); 31 of the 33 refused articles are dated by ganji (ganji_only); the 2 of 30 January 1874, where the NIKH ganji itself is off for the whole month, stay empty (ganji_unresolved). Scattered disagreements not explained by a month shift (1,306 articles) keep solar_date (ganji_typo). Metrics still use solar_date, as in 1.0; the ±1-day eclipse window covers the shift. A further 38 articles whose day reads '??' are counted as undated.

Astronomy

JPL DE406 planetary ephemeris read with skyfield. New moons, solar and lunar eclipses and inferior conjunctions of Venus were computed directly; ΔT from the Stephenson–Morrison–Hohenkerk 2016 long-term model. For every new moon the apparent discs of Sun and Moon were tested for overlap above the Seoul horizon (37.57° N, 126.98° E) at 2-minute steps. Apparent radii use 696,000 km for the Sun and 1,737.4 km for the Moon (with 695,700 km one grazing contact on 1528-11-22 drops out, giving 184); the `newmoon_jd` and `jd` columns are TT Julian dates. The 185 eclipses visible from Seoul in 1392–1863 so obtained all appear in the Espenak–Meeus canon, and all 1,125 of its eclipses match a new moon within ±1 day. Lunar eclipses come from skyfield's eclipse finder; visibility is judged by the Moon's altitude at maximum eclipse. Times of day were not compared in this version because of the timekeeping systems of the period.

Keywords and sample coding

Many metrics in series C, D and E select articles by Chinese-character keywords. Keywords also match passages where the characters mean something else ('日食一升' is a daily ration of rice), so for each keyword metric a stratified sample of hits was coded, and its precision with a Wilson 95% interval sits on the reliability card. All 332 eclipse articles were classified by type, and 34 were blind re-coded by another coder. Coding followed a codebook and was done by several language-model coders; every quotation was string-checked against the source. Codebooks and all judgments are published.

Reliability cards

Three kinds. Exact counts are read straight from the source and checked against external figures (Bak & Oh 2015; an independent 2026-09 index). External matches are rates of agreement with astronomical computation. Sampled precision is the share of keyword hits that mean what the metric says; multiply keyword counts by it.

Limits

• The Sillok records what was written down, not what happened. Frequency metrics (omens, death-penalty mentions, relief) reflect recording practice — the Sejong peak in death-penalty mentions turned out to be sentencing paperwork.
• NIKH's classification, comment tags and person links are modern editorial work of varying granularity. Person appearance counts are not used as metrics because of linking errors.
• The roster fill for clan concentration is 91% accurate, and after 1850 the clan of 20% of appointees remains unidentified.
• The series C lunar match rate (C2) is a lower bound without type classification; G2 in 1.2 gives the values by type.

1.2 — three sources and the sky

National Institute of Korean History, 'Korean History Database — Goryeosa original text' (data.go.kr 15053637, no restriction). All 4,568 articles of the Treatise on Astronomy (fascicles 47–49). NIKH's lunar date, Julian solar date and day ganji are used as given and cross-checked. National Institute of Korean History, 'Goryeosa-jeolyo original text' (data.go.kr, KOGL Type 1). 249 articles hit by the astronomy keywords; dates are monthly, and days restored from the ganji in the first clause are flagged in `date_flag`. National Institute of Korean History, 'Seungjeongwon Ilgi' (data.go.kr 15064218, KOGL Type 1): 297 XML volumes, 104,074 day nodes, 1,792,433 articles. Every weather cell and the 8,520 astronomy-keyword articles are used. The new articles sit in a separate table (articles_ext) so that the 1.1 Sillok counts are unchanged.

Sites and dates. Gaeseong (37.97° N 126.55° E) for the Goryeo sources, Seoul for Joseon. NIKH's Goryeosa solar dates are Julian; re-converting the 4,436 articles after 1000 with the calendar library reproduces 4,435, and the one exception (the eclipse of Gojong 36, 4th month, day 1) is a year typo in the NIKH solar date (1246→1249), corrected. The Ilgi stays lunar to the end; of its 2,747 well-formed '陽曆 m月 d日' notes after 1896 (12 irregular ones excluded), 2,743 agree with the conversion. Day ganji agree with the converted date on 103,870 days; three whole months (1699-2, 1709-11, 1751-11) are shifted by one day — the contemporary almanac started those months a day away from the modern table — and for them the ganji-following date (solar_date_ganji) is used. ΔT before 1392 is a model value with a standard error of about one minute around the year 1000 (Morrison & Stephenson 2004): no effect on day-level matching, but borderline visibility calls can move.

Coding. Solar codebook v2 fixed, before re-coding, the four boundary cases that split the 1.0 review (below-horizon computation → forecast; reporting orders → forecast; ritual preparation → ritual; undated calendrical method → discussion). The lunar codebook uses the same frame plus an after-midnight flag. The slaves keyword was re-coded with a v2 rule naming four contexts (punishment, interrogation, transfer, labor duty) as yes: κ 0.86 on 30 blind articles (v1 0.44), precision 0.68 (v1 0.54). Astronomy keywords and the weather token rule were registered before computing and not adjusted afterward. Coders were several language models (15 batches plus 3 blind); every quotation was string-checked against the source (check_coding.py). In 1.2.1 the blind verification found a systematic split between coder batches — the reply dispatches sent to the Qing Board of Rites after an eclipse had passed ('日食已過, 禮部回咨') were classed as forecast (reporting order) by some batches and as discussion by others — and an editing rule now places them under discussion (15 solar, 85 lunar articles; the coder's type is kept in type_coder and the coder files are untouched). The blind re-coding pairs (blind_recoding_pairs) and the judgments of the 1.0 precision samples (precision_sillok_samples) are published as well. 1.2.2: the second blind round found articles of the same kind that escaped the rule because their quote lacks 回咨; the ids, read in the original, were added to the rule (22 solar and 92 lunar articles now). Abstinence (齋戒) articles that coder batches split between ritual and forecast were unified under ritual (21 solar, 3 lunar). The rules file (overrides.json) is published. Editing rules are used only for splits found by verification after coding, and each one is written at the end of the codebook. 1.2.3: the third blind round again found articles that the quote-based rules miss, so the full text of every forecast- or other-coded Ilgi article in both bundles was searched for 回咨, 齋戒 and 相値 and each hit was read: 20 lunar reply dispatches and the eclipse-day postponements (5 solar, 11 lunar) were added by article id (回咨 now 22 solar and 112 lunar; 齋戒 26 solar and 14 lunar). The boundary articles left under forecast are listed in the codebooks.

The weather cell of the Ilgi

Lyu, Z., Ichikawa, K., Cheng, Y., Hayakawa, H. & Kawamoto, Y. (2024), 'Digitization of weather records of Seungjeongwon Ilgi: A historical weather dynamics dataset of the Korean Peninsula in 1623–1910', Geoscience Data Journal 11(4): 504–513, doi:10.1002/gdj3.227; data: Zenodo doi:10.5281/zenodo.8382243 (CC BY 4.0). Daily weather strings 1623–1910 with Sunny/Cloudy/Rainy/Snow/Hail/Wind dummies. Lee J.-W., Shin H.-J., Hong J. (2024), 'Precipitation data in Seoul, Korea during 1778–1907', Scientific Data 11:566; data: Zenodo doi:10.5281/zenodo.11076768 (CC BY 4.0). Ilgi weather category and rain-gauge depth by day. Those two datasets digitized the cell first. This site adds the day-by-day agreement with them and the lists of differences, a routine-entry index and the structure of the gaps. Shares of the cell describe entries in a secretariat form, not Seoul's weather — the comparisons with the rain gauge and with modern observations show why.

Not done. The manual screen check against the KASI historical-astronomy database (registered procedure: no automated collection, only article number and match flag) did not make it into this version. Public numbers that changed are listed with their previous values under Versions.

Attribution added in 1.2

The Goryeosa, Goryeosa-jeolyo and Seungjeongwon Ilgi texts are released by the National Institute of Korean History on the Korean public data portal (Goryeosa: no restriction; Jeolyo and Ilgi: KOGL Type 1). Espenak F., Meeus J., Five Millennium Catalog of Lunar Eclipses (NASA/TP-2009-214173). Its 1,538 umbral eclipses of 918–1910 are used for cross-checking only (the catalog itself is not redistributed). NASA/JPL Small-Body Database — perihelion times of the 376 comets with perihelion before 1911 (queried 2026-11-02). Only the names and perihelion dates used for identification are shown. Korea Meteorological Administration, ASOS daily data service (data.go.kr, KOGL Type 1): Seoul (108), 1907-10-01 to 1910-12-31 — daily precipitation, mean cloud cover, phenomena. Weather comparison data are the Zenodo datasets of Lyu et al. 2024 and Lee, Shin & Hong 2024 (CC BY 4.0); the files are not redistributed, only agreement rates and difference lists.

Prior work

New here: the article-level table, precision for keyword metrics, eclipse match rates by article type, like-for-like comparison of edition pairs, and character-level deletion counts for the Daily Records of Gwanghaegun. Articles per reign (Bak & Oh 2015), Joseon eclipse records (Ahn & Lee 2004), early-Joseon historians' comments (Cha) and Sillok criminal cases (Yu 2007) already existed and are shown next to the recount on the Prior results page.

Attribution

This work uses the 'Veritable Records of the Joseon Dynasty — original text' and '— supplementary person data' created by the National Institute of Korean History (Ministry of Education) in 2022 and released under KOGL Type 1 on the Korean public data portal (https://www.data.go.kr/data/15053647/fileData.do · /15053645/fileData.do), and the 'Successful candidates of the civil service examination' released by the Academy of Korean Studies (https://www.data.go.kr/data/15052752/fileData.do). Astronomy: NASA/JPL DE406; eclipse canon: Espenak & Meeus (NASA/TP-2006-214141).

Versions

1.0 (2026-11-02) — first release: 24 metrics, 30 editions, 25 reigns.
1.1 (2026-11-02) — Gojong and Sunjong added to the counts (33 editions, 27 reigns); series C and sample coding unchanged (Taejo–Cheoljong).
1.1.1 (2026-11-02) — tie wording fixed in the reign summaries: the ten reigns with no historian's comment were each shown as a solitary "rank 18"; they now say how many reigns share the value. No number, table or data file changed.
1.2 (2026-11-02) — adds series G 'Record and sky — three sources' (6 metrics) and series F 'The weather cell of the Seungjeongwon Ilgi' (4), 34 metrics in all; 13,337 astronomy articles of the Goryeosa treatise, the Jeolyo and the Ilgi and the Ilgi's 104,074 days join as separate tables. Public numbers that changed: (1) Sillok solar-eclipse observation reports, re-coded with codebook v2, go from 84 articles / 65.5% (1.0, kept on C1) to 77 / 72.7% (G1); (2) lunar match rates are now given by type (G2) instead of the 1.0 lower bound without types; (3) slaves keyword precision 0.54 → 0.68 (v2 rule, series E card); (4) column n_articles renamed n_dated_articles in the D and E tables (values unchanged). The prior digitization of the weather cell (Lyu et al. 2024; Lee et al. 2024) is stated at the head of series F.
1.2, English revision (before release) — wording, terms, US spelling and sentence-case headers on the pages carried over from 1.1.1 and on the new series. Where the Korean and English differed, the side supported by the codebook or data was kept (a few Korean lines were corrected), and the slaves headline, takeaway and caveat now follow the 1.2 re-coding (precision 0.68). No number or data file changed.
1.2.1 (2026-11-02) — blind verification (165 checks) applied. Public numbers that changed: (1) G3's computation epochs were 12 hours later than the method text (local 'noon' was 00 h of the next day, 'that night 18–06 h' was 06–18 h of the next day). After the fix the Venus medians move by less than 0.15° (Goryeosa 43.1°→43.0°, Jeolyo 42.7°→42.8°, Sillok 43.6°, Ilgi 42.3°→42.2°; random 30.6°→30.4°, the other controls likewise) but the 犯 separations moved — Goryeosa 454 pairs, median that night 4.59°→2.77°, ±1-day window 2.62°→2.21°, ≤1.5° (±1 day) 32.8%→42.7%; Ilgi 156 pairs 6.69°→3.13°, ±1 day 3.03°→2.65°; random 90.9°→90.8°. (2) F4 'missing days' recounted as calendar days − distinct dates (2,275) instead of calendar days − nodes (889, including Yeongjo −715 and Gojong −5); duplicate nodes (1,317) and undated nodes (69) are now counted separately. (3) F2 'false rain 1.7%' withdrawn — 631 of those 632 days are trace-precipitation days with a gauge entry of 0 (98% carry a rain or snow token); 雨 appears on 1 of the 36,279 days without a gauge entry. Longest 晴 runs recounted over distinct dates and broken by calendar gaps: Hyeonjong 42→35, Heonjong 47→40, Cheoljong 102→86 (the other reigns' longest runs unchanged; mean runs changed). (4) 回咨 editing rule: 15 Ilgi solar articles (14 forecast, 1 observed), 84 Ilgi lunar (forecast) and 1 Sillok lunar (other) move to discussion — Ilgi solar observation reports 54/59 = 91.5% → 53/58 = 91.4%, Ilgi lunar forecast 207→123 and discussion 158→242, Ilgi solar forecast 83→69 and discussion 293→308, G6 Ilgi→Sillok solar 12/50 = 24.0% → 12/49 = 24.5%. Blind κ is kept on the coders' judgments (solar 0.90, lunar 0.78) with the after-rule values (0.91, 0.87) alongside. (5) G5 adds 50 Sillok articles counted with the same pattern, 170 in all (the '42 by the 1.0 keyword' in the 1.2 text was a different count); the G4 external note '322 guest-star articles in Seonjo 37' corrected to 323 in Seonjo's reign (137 in 1604–1605), with the file published. (6) G6's 1623–1863 cut set to the calendar dates themselves (1623-01-01 to 1863-12-31) — counts unchanged. (7) Display rounding made half-up — G4 Goryeosa chance mean 2.9%→3.0%. Wording corrected: sunspots 'single digits per 50 years' (contradicted the table), the generalization about lunar ritual match rates, 41 Goryeosa occultations (40 treatise + 1 Jeolyo), 16 star names (15 positions), Lee 2023 single author and 1,443 of 2,799 days, ΔT uncertainty (about one minute around 1000), the 'few kilometers' station distance removed, the F1 agreement ranges, the Lee comparison denominator 45,678, the definition of the 2,747 solar-date notes, the G6 copying inference softened, the Yeomans & Kiang 1981 reference. New public files: per-article Venus elongations and control draws, 犯 control draws, Sillok sunspot and guest-star keyword hits, blind re-coding pairs, judgments of the 1.0 precision samples.
1.2.2 (2026-11-02) — second blind verification (190 checks) applied. Public numbers that changed: (1) the 回咨 rule widened (explicit article ids, the spelling 回資) — 15→22 solar and 85→92 lunar articles under discussion; Ilgi solar observation reports 53/58 = 91.4% → 52/57 = 91.2%, forecast 69→42 and discussion 308→315, lunar forecast 123→113 and discussion 242→249. (2) New 齋戒 rule — forecast→ritual for 21 solar and 3 lunar articles (Ilgi solar ritual 146→167, lunar 327→330). After-rule blind κ 0.91/0.87 → 0.96/0.90 (coder-based 0.90/0.78 unchanged). G6 Ilgi→Sillok solar 12/49 → 12/49 = 24.5%. (3) The G4 Sillok observation-report sample now uses the published final judgment (overrides included) — 50/64 = 78.1% → 54/69 = 78.3% (chance 16.7%→16.7%). (4) The F2 gauge comparison counted per calendar date instead of node rows — measured-rain days 8,364 rows→8,320 dates, 雨 6,057→6,038 = 72.4%→72.6% (95% 71.6%–73.5%); zero-amount days 993→989, days without a gauge entry 36,279→36,138 (雨 1); 243 of Lee's 8,563 measured-rain days have no weather node in the diary and are excluded. Wording corrected: 0 mm is a trace per the paper's wording and not the same set as flag e; precip_any is rain, snow or hail; F4's '2,256 empty-cell days' are nodes (2,185 distinct dates); the Jeolyo's one 'not an eclipse' row is a solar-eclipse procedure article, not an occultation; the cause of G4's chance rate (window coverage of the period); G5's 50-year table sums to 168 (two undated articles); F3's 'had stopped being an observation' softened; the Halley reference is reference only; the 1.2.1 entry's 'Venus medians unchanged', 'others unchanged' and '0.57→0.58' corrected. New public file: overrides.json (the editing rules). Ten sample rows of the ephemeris were recomputed with Skyfield (AI check): all 10 reproduce the CSV.
1.2.3 (2026-11-02) — third blind verification (185 checks) applied. Public numbers that changed: (1) the 回咨 rule now searches the full text rather than the quote — 92→112 lunar articles under discussion (solar stays at 22); Ilgi lunar forecast 113→82 and discussion 249→269. (2) The 齋戒 rule now covers eclipse-day postponements — 21→26 solar and 3→14 lunar articles under ritual; Ilgi solar forecast 42→37 and ritual 167→172, lunar ritual 330→341. Observation-report match rates (solar 52/57 and the rest) and G6 are unchanged. After-rule blind κ for lunar 0.90→0.91 (solar 0.96; coder-based 0.90/0.78 unchanged). (3) Four dates that appear twice in the Lee et al. 2024 file (1785-01-01 to 03, 1873-01-01) are reduced to the row with a gauge value — F1 Lee comparison 47,244→47,240 and 45,678→45,674 days (agreement rates unchanged), F2 weather nodes 45,636→45,632 and dates with two nodes 189→185, Lee's days without a gauge entry 37,927→37,923 (all rates unchanged). The English notes in the rules file were revised (judgments; internal item numbers removed). No new public file.
1.2.4 (2026-11-03) — page numbers of the cited literature and the chronological-table check applied. Public numbers that changed: (1) on the prior-results page, the Bak & Oh (2015) parenthesis '380,271 with Gojong and Sunjong' had mislabeled the corpus total stated in the paper's text; Table 1 sums to 380,723 for all 27 reigns, and the text's 380,271 differs from that sum by 452, an inconsistency within the paper (the 25-reign sum 347,926 stands). (2) F1's external check 'Lyu et al. 2024, 102,689 days' is the row count of the Zenodo release; the paper states 100,620 days with weather records, and both are now given. Non-numeric changes: page and table numbers and full references (volume, issue, pages, DOI) in the external-check sentences; a note that Cha's Table 2 sums to 3,426 against the stated 3,435; and that Lee et al. 2024's '0 mm = trace' is the Table 1 (flag i) definition. The article table gains solar_date_ganji and date_flag (Method, 'Dates'): the 33 articles the conversion package refused were not package errors but months the Sillok began a day apart from the modern table, and 31 of them are now dated by their ganji. No metric value changed.
1.2.5 (2026-11-03) — review of the public site applied. Statement corrected: the reigns above 100,000 characters per reign-year are now listed as the data show — not 'Sejong, Seongjong, Jungjong, Seonjo and Sukjong' but eleven reigns (Sejong, Seongjong, Jungjong, Seonjo, Jeongjo and others, plus the one- to three-year reigns of Injong, Yejong, Munjong and Danjong); Sukjong has 72,246 (metric values unchanged). Display: the language link now opens the same page in the other language instead of its home page; Korean pages write year ranges with a wave dash; English legends and axis titles start with a capital; the Venus phase summary table has row names and rounding; negative numbers use − (U+2212) and '−0.00' shows as 0.00; '±1 d' throughout the English pages; an English home-page card summary no longer stops at 'et al.'; a few English sentences. No metric value or data file value changed (two explanatory notes in overrides.json were reworded).
1.2.6 (2026-11-03) — navigation: breadcrumb, a list of the other pages in the same series (sidebar on wide screens, collapsed box above the title on narrow ones) and previous/next links (arrow keys) on metric and reign pages. No number or data changed.