Asserts that one entity exists, under one preferred name, with a type and a country.
Idriss Koudouss Koné · person · Côte d'Ivoire
What IWAC stores
The link and the record. The mention is not kept as an object of its own.
The workflow
Which steps are model calls, which are code, which are agents
Five steps, and who acts at each
Press of Burkina Faso, 8,045 articles, 20–22 October 2026
1
Extract
Model one call per article
Names of persons, organisations and places; one to six subjects.
2
Match
Code no model
Exact keys first. Ambiguous or fuzzy candidates are flagged, never linked.
3
Check in context
Model one call per article
Keep, choose or drop, only among the ids it was offered.
4
Write
Scripts no model
Dry run first. Each link annotated with its model: iwac:nerModel.
5
Leftovers
Agents then the curator
Reviewer agents propose new records. The curator approves each plan.
How the country is decided
Rules in the prompt
“Use the text first, the newspaper’s country second.”
“A body of another country is not this one.”
A contradiction is only “a different country, an incompatible calling … or dates that cannot fit”.
“Choose only among the candidates of that entry. Never invent an id.”
What the model is given for one name
Name as found
Kassoum Coulibaly
Newspaper’s country
Burkina Faso
67772Coulibaly Kassoum · person · Côte d'Ivoire No description
127886Kassoum Coulibaly · person · Burkina Faso Minister of state, minister of defence
KeepChooseDrop: leave unlinked
Illustration of the input’s shape, not the transcript of a real call
The first full run, Burkina Faso
Articles8,045
Matches kept in context88,250
Matches dropped, after the re-check1,776
Fuzzy or ambiguous names linked by the check522
Links written: 10,036 place, 40,763 subject50,799
New authority records from the leftovers674
Claude Haiku 5.5. Extraction 20 October (1 h 38, about $6.50); leftovers 21 October; context check and write 22 October. From the pipeline changelog and run files.
The other countries
Togo: 1,570 articles extracted and matched on 23 October; no context check, nothing written. Côte d'Ivoire, Benin, Niger and Nigeria: not run.
Three cases
Where the workflow went wrong or could not decide
Case 1: one acronym, three countries’ committees
Record 654 · Burkina FasoCNOPM32articles, most from Ivorian and Beninese papers
1314Scholarly reference Researcher, 2002 work on Sufi communities in Côte d’IvoireKept
188930Côte d’Ivoire Politician, Assembly vice-president 1989, foreign minister 2003New
188934Burkina Faso Lieutenant-colonel, spokesman of the failed coup of September 2015New
188954Côte d’Ivoire Deputy imam of Yamoussoukro, end-of-Ramadan prayer 1986New
No record“El Hadji Mamadou Bamba”, 1979: cannot be placedUnlinked
“Unknown” has to be representable
Authors and actors share one person set
Case 3: acronyms that are also words
The authority records hold 408 acronyms. Twelve are also common words in the articles.
Times written in lower case in the articles
UN · un43,823
FOI · foi, “faith”2,245
AIDE · aide, “help”971
Links on the FOI record (23699) that were the word foi317 of 410
Matches of “Foi” and “Aide” in the Burkinabè run, all dropped199 + 27
The rule now: capitals only
FOIkey
F.O.Isame key
Foino key
foino key
Lesson
An alternative label is also a matching key.
Danyi and Dankpen, read as Denmark
93of 124
place links on Danemark (428) were Togolese agency reports, 2020–2024, naming the prefectures of Danyi and Dankpen
Wrong links, removed75%
Who made the error?
A pass older than the iwac:nerModel annotation. Nothing records which model or script made these links.
Lesson
Provenance belongs on each link.
What stays human
One curator, some 50,000 links from one country run
Where the curator sits
Rules before runsWritten by the curator, rewritten after reading what went wrong
Ambiguity is never settled by codeThe model chooses among the records offered, or leaves the name unlinked
Every new record is approved731 proposed in the first pass: 18 held back as possibly existing, 39 flagged
Merges and splits are the curator’s59 merge pairs drawn from the reviewers’ remarks, withdrawn unapplied
Agents never writeScripts do: dry run, backup of each item, dump of every payload
Who validates?
The same model rereads
Not independent
Extractor and checker are both Claude Haiku 5.5.
A deterministic test
Independent, but narrow
A full-text count overturned wrong drops: 4 on the FOI, 10 on other acronyms.
A human sample
None yet
No hand-checked error rate for any country.
What escapes the loop is what no rule anticipated.
Four modelling decisions the cases forced
Country in the title
For generic bodies, identity includes jurisdiction.
Unlinked as an outcome
A mention may stay a mention when the source cannot settle it.
Provenance on each link
iwac:nerModel, iwac:descriptionModel: which model proposed this value.
Undated descriptions
Written today about articles decades old. Patched by a prompt rule; IWAC lacks a temporal marker.
Questions for the Alevi Archive
Where does the identity of a person or a cemaat live, and who decides that two factoids are about the same entity?
Can a probabilistic identification stay marked as such once it is in the graph?
Is a check by the same model a verification at all? What would count as an independent one?
When the source cannot settle an identity: an unlinked mention, a provisional record, or a “possibly the same as” relation?
Thank you
The curator approves records, merges and splits. The links still await a hand check.
Frédérick MadoreResearcher and Data CuratorAfrica Multiple Cluster of Excellence, University of Bayreuth
Scan for these slidesslides.frederickmadore.com
Appendix
Held back for questions
How step 2 matches names
Code, no model. Looser keys are tried only once every exact key has missed.
Exact keys: lower case, no accents, spaces or dashes
A key shared by two records is ambiguous, never resolved
Word order: “Coulibaly Kassoum” = “Kassoum Coulibaly”; never for initials
Plural fold for non-person subjects; accents kept, “Jeunes” ≠ “Jeûne”
Acronyms of 2–6 capitals get a case-sensitive key only
“(Country)” titles are also keyed bare within that country’s run
Twelve acronyms that are also words
UN43,823 in lower case
FOI2,245
AIDE971
AU
CAR
CET
AIT
UNIR
UL
SEMI
AD
UP
The other ten acronym records had 722 item links; 579 name the record in full or in capitals. The model read the other 143, and a full-text check overturned ten of its drops (seven “Nations-Unies”, two Tidjania, one bishops’ conference). 108 links were removed, 99 of them “aide”.
Event, place, person and organisation are types a language model blurs.
Further questions
Should provenance sit on every link, and what does that cost in a factoid model where links are the data?
How do you tell apart two people, or two ocaks, of one name in different regions? Is that decision a factoid, with a source and an author?
Does every factoid pass a person, and would it still if the data grew tenfold?
What the first extraction prompt picked up
Generic themes extracted as subjects13,582
of them “Islam”4,983
of them “Religion”1,115
Mentions of the article’s own source outlet1,700
The revised prompt, tested on 18 articles: subject entries fell from 185 to 127, and “Islam” from 12 articles to none. The Burkinabè run used the first prompt; Togo, the revised one.