HKADH 2027 · Education University of Hong Kong, 21 January 2027

Knowledge gaps and mediated access

AI infrastructure and practice for African research data

QR code linking to these slides online Scan for these slidesslides.frederickmadore.com

A catalogue should define its own terms

Three goals: the Cluster’s agenda, taken to the catalogue
  1. Say what a term means in the research we catalogue
  2. Shift who describes: definitions drawn from that research
  3. Keep each definition cited, dated and open to revision
The question Can language models draft these definitions from AMIRA’s own holdings, with every claim traceable to a source? Does that do better than borrowing a definition, or asking a model alone?

“It is our contention that this approach allows us to transcend the limits of thinking in fixed categories and binaries.”

Rüdiger Seesemann, the Cluster’s Dean, writing in 2020. University of Bayreuth African Studies Working Papers 24

A catalogue of research data held elsewhere

The Africa Multiple Interactive Research Atlas (AMIRA) is the research-data platform of the Africa Multiple Cluster of Excellence, University of Bayreuth (Germany). The material usually stays with the researchers, archives and libraries that hold it; the metadata is shared.

3,975Research items
106Projects
813Publications
28Languages
43Podcasts
142Videos

AMIRA MCP server, collection overview, 16 November 2026

data.africamultiple.uni-bayreuth.de Open AMIRA ↗

Why an MCP server, not just an API

An API
  • Written for programmers
  • Each new question needs someone to write code
  • AMIRA has one, from Omeka S; the MCP server reads it
An MCP server
  • Describes the same catalogue to a language model
  • Named tools in plain words; the model picks one, no new code
  • Instructions on connecting: how to cite, what is private
  • One open standard: Claude, ChatGPT, our test harness
Who decides what We decide which tools exist and what they return. The assistant and its provider decide what reaches the answer.

Two tools this pilot builds on

Linking keywords to Wikidata
  • CAREL, led by Jiayu Yang: linking without task-specific training
  • Rules settle the clear cases; a local open-source model, the ambiguous ones
  • 89.8% and 90.9% accuracy on Library of Congress headings

Yang et al., JCDL 2026, doi:10.1145/3805696.3846031; CIKM 2026, doi:10.1145/3799682.3840034

AMIRA MCP server
  • The server, not the model, computes counts, facets and related records
  • Curator mode: private records and full text, behind a personal token
QR code linking to the AMIRA MCP server repository on GitHub Code on GitHubgithub.com/AM-Digital-Research-Environment/amira-mcp-server

A label and a list, but no definition

data.africamultiple.uni-bayreuth.de/s/amira/item/3270 Open record ↗
Records
55
Definition
None

Readers also need to know what a concept means and how the research catalogued in AMIRA uses it.

What a definition would add What highlife means in the research AMIRA holds, each claim cited to an article.

Standard vocabularies identify a concept but rarely explain it

Library of Congress
Highlife (Music)
No scope note
Mbira music
A cataloguing instruction: “Here are entered compositions not in a specific form or of a specific type for solo mbira…”
Wikidata
highlife
“musical genre originating in Ghana”
timbila, a xylophone
Listed with the alias “thumb pianos”

id.loc.gov and wikidata.org, read 16 and 17 November 2026

The sources: ILAM’s journal African Music

The International Library of African Music (ILAM), at Rhodes University (South Africa), holds recordings, journals and photographs of music from across Africa.

1954–2023Years
45Issues
1,032Articles
935With full text
  • Definitions quoted from seventy years of research
  • Articles found by what they discuss
  • Vocabulary dated, article by article

AMIRA, ILAM collection page and full-text map, 16 November 2026. The full text is private.

data.africamultiple.uni-bayreuth.de Open the collection ↗

Highlife, defined from AMIRA’s sources

Highlife is a popular dance music of urban West Africa, above all Ghana and Nigeria, that fuses African and Western musical styles. Writers describe it as a compromise between African traditional forms and Western music, mostly sung in many vernaculars with simple, binary-form melodies. It emerged as a national rather than ethnic music in the 1950s.

ClaimQuote found in the text the model readSource
1“all represent this compromise between African traditional forms and Western music.”Smith 1962
2“The highlife, primarily vocal music, is sung in the various vernaculars.”Smith 1962
3“Highlife emerged in the 1950s as a national, rather than ethnic or regional-based music.”Emielu 2013

7 of 7 claims carry a quote found in the text the model read.

Claude Haiku 5.5, route D, run of 16 November 2026. Sources: articles in African Music (ILAM), linked to their AMIRA records.

Four models reading AMIRA on highlife

ModelOrigin givenQuotes foundEst. cost (USD)
Claude Haiku 5.5Debated: Konkomba singing bands, several streams7 of 70.014
GPT-6 LunaGhana, then across West Africa3 of 30.014
Gemini 3.5 Flash-LiteLocal forms with brass band, jazz and ballroom3 of 30.016
Mistral Small 4Local forms with ballroom, brass band and Latin American music1 of 50.029

Route D, one run per model, 16 November 2026. Costs estimated from list prices. Tool calls and what each definition adds: annexe.

One English name, three drums

Talking drum
57articles

One Wikidata item. Reading AMIRA, all three usable definitions name members: Yoruba dùndún, Akan atumpan.

Dùndún, Yoruba (Nigeria)
26articles

No Wikidata item. The Yoruba label “Dundun” sits on the Mande dunun, another drum.

Donno or dondo, Akan (Ghana)
9articles with “donno”

No Wikidata item. Alone, only one model of four placed it in Ghana.

Does one model keep the three drums apart? Verdicts “distinct”: model alone 3 of 12; reading full texts 8 of 12; passages only 0 of 12, each naming the dondo merged with a Yoruba drum.

Articles: full-text matches in African Music (ILAM). Wikidata read 17 November 2026. Definitions: four models, one run per cell, 17 November 2026. Open verdicts, no answer key: Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon, 17 November 2026; uncalibrated; agreed on 11 of 12 sets.

Reading AMIRA won where the models knew least

Blind verdicts, uncalibratedModel aloneReading AMIRA
Known well: highlife, timbila18 of 246 of 24
Three drums: dondo, dùndún, talking drum6 of 3327 of 33
The Cluster’s concepts: multiplicity, relationality, reflexivity9 of 3627 of 36
All eight concepts33 of 9360 of 93
The judges know what the models know They score from their own training, the knowledge the model alone draws on. No researcher has calibrated them.

Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon chose between each author model’s two definitions, shown as text only. Definitions 16 and 17 November, verdicts 17 November 2026; one run per cell. All three judges agreed on 20 of 31 pairs. Talking drum: 9 verdicts, one grounded answer unusable.

Wikidata, extended but never contradicted

Wikidata: multiplicity

“philosophical assertion that there is more than one geo-historical trajectory, developed by Edmund Husserl, Henri Bergson, and Gilles Deleuze”

Reading AMIRA: GPT-6 Luna

“Multiplicity is a way of understanding African and African diasporic life-worlds as heterogeneous, coexisting, and mutually influential rather than as one uniform object or fixed category. …”

57 verdicts by the same judges The grounded text extends or nuances Wikidata in every verdict; it never contradicts it, and never simply agrees. Multiplicity: another sense in 12 of 12. Talking drum: a category, not one hourglass drum, in 9 of 9.

Wikidata Q3860307, read 17 November 2026. Route D definition, 16 November 2026. Open comparison by the three judges, 17 November 2026; one run per cell.

Passages fit a small model’s window, at a price

Grounded runsFull texts allowedPassages only
Largest single call, median (tokens)46,13313,658
Runs whose largest call fits 32k8 of 3216 of 16
Est. cost, same 16 runs (USD)0.3770.121
Blind panel’s choice, 45 verdicts387
The price With passages only, all four models merged the Akan dondo with a Yoruba “dondon”: a short passage does not carry enough context to tell them apart.

Provider token counts. Full texts 16–17 November; passages (no whole texts, at most 12 tool calls) and panel 17 November 2026. One run per cell; panel uncalibrated; no open-weight model yet.

Three limits of the pilot

The judges know what the models know

They score from their own training, as the model alone writes from it, and no researcher has calibrated them. One marked down a 1962 spelling the model had quoted faithfully.

Keyword search finds a term only as written

Another language or spelling is missed unless the record carries it. Retrieval by meaning is not built yet.

Older sources, older words
6 to 7uses of “tribal” per 10,000 words, 1957 and 1962

Under 1 from 2007 to 2023, apart from one sparse year. Spellings date too.

“Tribal”: AMIRA MCP server, full-text map of 657 African Music articles, 16 November 2026. Every result in this talk: one run per cell.

What comes next in the pilot

1 About 50 concepts

Chosen by family: lamellophones, xylophones, drums, popular genres. Three runs per cell.

2 A local open-weight model

On the passages variant, which fits a 32k window: a 27-billion-parameter model, run locally.

3 Review by researchers

Subject specialists assess the drafts and calibrate the judges.

“Would you put this on the record?”
4 Retrieval by meaning

Behind the same server, so other languages and spellings are found.

Thank you

Reading the archive pays where it knows more than the models.

QR code linking to these slides online Scan for these slidesslides.frederickmadore.com

Appendix

Held back for questions.

What happens when someone asks

1 A person asks

In Claude, ChatGPT or another assistant: “What does AMIRA hold on highlife?”

2 The assistant picks a tool

From the tools’ descriptions: search, count, read a record.

3 The server answers

Records, counts and passages from the catalogue, each linked to its page.

4 The assistant writes

In its own words, ideally citing the records it received.

Who decides what We decide which tools exist and what they return. The assistant and its provider decide what reaches the answer.

Four models reading AMIRA on highlife

ModelTool callsFull texts readClaims with a quote foundEst. cost, list prices (USD)
Claude Haiku 5.51057 of 70.014
GPT-6 Luna1863 of 30.014
Gemini 3.5 Flash-Lite633 of 30.016
Mistral Small 41061 of 50.029
Read with care One run per model: anecdotes, not measurements.
Mistral Small 4: three quotes over the 25-word limit, one not found in the text it read.

Four definitions of highlife, read side by side

All four: an urban popular music of West Africa, chiefly Ghana and Nigeria, mixing African and Western forms.

ModelOriginWhat it addsQuotes found
Claude Haiku 5.5Debated: Konkomba singing bands, several streamsVernaculars; a national music by the 1950s7 of 7
GPT-6 LunaGhana, then across West AfricaVariation by place and period3 of 3
Gemini 3.5 Flash-LiteLocal forms with brass band, jazz and ballroom“Intertribal” music; love and social commentary3 of 3
Mistral Small 4Local forms with ballroom, brass band and Latin American musicHarmony in thirds or sixths; mid-twentieth century1 of 5

Route D, one run per model, 16 November 2026. Definitions of 80 to 102 words.