Open data

Take the graph

The website is a window; the knowledge graph is the actual project. It is published in full so the reasoning can be audited rather than believed. Free to use with attribution — if you build on it, cite DeepTunisia and the underlying sources.

dataset.json The complete graph: entities, positions, relationships, events, sources, gradings. JSON · resolved intervals · ~320 KB positions.csv Every officeholding as a flat row. The most useful file for spreadsheets and stats. CSV · 417 rows relationships.csv Every edge with type, basis, attribution and source ids. Loads straight into graph tools. CSV · 231 rows sources.csv Every citation with tier and archive lookup, so links can be re-checked. CSV · 653 rows

How to read a record

Two fields carry most of the epistemic weight. basis says what kind of claim this is: documented, reported, inferred or unsubstantiated. The interval carries four date fields rather than two, because historical personnel records are rarely precise and collapsing them would invent certainty. raw preserves exactly what was authored, so you can see the difference between a decree date and an estimate.

{
  "id": "p-dgsn-saidane",
  "role": "dg-national-security",
  "holder": "mourad-saidane",
  "basis": "documented",
  "datesInferred": false,
  "interval": {
    "startEarliest": 1653955200000,
    "startLatest":   1653955200000,
    "endEarliest":   1759276800000,
    "endLatest":     1785110400000,
    "startPrecision": "day",
    "status": "last-verified",
    "raw": { "start": "2022-05-31", "end": "verified:2025-10-01" }
  },
  "review": { "by": "Initial compilation", "date": "2026-07-26",
              "method": "Checked directly against the JORT decree text" },
  "sources": ["jort-2022-546", "lapresse-saied-securite-2025"]
}

Here status: "last-verified" means the officeholder was confirmed in post on 1 October 2025 and may well have continued. It does not mean they left. That distinction is the difference between a dataset and a guess, and it is why the Chronicle draws that bar with a soft right edge.

Entity model

Type	Purpose	Key fields
Source	A citation. Everything else points at these.	id, title, publisher, date, url, archive_url, tier (1–5), lang, excerpt
Institution	Ministries, services, parties, companies, families, foreign states, organisations.	id, name_en/fr/ar, abbr, type, layer, parent, active interval, sources
Role	A canonical office. Normalising titles kills transliteration drift.	id, title_en/fr/ar, institution, authority (published weight table)
Person	Names in three scripts plus alias spellings, for entity resolution.	id, name_en/fr/ar, aliases, birth, death, layers, trajectory, basis, sources
Position	The join that makes time-travel possible: person × role × span.	id, role, holder, interval (4-field fuzzy), basis, datesInferred, reasoning, falsifiable_by, disputes, review, predecessorDerived, successorDerived, sources
Relationship	A typed, dated, graded edge between any two entities.	id, from, to, type, subtype, interval, basis, attributed_to, reasoning, falsifiable_by, disputes, sources
Event	Dated events, with competing characterisations held side by side.	id, date, title_en/fr/ar, category, rupture, actors, institutions, contested, disputes, sources
Era, Question, Hypothesis	The analytical layer: periodisation, open research tasks, falsifiable claims.	era: interval, thesis, formula · question: kind, status · hypothesis: support, reasoning, falsifiable_by

Caveats you should carry with the data

215 records await a primary source. They are marked. Do not cite an inferred span as a date.
7 contradictions are unresolved. Where sources disagree, the dataset records the disagreement instead of choosing. Check the disputes array before using a figure.
The pre-2011 police chronology is the weakest area. The sequence of national police chiefs under Ben Ali rests largely on secondary sources.
The authority weights are editorial. They are a judgement about formal authority, published on the method page so you can substitute your own.
Only 27 of 719 records have been through independent human review. Treat the rest as compiled but unaudited.

Rebuilding it yourself

The source of truth is YAML, not this JSON. The build validates every reference, rejects any claim without a source, rejects an inference without stated reasoning and a falsifier, rejects an unattributed low-confidence claim, and derives all succession chains from the position records so gaps surface instead of being smoothed over. If it builds, the referential integrity holds.