Loading…
Who is accountable for this record, how to reach them, and what the automated contributor is and is not allowed to do.
A dataset a newsroom is asked to cite has to have someone behind it. The register below is the editorial contributor list as it stands in the database, reproduced as recorded — including where a field is empty, which is shown as empty rather than filled in with something plausible.
CTF-C-BALDWIN-INDUSTRIESCTF-C-AUTOMATED-OFFICIAL-001The accountable publisher is the party recorded in the register above. Editorial decisions — what is included, what is published as verified, what is withdrawn — rest there, and a correction that is refused is refused by a person who will say so.
One entry in that register is not a person. An automated verifier holds a contributor record so that machine verification can be told apart from human verification in the review ledger, and its published limits are reproduced from its own record rather than summarised here. It may confirm literal facts against official systems; it may not approve motive, benefit, causation, disputed interpretation, or anything sensitive about a living person.
What the register does not yet contain is worth stating as plainly as what it does. The accountable publisher is entered under a role and an organisation rather than as a named individual, and no separate fact-checker, French-language editor or legal reviewer is listed. Correspondence sent to the addresses below reaches the editorial desk, and a reporter who needs to attribute a decision to a person by name should ask for one.
The contribution form is the channel that reaches the editorial queue with a tracking reference attached, and it is the fastest route for a correction. The addresses below are monitored and are the right choice when you need to attach documents or keep an email trail.
Corrections, missing events, a source we should hold, and anything that should enter the editorial queue with a tracking reference.
[email protected]Corrections
A factual error in a published record, particularly one you can point to a contradicting source for.
[email protected]Press
Questions on deadline, checking a figure before publication, and requests to interpret what a record does and does not establish.
[email protected]Editorial
Licensing and data requests, research partnerships, and anything about the record as a whole rather than a single entry.
Candidate generation: proposing that an event, a participant or an outcome exists and is missing from the record.
Entity matching: proposing that two spellings, two records or two datasets refer to the same person, party or riding.
Optical character recognition: turning a scanned page of a journal, a debate or a return into searchable text.
Link checking: confirming that a cited source is still reachable and still says what it said.
Translation drafting and terminology: producing a first French or English version of project-written text.
Machines find candidates, match entities, read scanned text, check links and draft translations. They do not assert facts, approve motive or causation, or decide what gets published — and every use is recorded on the row it touched.
Machine assistance is allowed to narrow a search and forbidden to end one. It may propose that a record exists, find a candidate source, match two spellings of a name, read a scanned page, notice that a link has broken, or draft a translation. It may not assert that something happened, decide that a source supports a claim, judge why a person acted, or move a record to published.
The distinction is not about how capable the tool is. It is about what a reader is entitled to assume when a record says a thing is verified: that a person read a source, at a named place inside it, and put their name to the judgement.
This restates a limit the project has already published. The automated validator that appears in the editorial register carries a biography stating that it validates literal dates, affiliations, offices and election results against explicit locators in independent official systems, and that it is not authorised to approve motive, benefit, causation, disputed interpretation or sensitive biographical claims. Nothing on this page extends that mandate.
A disclosure that lives only on a policy page cannot be checked. Machine assistance is therefore recorded on the rows it touched: a research candidate stores which kinds of assistance produced it, and a translation stores how much of it a machine drafted. Both carry the reviewer who accepted or rejected them, and a research candidate that becomes a published record carries the revision that promoted it.
This means the question "was a machine involved in this record" has an answer in the data, not an answer on a page. It also means the disclosure survives a change of policy: the rows keep their history whatever a future page says.
Both languages are primary. French is not a translated afterthought of English, and a French page that reads as translated English is treated as a defect and reported through the contribution form like any other error.
A machine may draft a translation or propose terminology. It may not approve one. Each translation row records how much machine assistance produced it, which reviewer accepted it, and whether it has gone out of date because the source text changed underneath it. Editorial review has a dedicated French review type, separate from fact-checking, so that language quality is signed off by someone reading it as French.
The prohibitions below are absolute, and each of them exists because it was breached once. An earlier phase of this project published invented vote counts, unsourced quotations and generated narrative essays; the account of what happened is in the sourcing standard. These rules are the shape of that failure, written as constraints.
Text last checked against the database: