Methodology

Language models are used as extraction tools in building this database, never as the authority for a fact. A model may locate a passage in a source and pull a value out of it; it does not get to supply the value. If no source states a figure, the field stays empty. An empty field is information — it says the evidence does not exist — and filling it with a plausible guess destroys that.

Evidence is attached per field, not per page. A single wreck may carry separate evidence for its depth, its loss date, its protection status and its typical visibility, each with its own source and its own confidence rating. This is why the sources section on a wreck page lists which field each source supports.

Each evidence record has fields for:

• source name, and a source URL where one exists
• source date — when the source itself was published, not when it was read
• the supporting excerpt, so the claim can be checked without leaving the page
• a confidence rating

Those fields are not populated to the same depth. Source name, source URL, excerpt and confidence are present on every evidence record; source dates are not, and most evidence records are undated. A missing date is left missing: inferring one from when the source was read, or from the page it sits on, would manufacture exactly the freshness signal this section claims to provide. The current figures are published on the coverage page.

Verified

Supported by an authoritative source — an official heritage register, a marine survey dataset, or agreement between independent published accounts.

Probable

Supported by a credible single source, typically a dive operator account, club record or established wreck database, with nothing contradicting it.

Unknown

Recorded but not corroborated, or drawn from a source whose basis cannot be established. Treated as a lead, not a fact.

A rating describes the strength of the evidence, not the importance of the fact. A verified rating on a depth figure means the sources agree, not that the seabed has been measured to the centimetre.

Sources conflict most often on maximum depth and on position, and less often on casualty figures and loss dates. Where a conflict exists it is stored as a conflict: both values, both sources, and a note on the discrepancy. The wreck page displays it rather than hiding it.

This is deliberate. A database that silently picks one of two depths looks more authoritative and is less useful — the divergence is often the most important thing a diver planning the dive needs to know.

A wreck page shows a conflicts section when its record carries one, naming both values and both sources. The first recorded conflict is on the Koln, where the same authoritative designation prints a latitude/longitude and a national grid reference that disagree by roughly a kilometre; the grid reference is used for the stored position and the disagreement is kept on the page rather than quietly dropped. How many conflicts are currently recorded is reported on the coverage page, which reads the number from the database rather than restating it here.

Not every position should be published. A record’s visibility state determines what is shown:

Public exact

Position is widely published and the site carries no protection that publication would undermine.

Public approximate

Position is reduced in precision — enough to locate the general area, not enough to drop a shot line.

Register only

A precise position exists in an official register; it is not republished here.

Restricted

Position withheld because the site is protected, licensed, or otherwise vulnerable to interference.

Hidden

No position shown. Applied to war graves and sites where publication would present a clear risk.

Where a record has no visibility state set, the position is withheld. The site shows coordinates only for records explicitly marked public exact or public approximate, so an unset or unrecognised state fails closed rather than defaulting to publication.

Records are revised when a new source is found, when a source is superseded, or when a correction is received. Where an evidence record carries a source date, that date is shown on the entry, so the age of the underlying claim is visible rather than implied by the age of the page. Where it does not, no date is shown and the age of that claim is genuinely unknown.

Wreck sites are not static. Several records in this database describe structures that have collapsed further, been salvaged, or had their legal protection changed since the sources were written. Where a source is old and dated, its date says so; where it is undated, the record cannot tell you how old the claim is, and this page would rather say that than imply currency it cannot support.

No original survey, no diver testimony, and no guarantee of currency. The database compiles what has been published and shows you where it came from. For conditions on a given day, and for the present legal status of a protected site, a local operator or the relevant authority is the source — not this site.

The same standard applies to the project’s own retrieval benchmark, where measurement limitations are published alongside the results.