DeepTunisia
An open-source public-interest research project documenting Tunisia's political, institutional, economic and historical power structures through verifiable evidence. The goal is to make the relationships between people, institutions and events legible to citizens, journalists and researchers, while keeping verified information rigorously separate from reporting, inference and allegation.
What this is not
It is not a website exposing a deep state. That framing presupposes its own conclusion, and a project that begins from a conclusion cannot be trusted to test it.
The narrower question is answerable: which institutions, individuals and economic groups retain influence when the formal political regime changes, and which relationships are genuinely continuous rather than reconstructed after each rupture. That question has an answer the evidence can support or refuse, and on the evidence page you can see it doing both — two of the six hypotheses currently come out badly, including the strong version of the deep state thesis.
The rule that matters most
An inference is never allowed to become a fact. Every claim in the dataset is filed under one of four bases, and they never render identically:
- Documented — an official record states it.
- Reported — credible publications report it, with attribution.
- Inferred — nobody states it; it is reasoned from documented structure, and it must carry both the reasoning and what would refute it.
- Unsubstantiated — it circulates without reliable evidence. Recorded, never presented as evidence.
The last category is the one most projects get wrong. Deleting widely-held claims does not make them go away; it just means the map cannot show anyone why the popular account is mistaken. So they are kept, labelled, attributed to where they circulate, and paired with what the evidence actually shows. Several are contradicted by this project's own data.
Everyone gets the same standard
Bourguiba, Ben Ali, Saied, Ennahda, the opposition, the military, the police, the business families, the unions, the European Union, France, Italy, Algeria, the United States, Russia, the World Bank and international NGOs are all in the dataset under identical evidentiary rules. If the evidence cuts against a preferred narrative, it gets published anyway.
This is not neutrality for its own sake. It is the project's only real protection. Credibility that depends on people agreeing with the conclusions collapses the moment they don't. Credibility that depends on the evidence being accessible and reproducible does not.
Governance and editorial independence
- No funder may determine editorial conclusions, research targets, publication timing, personnel decisions, or how any claim is classified.
- Every donor is published below with amount, period, purpose and restrictions.
- The dataset is plain text under version control, so every factual change is a dated, attributable diff.
- Corrections are published, not quietly applied — every one of them is listed on what has changed, generated from that history rather than curated.
- No user accounts, no comments, no tracking. There is nothing here to compromise and nothing to leak.
Who funds DeepTunisia
| Donor | Amount | Period | Purpose | Restrictions |
|---|---|---|---|---|
| No external funding received to date. This table is published empty rather than omitted, because a transparency commitment made only after the money arrives is not a commitment. Any future grant appears here with its full terms before it is spent. | ||||
Current state, honestly
That review figure is deliberately unflattering. The project's stated architecture is that machines propose and humans verify, and at this stage almost nothing has been through a second pair of eyes. Publishing the real number is the only way that promise means anything.
Where that review went
A single percentage treats verifying a gazette-dated appointment as interchangeable with verifying an unsubstantiated allegation about a named living person. Those are not the same risk, and averaging them lets the figure look identical whether the effort went somewhere useful or somewhere safe. Broken out, it is worse than the aggregate suggests: every review so far has landed on the best-evidenced claims, and the records most capable of doing harm have had none at all.
| Claim type | Reviewed | Coverage |
|---|---|---|
| Unsubstantiated — circulates without reliable evidence | 0/7 | |
| Attributed — a named source's contested claim | 0/57 | |
| Inferred — this project's own reasoning | 0/2 | |
| Reported — credible publications, with attribution | 19/517 | |
| Documented — an official record states it | 8/136 |
How much of this exists in three languages
The interface is fully translated. What is written into the dataset itself — the one-line identifications, the summaries, the reasoning behind an inference, the statement of each hypothesis — largely is not, and until it is, an Arabic or French reader meets English inside a page that is otherwise theirs. Where that happens the text says so rather than passing silently as translated.
A translation is a claim about what the original says, so it carries its standing the same way every other claim here does. Model-reviewed means a model produced it and a second model pass checked it against the source; no person who reads the language has seen it. Only human means that. The two are reported separately and never added together, because a single "translated" figure would let unreviewed text read as finished work.
| Prose in the dataset | French | Arabic |
|---|---|---|
| Fields translated | 649/2141 | 649/2141 |
| Checked only by a model | 649 | 649 |
| Checked by a person who reads the language | 0 | 0 |
What the map does not contain
Every figure above measures the quality of records that exist. None of them can see a record that was never entered — you cannot flag the absence of a row. So this audit asks a different question of each president: which kinds of relationship has anyone recorded for them, and how many people are reachable through family ties alone.
The answer is the most uncomfortable number on this page. Coverage tracks when evidence became public, not who exercised power: the dataset is dense around a deposed president, whose clan was documented by courts and confiscation records after 2011, and thin around the sitting one. Left unstated that reads as a claim this project never intended to make — that the removed were corrupt and the incumbent is clean. It is instead a statement about archives.
| President | Ties recorded | Family network | No records of |
|---|---|---|---|
| Kais Saied | 21 | 3 | party |
| Zine El Abidine Ben Ali | 17 | 8 | political-conflict, reported-influence |
| Habib Bourguiba | 13 | 3 | — |
| Beji Caid Essebsi | 9 | 1 | reported-influence |
| Fouad Mebazaa | 5 | 0 | appointment, family, political-conflict, reported-influence |
| Moncef Marzouki | 4 | 0 | appointment, family, political-conflict, reported-influence, security |
| Mohamed Ennaceur | 3 | 0 | appointment, family, institutional, political-conflict, reported-influence, security |
Two former presidents currently have no recorded relationships at all. That is a hole in the map, not a finding about them. The categories listed are derived, not authored: a category counts as expected once this project has managed it for any one president, so the standard is only ever “this has been done here and not there”.
Language coverage
- English — interface 100% translated.
- Français — interface 100% translated. Analytical prose still English-only.
- العربية — interface 100% translated. Analytical prose still English-only.
Entity names, titles and institutions are stored in Arabic, French and English in the data itself, and all three are searchable — searching an Arabic name or an alternative transliteration finds the right record. Long-form analytical text has deliberately not been machine-translated: a machine-translated methodology page would undermine the whole point.
Roadmap
The knowledge graph is the product; the website is a window onto it. Layers are added in order of how much they depend on the graph being right first.
- Now — the graph, the visual system, the evidence architecture, structured search, open data export.
- Next — broader coverage, particularly the pre-2011 police chronology and the business-family map; a second reviewer.
- Then — an assisted research layer that answers questions by traversing this graph and citing records, never by generating prose.
- Later — continuous ingestion with human verification in the loop; international coverage of how Tunisia is discussed abroad; secure evidence submission.
- Eventually — research fellowships, journalism grants, an institutional home.
The ingestion pipeline is designed but deliberately not built yet. Discovery, extraction, entity resolution and contradiction detection are all straightforward; the part that actually determines whether the project is worth anything is the human verification step, and building the automation before the verification capacity exists would produce volume at the cost of the only thing that distinguishes this from the material it is meant to correct.
Corrections
If something here is wrong, the useful correction is a primary source: a decree number, a gazette reference, an official biography, a corporate filing. The open questions are the current research agenda, and closing one of them is the fastest way to improve this map. The open data page has the full graph, so anyone can check the reasoning rather than take it on trust. Everything that has already been changed is on what has changed.