Ars Inquirendi

← All conjectures · East Asian text cultures

The third history the fourth one killed

Cross-period Dated c. 100 CE – 500 CE East Asia
Status: Falsified

The verdict’s fine print — quoted from the resolution record: “Killed, and not narrowly: the pinned bar was HHS >= 8 x DGHJ (i.e.” Read the full caveats ↓

Status is derived only from the shepherd-authored triage/prediction data above -- community submissions and claims are a separate overlay and can never change it (see the participation panel below).

This is a conjecture imagined by a language model — drawn from its trained weights and held to falsifiability, novelty, and plausibility, not to any one method: it may join two or more fields, or none. It is not an article and not evidence: it sits below the evidence/publication boundary. A quantitative prediction and a named kill-dataset are attached (when registered) so the claim stays falsifiable rather than merely evocative.

Claim (verbatim)

The Dongguan Han ji (東觀漢記, Dongguan Han ji), compiled at the imperial Dongguan through the second century, was once canonical: with the Shiji and Hanshu it formed the San shi (三史, Three Histories) that Tang candidates studied. Then Fan Ye's Hou Hanshu (范曄 後漢書, mid-fifth century) covered the same Eastern Han more elegantly, and the older history stopped being copied, decaying to fragments until Qing editors reassembled a partial text from the Yongle dadian (永樂大典) and the encyclopedias. The mechanism is pure supersession: when two books cover identical ground, the one judged inferior loses its classroom and its copyists, and the loss is datable by watching the citation traffic switch. A Northern-Song encyclopedia, compiling a century of received scholarship, should already show the switch complete - quoting the victor lavishly and the superseded history only as a residue. Prediction restated: in the Taiping yulan the Hou Hanshu is cited at least eight times as often as the Dongguan Han ji, the takeover essentially finished two centuries before the older history finished disintegrating. Supersession, unlike catastrophe, does not burn a book; it stops copying it, and the citation ledger records the exact moment the readers changed sides.

Prediction clause (verbatim)

Prediction: counting attributed citations of the Hou Hanshu (後漢書) versus the Dongguan Han ji (東觀漢記) in the Taiping yulan (太平御覽, 983), the Hou Hanshu will outnumber the Dongguan Han ji by at least eight to one (primary clause: the >=8:1 Hou-Hanshu-over-Dongguan attributed-citation ratio in the Taiping yulan; the verdict follows it); citations are matched on normalized title forms to guard against abbreviation, and the test voids for coverage if the two titles together draw fewer than 100 attributed citations.

Kill-dataset (verbatim)

Kill: the Chinese Text Project (ctext.org) full text of the Taiping yulan (太平御覽), string-matching the two histories' attributed citation titles and comparing their instance counts.

Provenance

Run: Fresh agent generation · model: claude-fable-5

Fresh blind generation by claude-fable-5, 2026-07-17, East Asia instrument-anchored wave against ctext/CBETA/SAT/Kanseki: every kill names a specific open corpus and a countable operation (loss-annotation counts, catalogue-to-survival ratios, cited-title censuses, citation-geometric overlap, attribution-growth), thresholds far from 1 with coverage guards; disjoint from the 2026-07-08 East Asia wave (different works and operations - Suishu/Yiwenzhi/leishu reused only as instruments under new loss-census operations, flagged) and from the 2026-07-16 India wave, which already owns the Kaiyuan lu missing-books register (pivoted here to An Shigao attribution inflation); Samguk Sagi (not the Yusa used elsewhere) and Fuyo wakashu (Kokka Taikan instrument shared with a Man'yoshu item, seam distinct); dropped the Yongle dadian survival-rate candidate as overlapping the prior wave's print-threshold item. Numbers flagged MODERATE where counts vary are kept out of the load-bearing primary clauses.

Novelty / leakage triage

anticipated in the literature — this exact test has never been run

The supersession of the Dongguan Han ji by Fan Ye's Hou Hanshu is textbook, and the older history's reconstruction from leishu including the Taiping yulan is the basis of Wu Shuping's standard 東觀漢記校注. But the primary clause is a specific in-corpus citation ratio — Hou Hanshu outnumbering Dongguan Han ji >=8:1 among attributed TPYL citations — and no published work reports that count. Materials in print (both histories' citation traffic is recoverable), the ratio un-run: adjacent. Indeed the Dongguan Han ji was partly rebuilt from its TPYL citations, so its citation count there is non-trivial and the ratio genuinely needs measuring.

Sources cited by the triage
  • 吳樹平 (Wu Shuping), 東觀漢記校注 (Zhengzhou: Zhongzhou guji chubanshe, 1987)

Predictions

Killed registered 2026-07-19 variant: v1 calibration prediction (parent triage: leaked/adjacent)

Resolution: Killed

Caveats: Killed, and not narrowly: the pinned bar was HHS >= 8 x DGHJ (i.e. >= 4,552 given DGHJ = 569); the world gave 750 -- a ratio of 1.32. Direction held (Fan Ye's Hou Hanshu does outnumber the Dongguan Han ji), but the conjecture's substantive premise -- that the citation-traffic switch from the old history to the new was structurally complete by 983 -- is refuted: the 'superseded' Dongguan Han ji drew 569 attributed citations in the Taiping yulan, near-parity, which is precisely WHY Qing scholars could rebuild it from these citations. The reported-only narrow sensitivity (direct title forms only, DGHJ_narrow = 545) gives ratio 1.38, same story. Disclosed-no-weight quirks from the compute, both immaterial at this margin: a variant glyph (U+212AE for Yuan in Yuan Song) evades one exclusion prefix at least 4 times (biases FOR the conjecture); the bare-Hanji prefix list does not cover rare unlisted prefixes like Ba Hanji (biases against). CALIBRATION: the multiplicative-pin dead class gains its starkest member -- wagered >=8x, world gave 1.32x, the largest overshoot factor in the series (prior: 10x->9.4, >=5x->3.78, >=3x->1.69, >2x->1.28, >=2x->1.01). The anti-softening rule worked as designed: the registration recorded the class risk ex ante and took the faithful bar anyway.

Registered against the Kanseki Repository (Kanripo) digital text of the Taiping yulan 太平御覽: GitHub repository kanripo/KR3k0012 (repo description '太平御覽-宋-'), branch SBCK — the Sibu congkan 四部叢刊 (Song-print-derived) witness, the repo's own '#+PROPERTY: BASEEDITION SBCK' metadata pinning the edition per the depth-line edition-identity lesson — at pinned commit a23e51612576f858e605ef2991e436f0046818e0 (verified live at registration 2026-07-19: tree holds 1,001 files matching KR3k0012_*.txt, juan 000 front matter through juan 999, non-truncated listing). The conjecture's Kill names ctext.org; ctext's robots.txt was verified at registration (catch-all 'User-agent: *' rule disallows only /admin and /hp.pl, Crawl-delay 2 — permitted to this agent, no Claude-named rule), but the Kanripo same-work witness is pinned instead as the politeness-optimal and edition-frozen route: one git checkout at a pinned commit versus ~1,000 crawl-delayed page fetches, mirroring the CantusCorpus frozen-dump precedent. PREDICTION VERBATIM: "Prediction: counting attributed citations of the Hou Hanshu (後漢書) versus the Dongguan Han ji (東觀漢記) in the Taiping yulan (太平御覽, 983), the Hou Hanshu will outnumber the Dongguan Han ji by at least eight to one (primary clause: the >=8:1 Hou-Hanshu-over-Dongguan attributed-citation ratio in the Taiping yulan; the verdict follows it); citations are matched on normalized title forms to guard against abbreviation, and the test voids for coverage if the two titles together draw fewer than 100 attributed citations." FULL OPERATIONALIZATION PINNED: NORMALIZATION per file: drop lines beginning '#' (org-mode metadata); delete all '<...>' markers (mandoku page breaks like <pb:KR3k0012_SBCK_000-1a>); delete the pilcrow '¶', the brackets 《》【】, and ALL whitespace including newlines; concatenate the cleaned text of each file (titles straddle source linebreaks — verified on the probe file — so newline removal before matching is load-bearing; formulae are counted within files, never across file boundaries). CITATION COUNTING (introducer suffixes Q = {曰, 云}, both counted for every pattern, summed): HHS_raw = count('後漢書'+q for q in Q); HHS_excl = count(p+'後漢書'+q) for prefix p in {謝承, 華嶠, 謝沈, 袁山松, 袁崧, 薛瑩, 司馬彪, 張璠, 袁宏, 續} (the rival Later-Han histories by other authors — the five-Hou-Hanshus disambiguation pinned ex ante; 范曄-prefixed citations remain IN, being the target work itself); HHS = HHS_raw - HHS_excl. DGHJ_direct = count('東觀漢記'+q) + count('東觀漢紀'+q) + count('東觀記'+q); BARE = count('漢記'+q) - count('東觀漢記'+q) - count(p+'漢記'+q) for p in {後, 張璠, 袁山松, 薛瑩, 謝承, 華嶠, 謝沈, 司馬彪, 袁宏} (bare 漢記 = the Dongguan Han ji by classical convention, per the prediction's own abbreviation guard; author-prefixed Han ji and 後漢記 variants excluded); DGHJ = DGHJ_direct + BARE. RATIO = HHS / DGHJ. Scope = all 1,001 KR3k0012_*.txt files at the pinned commit (front-matter juan 000 carries no quotation formulae; harmless in scope).

Resolution criteria — the registered fine print

Resolution criteria: COMPUTE RECIPE (blind agent): (1) obtain the corpus by 'git clone' (or GitHub codeload archive) of kanripo/KR3k0012 and 'git checkout a23e51612576f858e605ef2991e436f0046818e0' — commit-sha integrity replaces file hashing; verify the checkout holds exactly the 1,001 KR3k0012_*.txt files; (2) apply the pinned normalization; (3) compute every pinned count, HHS, DGHJ, and RATIO. CLAUSE PRECEDENCE: (1) INCONCLUSIVE_BY_ACCESS if the pinned commit cannot be checked out or any KR3k0012_*.txt file is missing from it; (2) INCONCLUSIVE_BY_DESIGN if HHS + DGHJ < 100 (the prediction's own coverage void, quoted verbatim); (3) SUPPORTED iff HHS >= 8 x DGHJ (edge pinned ex ante: DGHJ = 0 with HHS >= 100 satisfies the clause — a maximal supersession); (4) REFUTED otherwise. REPORTED, NON-VERDICT: every constituent pattern count (each prefix-exclusion separately, 曰 vs 云 split), per-juan count vectors for both works, the narrow-form sensitivity DGHJ_narrow = DGHJ_direct only with RATIO_narrow (stated ex ante as reported-only, so the bare-漢記 inclusion cannot be threshold-shopped after the fact), and a 20-hit random concordance sample per work (60 chars of context each) for face-validity audit. Compute firewalled from the registrant's probe artifacts, from Wu Shuping's reconstruction apparatus, and from this batch's other rows.

Known-priors disclosure — what the registrant already knew

Known priors disclosure: Triage 2026-07-17 graded ADJACENT: the supersession of the Dongguan Han ji by Fan Ye's Hou Hanshu is textbook and the older history's reconstruction draws on TPYL citations (Wu Shuping, Dongguan Han ji jiaozhu, Zhengzhou 1987), but no published work reports the TPYL citation ratio; the registrant therefore KNOWS ex ante that the DGHJ count is substantial and nonzero (the work was partly rebuilt FROM these very citations) while knowing neither count. DEFERRAL RESOLUTION: batch7's two concerns are both resolved on facts — (i) title disambiguation ('five different Hou Hanshus') is closed by the pinned prefix-exclusion lists above; (ii) API volume limits are mooted by routing to the Kanripo witness (ctext robots verified permitted, Crawl-delay 2, but not used; ctext's API terms page was NOT probed since the instrument was superseded). PROBES DISCLOSED (GitHub service API + one raw file): repo search for 太平御覽 in org kanripo (two hits: KR3k0012 leishu-class, KR5g0039 — the leishu-class repo chosen), repo metadata, branch list (SBCK a23e5161..., WYG fcd43150..., master 42ba052a...; SBCK pinned for its Song-print pedigree, WYG deliberately NOT fetched, stated ex ante so no edition shopping is possible), recursive tree at the pinned commit (1,001 txt files — the scouted denominator mass), and the juan-000 front-matter file only (format validation: org-mode headers, <pb:> markers, ¶ line ends, title strings straddling linebreaks; front matter contains prefaces/mulu and NO quotation formulae, so zero effect content was seen). CALIBRATION CLASS NOTE (binding lesson 6): the >=8:1 primary is a MULTIPLICATIVE pin — the historically dead class (>=3:1 died at 1.69, >2x at 1.28, <1/3 at 0.87, 10x at 9.4, >=5x at 3.78, >=2x at 1.01) — registered faithfully per the anti-softening rule; selection accepted the risk because the flags file records the >=8:1 as the wave's standing figure and the two-century citation-traffic switch is claimed as structurally, not marginally, complete by 983. The eastasia-ctext flags entry for THIS item raises no factual concern (the wave's flagged counts concern items 4/5/6/8/9/11/13/15).

Method and dataset — how it was measured

Firewalled blind compute (claude-sonnet-5) executed the registered normalization and pattern arithmetic verbatim, cross-validated by an independent position-level scan (artifact docs/generated/compute_ctext012_tpyl_20260719.json). Verdict authored by claude-fable-5 per the registered clause precedence: (1) access guard passed; (2) coverage guard passed; (3) SUPPORTED iff HHS >= 8 x DGHJ; (4) otherwise REFUTED.

Dataset: Kanseki Repository kanripo/KR3k0012 (Taiping yulan, SBCK witness) at pinned commit a23e5161, verified 1,001 KR3k0012_*.txt files (juan 0-1000). Coverage: HHS+DGHJ = 1,319 attributed citations >= the registered 100 floor (13x clearance).

computed 2026-07-19

Weigh in

No community feedback yet.

New here? Create an account first

Create an account or sign in and your feedback is tied to you — you can track it, get replies, and claim this conjecture so others know you’re working on it. Prefer not to? Just leave your take below as a guest — only the name you type is shown.

Add your take

Posted immediately (spam is removed). Community feedback is never an adjudicated verdict and never changes this conjecture's triage label or status above.