How a grade is given, and how it is taken away
This archive holds two kinds of claim. Game-side claims are what people have said about a title that has not shipped: 53 dated entries. Tradition-side claims are what the record says about Zhong Kui himself: 51 passages and 50 pictures. Neither kind is allowed to borrow the other's confidence, and neither is carried by adjectives. It is carried by a seal, a capture, and a link that was open when this archive was last compiled (23 Sep 2026).
01 The chain, and why it runs one way
-
A question becomes a URL and a query.
Not «find something about the release date» but «which monthly publishing-licence lists exist for this window, and does this title appear in them». The query is written into the file, so a capture records how it was found as well as what it said.
-
The page is downloaded to researchPack/sources_raw/.
One text file per document, named <subject>__<first-8-of-digest>.txt, headed with its URL, its access date and the query that reached it. 425 captures are held.
-
The digest is written down.
sources_raw/manifest.json keeps each file's byte size and SHA-256. A quote that cites a capture by size is checked against the filesystem, so a note cannot quietly claim a longer document than we hold.
-
Rows name their captures; a builder folds them into the ledger.
The harvest files under materials/ carry the quotation and the snapshot name together. scripts/build_ledger.py produces researchPack/facts-ledger.json and the src/data/*.json the pages read.
-
A gate reads the string back out of the capture.
scripts/verify_quotes.py walks every quoted line and every figure on every page's data and looks for it in the named file. A string that is not there is a build failure, not a note for later.
Nothing flows upward. A page is never the fix for a wrong claim: either the capture changes, or the sentence is deleted. Hand-editing src/data/ would be the fastest way to make this site look checked while it was not.
02 The game-side scale, counted
What the five seals mean is set out on the standing page; what decides which one a row gets is a domain test, not a judgement of the reporter. A row is promoted to Confirmed Confirmed only if the studio published it on a declared domain or a registry issued it: gamesci.cn, youtube.com (the studio's own channel), nppa.gov.cn, esrb.org. A trade report that repeats an official line correctly stays one grade below it, and an aggregator card stays an aggregator card however precise it looks. The current distribution over 53 rows:
- Confirmed Confirmed 17
- Attributed Attributed 21
- Reported Reported 3
- Not found Not found 8
- Denied Denied 4
03 The tradition-side scale
A quotation and a claim about a quotation are different objects, so they are stamped differently.
- Confirmed Confirmed The line was read against the captured edition of the book named on the slip. 45 of 51 passages.
- Inferred Inferred The quotation is there but the framing is not: an attribution of the book unverified, or an edition whose volume boundaries differ. 4 passages.
- Reported Reported A later scholar's assertion we could not recover in any accessible text. 1 passage, kept and labelled rather than dropped.
- Denied Denied A citation that fails when it is tested. 1 passage: the encyclopaedias' supposed Tang source does not contain the name.
Pictures carry a third scale, set by the holding record's own wording: firm statement, hedged attribution, or 相传, which drops the plate to Reported Reported . 50 of 63 catalogued plates have a local file; the rest are listed on the gallery page as catalogued and not shipped.
04 What may sit inside quotation marks
A string gets quote marks and the gold bar only if the whole string appears, in order, in a capture the row names. That is a verbatim record, and on the ledger it is 37 of 53 entries. In the classical-text engine it is every line printed in 楷體 under a work title, and those lines are what the verifier compares against the file.
Everything else is ours: a query we ran, a dating we derived by reading two documents, a summary of a paragraph we could not quote whole. Those sit on a dashed 石绿 slip labelled «our record, not a quotation», 16 entries on the ledger. The distinction exists because a paraphrase wearing quotation marks is the one error a research site cannot repair with an apology.
verbatim · their words (BKZ-053) 据了解，游戏科学希望《黑神话：钟馗》的核心主题能够更加普世，能够吸引任何类型的玩家，故事的叙述将更加直接，并以更加简单易懂、更容易被各种玩家群体接受的方式呈现。 held record press-gamersky-3djuegos-sales-interview__cb8748b7.txt
our record, not a quotation (BKZ-032) METHOD RESULT (13 monthly lists, Aug 2025 -> Aug 2026): rows read = 2111. '钟馗' = 0. '黑神话' = 0. '游戏科学' = 0. | POSITIVE CONTROL row: 33号避难所 | 广东音像出版社有限公司 | 广州龙畅网络科技有限公司 | 国新出审〔2026〕1917号 | ISBN 978-7-498-17356-0 | 2026年08月28日 held record official-nppa-banhao-negative-scan__570a3e4a.txt
Both are printed exactly as the ledger prints them. The gold bar means the string was found inside the file named under it; the dashed slip means it was not, because it never claimed to be anybody's words.
05 What is not a source here
Fan wikis, content farms and AI summaries are excluded, as are the sites in the same niche that publish release guesses for this title. The reason is not taste. Their dates are inferences dressed as news, several of them cite each other, and a citation hands the loop a backlink. A figure that appears only there stays out of the record, and the gap is left visible.
6 rows sit at the lowest aggregator tier, and that tier is the grade they are printed with. Where a trade page and an official FAQ disagree, both wordings are printed side by side: the difference is the finding, and smoothing it away is how a record becomes a press release.
06 Three times the method overruled the draft
Case 1
The homepage draft said the classical corpus ran «from Tang rubbings to Meiji prints». No Tang plate is shipped, so the sentence was describing an archive we did not have. It was replaced with counts derived at build time, and the same rule now applies wherever a page prints a number.
Case 2
A ledger row about disputed sales figures named two captures. Both were failed fetches, which the verifier flags as «every named snapshot is a failed fetch». A live mirror was captured on 23 Sep 2026 and the row was re-anchored to it, which also produced a studio design-intent statement that had not been in the ledger before. A separate row that recorded a search result in its own note was kept as a stub on purpose: for a negative finding, the failure is the record.
Case 3
When the verifier was first run against the ledger it came back with a finding on row after row: lines labelled «quoted verbatim» that were really a join of two documents, or a note about how we looked something up. The fix was not to loosen the checker. It was the two-treatment system in section 04, which is why the site now shows you which words are anybody's and which are ours.
07 Standing limits
There is no store page for this title, so there are no frame timings, drop rates, item counts or stat tables on this site, and there will not be any invented as «early data». Numeric game data enters only through a capture.
This is a display archive. While the game is unannounced for release, nothing here is a calculator, a build planner or a checklist tool: those are built against facts that do not exist yet, and a tool is the easiest object on a wiki to make look authoritative while it is wrong.
A page that throws during the build makes the static generator exit without a word and write no HTML, so every build is checked by counting the files it produced, and the layout asserts the shape of everything it iterates.
python3 scripts/build_ledger.py
python3 scripts/verify_quotes.py --strict
npm run build
find dist -name '*.html' | wc -l
Corrections: contact@blackmyth-zhongkui.wiki, with the entry id and the link. An entry is amended by adding a row, never by rewriting the old one.
Source record: every claim in this dossier carries the archived source it was read from, linked in place. No source, no row.