The path of a claim
- The exact proposition and its period, place and register are defined before any search.
- The work, the specific edition, the physical or digital witness and the route of access are recorded separately. Restricted text is not reproduced.
- The passage is examined. We cite the printed page and the PDF page where they differ, the heading and paragraph for web pages, the passage identifier and line span for corpora, the timecode and recording identifier for audio. A landing page abstract does not count as reading the article.
- Original writing and editorial interventions are recorded separately. A normalised form is not a fresh attestation.
- Support is evaluated at the level of the claim: does it apply to this period, this place, this register? A citation alone does not validate an interpretation.
- Dependency is mapped. Reprints, translations, corpus copies and reviews can derive from one witness. "Corroborated" needs two independent sources, and independence needs a stated reason.
- Human scholarly review, the rights decision and the publication decision are separate steps. Agreement from a machine is not evidence.
Six independent statuses
| Axis | Values | Meaning |
|---|---|---|
| Source access | passage, abstract, metadata, mediated, inaccessible | How much of the source was actually examined |
| Claim support | unverified, supported, corroborated | How well the claim is supported |
| Dispute | none recorded, contested, unresolved | Whether alternatives conflict |
| Editorial | candidate, researching, review requested, reviewed, rejected | Where the work stands and who did it |
| Publication | withheld, approved, released, withdrawn | Decision and exact version |
| Rights | not assessed, restricted, permission required, permitted for a specific use | A decision for each use, never a general "open" flag |
A reviewed claim can remain contested. A public discussion can explain the alternatives, but it does not present an unresolved form as settled.
Absence is not evidence
An inscription is direct evidence of its written reading, not a recording of a population. A corpus of mostly literary or translated texts cannot give everyday frequency without a designed sample. When a search finds nothing we write "not found in this bounded search", never "not used in this dialect". Automatic labels and parsers are aids and are checked against the edition.
Rights and living communities
Text, manuscript image, transcription, translation, metadata, code and audio are separate rights objects. The First1K repository states CC BY-SA 4.0 and Scaife's documentation states an MIT licence for its code B09 Scaife Development B10 README. Neither is blanket permission for every linked work. The TLG restricts downloading and redistribution of texts B07 FAQ.
For any oral work, collection, analysis, quotation, voice playback, download, attribution or pseudonymity, and withdrawal are each defined separately. Precise locations and personal details are protected where needed. Speakers are people, not specimens, and are credited by name and place only with their consent.
Artificial intelligence in the work
Ellinofonia uses artificial intelligence tools. OpenAI's Codex and ChatGPT were used to draft the research framework and to organise and record the sources. Anthropic's Claude was used for independent critical reading of the work and to build the website. Artificial intelligence is not a source: every claim rests on a published source with its page, no word or form is generated by a machine to fill a gap, and a model's agreement does not count as evidence. People decide what is published.