Methodology: how the map is made
Updated
Benmer summarises what published sources have reported. We do not investigate and we make no findings of our own. This page is for journalists, researchers and anyone who wants to judge how far an entry can be relied on.
The name Benmer is short for benang merah, "the red thread": the connections that tie people, institutions and cases together behind the news.
What is on the map
The map holds people, organisations, events and cases in Indonesian public life, and the connections between them. Every map page shows the number of entries, the number of connections and the date of the newest source.
Entries cover only public office, court proceedings, official investigations and the conduct of organisations. Home addresses, private contact details, health information and family members with no public role are not included.
People and organisations from outside Indonesia are included only when they are linked to an Indonesian case or official. So a name that Wikidata places outside Indonesia can be suggested from the map's search only together with the Indonesian case or official it is linked to and a news link that shows the link.
Sources
Connections (the lines on the map) are made only from primary sources: media verified by the Press Council (Dewan Pers), news agencies, court documents, official documents of state institutions, or recognised reference works. The details are in How entries are sourced.
How statements are written
The note beside each citation stays within what that source reported. We add no background from elsewhere, do not inflate claims, and do not turn "will" into "has".
What officials, lawyers, police spokespeople, the KPK or parties say is attributed to them, not stated as fact. Where sources differ on a date or a number, the entry names the source it uses or shows the range.
Allegations and established facts
A connection is marked Allegation ("Dugaan") and drawn as a dashed line for as long as no court ruling has upheld it. This covers accusations, police reports, being named a suspect, arrest, detention, indictment and investigation.
Every reported denial or response is shown beside the allegation. Where the cited article includes no response, the note says so.
Verdicts, acquittals, appointments to office, family relationships and election results are not marked as allegations.
What the marks mean
| Mark | Meaning | Not included |
|---|---|---|
| Thin ring: Convicted | A source reports a criminal verdict. The entry says whether the verdict is final and binding (inkracht) or still under appeal, cassation or judicial review (PK). | Being a suspect or defendant, an investigation, or a civil judgment |
| Closed padlock: Detained | A dated status line reports the person in prison, a detention centre or custody. | Not a finding of guilt; may lag behind events |
| Open padlock: Detention suspended | The status line reports suspended detention, city arrest or house arrest. | Not a finding of guilt; may lag behind events |
| Grey: Deceased | A source reports that the person has died. | Allegations against a person who died before a final verdict are marked as allegations |
Collection and reading with AI
We collect articles from the listed media every day. Our collector identifies itself, respects each site's robots.txt and content signals, pauses between requests, and stops collecting from a site as soon as that site refuses. Articles from media that have chosen manual recording are never collected.
An AI reads the article and proposes facts, each with a quote from its article. A second AI check, separate from the first, does not read the whole article: it sees the quote, the article's headline and the outlet's name, the proposed statement with its summary and date, and the reported response if there is one, and judges whether the quote really supports the statement.
The AI's proposals are not taken on trust. Our code, not the AI, checks every proposal:
- the quote must appear in the article exactly as written; a quote that has been changed or stitched together is rejected;
- every name must appear in the article, and people must be named in full; a name given only as initials, as in KPK or prosecutors' press releases, is never linked to a person;
- the legal status (witness, suspect, defendant, convicted) must be written in the quote for the person concerned;
- deciding which entry a name refers to is done by our code. If a name matches more than one entry, an editor decides;
- a fact published without an editor must be written in its own quote. A hearing date must be written with its day and month; where the year is not written, the date must be the first one after the article's date, and a quote saying the hearing was "ditunda", "diundur" or "batal" (postponed, put back or cancelled) is read by an editor. A position must be written immediately before the person's name, or immediately after it with only a comma, "selaku" or "sebagai" ("as") in between, followed directly by the name of the institution or party. A quote containing a word such as "mantan", "pernah", "mundur" or "dipecat" (former, once, resigned or dismissed), and a position preceded by a word such as "wakil" (deputy) or "Plt." (acting), is read by an editor. These checks read the order of words, not their meaning, so we choose to send more facts to an editor.
Before publishing without an editor was switched on, the whole chain was tested with a set of fictional articles whose results were known in advance, including namesakes, initials, denials and articles containing hidden instructions for the AI.
Publication tiers
After those checks, our code, not the AI, puts every fact in one of three tiers. A fact that fails even one condition of its tier is read by an editor.
Some conditions require two outlets with different owners: the same fact reported by two outlets that do not belong to one owner group, with the two articles published no more than 30 days apart. Until that second report arrives, the fact waits for an editor, who can decide it at any time. After 30 days nothing happens automatically: the fact stays with the editor.
Tier A: published at once
- Offices and roles. Appointments to office, election results and roles in a party, from a person already on the map to an institution or party an editor has marked, with a position name from a list the editors have approved. The person must already have had their identity confirmed by an editor, or already have a published connection to the same institution or party, because two people can share a name. This also includes a new article reporting such a connection that is already published, which is added as a further source, and hearing dates in cases already on the map, where the date is written in the quote and had not yet passed when the article was published. For these, one article is enough.
- A new person, because of their office. Someone not yet on the map can be added without an editor only because of an office or role as above, and only when two outlets with different owners report the same office within 30 days, the newer of the two articles published within the last year. Both articles must name the person in full, and there must be no other entry or proposed entry, including one an editor turned down, with the same name, the same alias, or the same first and last name. The new entry's role line is taken from that fact. A single source never creates a new person automatically.
- Family links between public figures. A child, father, mother, wife, husband or sibling ("anak", "ayah", "ibu", "istri", "suami", "saudara kandung", "adik", "kakak"), only between two people already on the map who each already have a published office or role connection, and only when two outlets with different owners report it within 30 days, the newer of the two articles published within the last year. Each of the two must have had their identity confirmed by an editor, or the two must already have a published connection to each other, because two people can share a name. Where the two people are already linked by any family relationship, in either direction, an editor reads it. In-laws, such as a son- or daughter-in-law, a brother- or sister-in-law, or a parent-in-law ("menantu", "ipar", "mertua"), and a former wife or husband ("mantan istri", "mantan suami") are always read by an editor.
A Tier A fact may not be an allegation, and may not carry a legal status or a response. It is published in a standard sentence built from the data, not written by the AI, with a link to the source article.
Tier B: held for 48 hours, then published unless an editor vetoes it
- Official proceedings. The status of suspect, defendant or convicted ("tersangka", "terdakwa", "terpidana") on a connection, where the quote names the KPK, the Kejaksaan (prosecutors), the Polri (police) or the court that set it. Our code confirms that the body is in fact written in the quote. Suspect and defendant must be written as their own word; convicted may be written as a verdict (for example "divonis", sentenced). Other allegations, including any that do not name an official body, are read by an editor.
- Business ownership. Owner, shareholder or beneficial owner ("pemilik", "pemegang saham", "beneficial owner") of an organisation, reported as a fact, not as an accusation. Where a cited article carries a denial or response about that ownership, an editor reads it.
Of legal statuses, only one on a connection can be in this tier. A verdict reported as a change of status of a person, for example a conviction that gives them the Convicted mark, is always read by an editor.
The other conditions: both ends of the connection are already on the map; two outlets with different owners report it within 30 days, and both articles were published within the last 30 days, both when the fact is held and again when it is published, so an old report is never shown as current; each person in the connection has had their identity confirmed by an editor, or already has a published connection to the other end; the person carries no status mark (such as Convicted or Detained), and no connection on the map already gives them the same or a later status in the same case (suspect, then defendant, then convicted), since an editor decides any change of status; the second AI check has no doubt about it; and a third reading by AI, which looks only for reasons not to publish (a namesake, a denial, an acquittal, a retraction, an outdated status, an actor that is not an official body, or a quote about someone else), has no doubt about it either.
For 48 hours the fact is held and shown in the editors' queue, where an editor can veto it or publish it early. When the time is up, every condition is checked again at that moment, and the fact is published only if all of them still hold; otherwise it goes back to an editor.
The summary of a Tier B fact is the outlets' own words, attributed: "Menurut (first outlet) dan (second outlet), …" ("According to … and …"), followed by the quote from the article, not a paraphrase. Suspect and defendant statuses are still marked Allegation. For a suspect or defendant status, a response reported by either of the two cited articles is shown beside the allegation; where neither article carries one, the page says "Belum ada tanggapan dalam sumber yang dikutip" (No response yet in the cited sources). A conviction, like an ownership, carries no response line: where a cited article carries a response to it, an editor reads it.
Tier C: read by an editor
Every other fact is read by an editor before it is published, and always so for: findings of NGOs or journalists, anonymous sources, one politician's accusation against someone else, ordinary people with no public role, children, victims, sexual offences, and a person's health, religion or ethnicity. The AI marks such facts. For children, victims, sexual offences, health, religion and ethnicity, our code also looks for words that point to them on its own; the code can only add that mark, never remove it.
An editor can also mark an entry always review: no fact about that entry is published without an editor.
Safeguards
- A reader's report on a Tier B fact. A report through the correction form, from the Report link on a connection published through Tier B, hides that connection at once until an editor decides. If the editor rejects the report, the connection is shown again, unless another report on it is still open or an editor has changed the connection since the report came in. If the editor resolves the report, the connection stays hidden until an editor shows it again. The only exception is where hiding the connection would leave an entry's summary citing a source no other visible connection supports: the connection stays visible, and the report goes straight to an editor.
- A rejection stands. A fact an editor has rejected or undone (the same two entries and the same relation) is never published automatically again: every later report of it is read by an editor.
- Tier A facts are not hidden because of a report, so that a report cannot be used to strip someone's public office from the map. A report about a Tier A fact is still read by an editor like any other.
- Automatic pause. When an editor undoes or hides a connection or an entry published without an editor, or logs a correction to it in the Corrections log, Tier B stops at once and the owner is told by email. While it is stopped, new Tier B facts are read by an editor and facts already held stay held. Tier B runs again only when an editor switches it back on.
- Spot checks. Every morning we receive an email listing the facts being held, the facts published from the hold, and a random 10% of the Tier A facts published the day before, to be read again.
- Re-checks. Suspect and defendant statuses, detention, and verdicts that are not yet final and binding are searched again for newer reports at regular intervals.
- Tested before running. New people, family links and Tier B run only after the set of fictional articles above, now with cases for every tier, passes: every case must pass at least once in the five newest test runs of the last 14 days read wholly under the current instructions, and in none of the test runs of those 14 days may a fact that should be read by an editor be in Tier A or B, or a Tier B fact in Tier A. The newest loaded set must be among those runs and be fully read.
Every fact published without an editor is logged and can be undone by an editor at any time.
Right of reply
The right of reply applies equally to facts published without an editor. Anyone named in an allegation can send a response through the Corrections page, and we add a fair summary of that response beside the allegation. Anyone can also ask there for a correction or removal, within the target times that page gives.
Checking and "Last checked"
Apart from the Tier A and Tier B facts described above, an editor reads every fact against the cited article, including its headline and date, before it is published. Allegations and legal statuses outside Tier B, changes to criminal status, family relationships outside Tier A, and every new entry other than a person added because of their office (organisations, parties and cases alike) are always read by an editor. So is every statement the second check has doubts about, and every name that may belong to more than one entry.
Automatic checks also make sure every connection has a source and every link still works.
Last checked on an entry is the date we last searched for newer reports. It is no guarantee that nothing was missed.
Limits
Entries can lag behind events, and the map is not complete. Not appearing on the map means nothing. Appearing on the map is not a finding of wrongdoing: many entries are officials doing their jobs, witnesses, victims, or people named because they are connected to someone else.
Citing Benmer
For a fact, cite the original source after reading it yourself. Cite Benmer for its map or summaries, using the entry's permanent address, for example benmer.wiki/tokoh/…. New reports are listed on the Updates page and in the RSS feed.