What “Verified” means here

Condition of the law corpus against its sources

Our text is extracted from PDFs the agencies publish. Extraction can fail, and we measure it with three methods whose blind spots differ. The figures below are counted live from the corpus.

Checked by three methods; no defect found

2,087 / 2,703Three methods: embedded-Latin scan, second extractor, glyph-order cross-check

Defects found

46 / 2,703The per-document count is shown on that document's page

Not verifiable with the methods available

570 / 2,703Each has a named reason — most often no reference font to compare against

"No defect found" is not "correct". Two of the three methods read the PDF's own character table, so when that table is wrong both are wrong together. The third method sees that class but does not cover every font — which is why some documents sit in "not verifiable".

What we deliberately do not do

• ✕No truth score, no confidence percentage, no Verified: YES/NO badge.
• ✕Colour never signals truth — only computable states: no source, stale, sources disagree.
• ✕The word “conflict” is never used of scripture, and no reliability metric is shown on scripture.
• ✕No figure that cannot be proven appears on the home page or anywhere else.

The principles every screen must pass

1. 01Truth Before AI — the fact is the protagonist, not the model
2. 02Evidence Before Opinion
3. 03Cite or Refuse — no evidence, no answer
4. 04Separate Facts from Interpretation — visibly, always
5. 05Every Source Is Traceable — every sentence leads back to an original
6. 06Intelligence Is Collaborative — readers can check and build on it
7. 07Never Claim More Than You Can Prove — every metric must explain its own computation

Licence status — the test we apply

Every edition carries two things, not one: a status and the evidence for it. A status without evidence is a claim, and this page exists so that we never claim more than we can show.

The four bases actually in use

• Granted in writing by the rights holderIt means A standard licence published by the rights holder themselves — CC0, CC BY, the Public Domain Mark, the DGA Open Government Licence — and the evidence is that publication.It does not mean That we vouch for the accuracy of the translation or compilation. A licence speaks about reuse rights, not about quality.Editions on this basis: business · realEstate · trade · economy · pli-ms · en-sujato · en-brahmali · th-dhiranandi · th-jayasaro · th-korakot · en-kelly · en-soma · en-suddhaso · en-kovilo · en-anandajoti · grc-sblgnt · ar-uthmani
• Declared public domain by the sourceIt means The publishing source states on its own page that the work is in the public domain, and we record the link to that statement.It does not mean That we have verified the copyright status in every jurisdiction. We record what the source says, not our own finding.Editions on this basis: he-wlc · en-jps1917 · en-kjv · sa-gita
• The operator’s own reading of public factsIt means No one granted anything and no source declared anything. The operator read checkable facts — the publisher and year of printing, or the fact that the text was promulgated in the Royal Gazette — and concluded it may be republished. The reasoning is recorded per edition and is revisable.It does not mean It is not counsel’s opinion and not a grant from anyone. It is the weakest basis we use, and it is labelled that way on purpose — if a rights holder disagrees, we take the text down.Editions on this basis: th-krisdika-consolidated · th-gazette-agency · th-gazette-ocs · th-gazette-ect · th-chabab-luang · pli-siamrath · en-pickthall
• Permission asserted, document not yet on fileIt means There is a record that the operator obtained permission, but no document we can show publicly yet. We display the text with this label rather than hiding the state.It does not mean That this equals a licence. Until the document is on file it is a one-sided assertion.Editions on this basis: th-attha-84000 · th-mahachula

Comparing editions: what evidence each match rests on

Editions of the same text divide and number their passages differently, so we never carry an item number across editions — we match on evidence and always say which kind.

• Titles match2,970 texts · 41.0%both editions give the same title
• One title contains the other290 texts · 4.0%one edition prints a long title that folds in its sub-headings while the other prints a short one, and the two point at one text without ambiguity
• Position in the volume3,721 texts · 51.4%both editions divide that volume into the same number of texts, so positions line up — but the titles differ, so this is not confirmed to be the same text
• Refused263 texts · 3.6%cannot be compared; the system does not guess

The first two count as verified · the third is not, and is labelled on every page that shows it.

These numbers are not a score. No target is set on them and nothing is optimised toward them — they say which kind of evidence supports each match, and move only when the data or the method moves.

Census change log

• ๗ กันยายน ๒๕๖๙ · method — Logging of causes begins — movements before this date were not recorded

Reference units per religion

We do not force one unit on every scripture — a unit is what that tradition itself uses to cite its text, and imposing a foreign unit changes what a reference number means.

Buddhism

texts (sutta/kaṇḍa) and numbered items of the printed edition

Christianity

verses within chapters of a book

Judaism

verses (pasuk) as divided by the edition that covers the whole canon

Islam

āyāt within sūrahs

Hinduism

ślokas within chapters

The unit-label contract (short form)

Whether a unit is labelled “text” or “item” comes from the collection bundle (per-edition counts made at build time), not from the page · G1a = a title the source gives, witnessed by a source field · G1b = a source title we transliterate under a declared scheme · G2 = a title we set or translate, which must be labelled as ours — the gate recounts from the documents before every release.

Every modification of a text

Every edition whose copy differs from its source is listed here and at the foot of every page that shows it — read live from the edition registry, so a line disappears by itself once the text is restored.

No edition currently differs from its source

Licence instruments the source registry leaves unspecified

• ธนาคารแห่งประเทศไทย — ตารางสถิติ BOTWEBSTAT: the registry states “Creative Commons Attributions” without a version — pending a reply we follow the strictest obligations (CC BY 4.0) · inquiry not yet sent · https://creativecommons.org/licenses/by/4.0/

The “summary” rows of the Mahachula edition

The Mahachula source has 8 units titled “summary” with nothing in the source distinguishing them from the text — we show them as the source has them and record this as a limit of the source, not a judgement of ours.

When we refuse to compare

A real case: in volume 30 one edition has 19 texts and the other 37 — when the division differs, comparing by position fails for the whole volume, not for some texts in it.

So the system offers no comparison in that volume and takes you to the target edition’s contents instead — declining is itself an answer.

What we cannot verify yet

51.4% of all cross-edition matches are “matched by position” — those volumes divide into the same number of texts, so positions line up, but we have no evidence yet that the texts are the same.

We show this figure rather than folding it into the verified count to make the numbers look better — hiding it would make the rest meaningless.

Why there is no truth score

We show no truth percentage, no credibility score, and no binary “verified ✓” badge, because such numbers cannot be shown to be computed — they are opinions dressed as measurements.

We show the raw materials of credibility instead, for you to judge: number of sources · their tier (primary/secondary) · when the data was fetched · the source document’s date · coverage · agreement between sources.

How to audit us

Automated checks run before every release:

• ✓the page served must be the edition the contents pointed at — checked for every volume of every edition
• ✓links we have published must not die — a broken citation destroys the thing we are for
• ✓an unverified match must not become “verified” unless the evidence changes
• ✓every figure on the site is read from the system at render time, never written in advance

Each check is tested by deliberately breaking the thing it guards and confirming it catches it — a check never tested that way does not count as a check.

The scope of this page today

This page describes methods actually enforced on the collections already ingested — a collection not yet ingested shows that status and carries no figures.

Current status of every collection →

Changes

Every change to these principles and methods is recorded here with its date.

Last updated ๗ กันยายน ๒๕๖๙

Permanent identifiers — and the time we broke ours

Every unit has a permanent identifier such as arnon.in.th/id/religion/luang-b1-8 — collection then text key — computed from the text’s own key alone and not from the page layout. The layout may change; the identifier does not.

Before 13 August 2026 the canon’s identifier was the page URL, so moving the reader’s base changed it once. We verified that every earlier address still resolves, but a changed identifier means a machine holding the old form as a key sees a different thing for the same text. Every earlier form is therefore declared on each page (sameAs) so the two can be matched — and we do not erase the record of it.

Names of world scriptures — which are ours

All 220 Thai names of world scriptures are ours (transliterations or names already in common use) — none of our sources is a Thai edition. Source-script names of the Tanakh and Bible have no witness yet for which file they were copied from (the WLC OSIS files carry only the collection title). Arabic sura names match Tanzil for 108 of 114 (six differ in hamza spelling); the English sura names follow our own transliteration scheme, not Tanzil’s. Every page declares the kind of each name in its colophon and in the machine-readable data.

World-scripture corpus digest — a history, not a current value

corpusChecksum is the sha256 over every segment of text. It may move only when the text a reader sees changes, or the method changes — and each move needs a new row in the register first: the bundle builder refuses to write a value that is not the last row, and old rows are never edited. Shown are the first 16 of 64 characters.

1. 711c8a9713f07029… · 2026-08-29 · first build — บันเดิลแรก — ทานัค · ไบเบิล · อัลกุรอาน · ภควัทคีตา ๒๒๐ เล่ม ๖๑,๒๕๔ ข้อ (ai-religion #13)
2. 4308187b8a11269c… · 2026-09-10 · text changed — P3.3–P3.5 คืนสิ่งที่ท่อดึงเดิมตัดทิ้ง: สัญลักษณ์ apparatus ของ SBLGNT ๘,๗๐๕ ตำแหน่ง · qere ของ WLC ๑,๒๕๔ รายการ (เข้า x.variants · ตัวบทที่เดินยังเป็น ketiv) · เครื่องหมาย פ ס ׆ ของ WLC — จำนวนข้อ ๖๑,๒๕๔ เท่าเดิม

The Tipiṭaka fetch pipeline from SuttaCentral — its fetch step cannot be re-proven offline

The Pali/English texts and the table of contents of the Tipiṭaka were fetched from SuttaCentral once, by three scripts (ingest-canon · add-translations · build-canon-sqlite), and stored as a database committed to the repository. The final artefacts are proven byte-for-byte by rebuild gates, but the fetch step itself cannot be re-proven without the network: the stored cache holds only 388 menu files, not the 7,824 texts. We choose not to re-fetch from a volunteer project merely to prove our own script, and will port this pipeline only when the source publishes a new edition — this declaration is withdrawn on the day the pipeline is ported and proven.

How to cite this site

Scripture pages carry a “copy verse link” button on every verse, giving an edition-pinned link (?ed=). Law pages cite per section via the section link (…/12) pinned with ?v=year. Cite the pinned link, not the bare page. A copy button on every unit of every collection is coming.