Skip to content
Investigative Journalism
ARKA
Unclessify — Journal of Investigation and Declassification, Founded by Graziano Costantino
ARKA

Unclessify's investigative engine

ARKAARKA — Advanced Reporting Knowledge Architecture

Advanced Reporting Knowledge Architecture

The Ark of knowledge: it traces every sentence back to the record that supports it, counts sources by issuer and delivers the evidence together with the story.

What it does, and why it is different

A search engine finds those who comment on a record. An automatic writing tool summarises what it finds, and invents dates, names and numbers with the same confidence. ARKA is built the other way round: it starts from the primary record, keeps it with the fingerprint of its bytes, extracts only what can be found word for word, traces the records that record cites, and counts how many truly independent voices confirm each fact.

The output is not a text: it is an evidence register, from which come the investigative story and the case file that proves it. If a sentence has no evidence, it does not go out.

Measured, not declared

measured on 2026-10-19, at every publication

investigations published with ARKA
52
investigations published with ARKA
documents archived with a SHA-256 fingerprint
801
documents archived with a SHA-256 fingerprint
institutional domains of origin
52
institutional domains of origin
passages quoted word for word
5148
passages quoted word for word
distinct issuers in the registers
84
distinct issuers in the registers
factual sentences anchored to a record
100%
factual sentences anchored to a record

Architecture

Twelve stages, in the order the engine runs them

ARKA is not a search tool with a writer on top. It is a chain of stages in which each one produces a verifiable object — an archived document, a passage, a claim, a register — and the next stage can only use what the previous one proved.

  1. 01Acquisition
  2. 02Multiplication
  3. 03Evidence
  4. 04Writing and checks

Acquisition

  1. 1

    The seed

    A primary record: a court ruling, a parliamentary report, a decision by an authority, a declassified collection. ARKA reads it page by page, even when it is a scan.

    It always starts from a document, never from a topic.

  2. 2

    Chain of custody

    Every document is downloaded from the official source, respecting its access rules, and logged with its address, date, channel and the SHA-256 fingerprint of its bytes.

    The copy we read is the copy we cite: a different copy has a different fingerprint.

  3. 3

    Linked records

    From the record's page ARKA collects the linked documents on the same institutional domain (files, texts, annexes, including those the site opens through scripts) and acquires them with the same chain of custody.

    Context comes from the institution that issued the record, not from those who comment on it.

Multiplication

  1. 4

    The atoms

    In the text it recognises the strings that identify other records — ruling numbers, laws, decrees, EU regulations — with their exact position.

    Every reference has a verifiable address inside the record.

  2. 5

    Cited records, from their publisher

    For each atom it rebuilds the official address at the publisher of the record (the Constitutional Court, the national legislation database, EUR-Lex) and acquires it. One channel at a time, because distinct issuers are what count, with sampling that learns which channels pay off.

    A cited record is kept only if its text really contains the number and year of the reference.

Evidence

  1. 6

    Claims, word for word

    From each record it extracts verifiable claims, each with the passage that supports it.

    A claim stays only if its quotation is found word for word in the archived document.

  2. 7

    Primacy from structure

    It separates what a record decides (operative part, reasoning of the issuing body) from what it reports (a party's argument, the content of another record).

    A party's argument never becomes a fact: it stays attributed to whoever makes it.

  3. 8

    Corroboration

    It counts sources by issuer and by the routes they come from, and gives every claim a state and a level (the equations are below).

    Ten documents from the same office count as one source.

  4. 9

    The evidence register

    Claims, literal passages, issuers, routes, document fingerprints: a single register, with its own fingerprint.

    Everything published comes from the register and can be traced back.

Writing and checks

  1. 10

    The story and the case file

    Two sections come from the register: the investigative story, in which every factual sentence contains the passage that supports it and names the body that states it, and the evidence file, generated from the register, also as a PDF. The piece is published in Italian and English.

    Readers see the evidence next to the story, and can retrace the path.

  2. 11

    Checks that stop a piece

    Sentence coverage, copying from the records, attributions, people and minors, republication rights, images, length measured on the served page.

    If a check fails, the piece does not go out, and the reason is written down.

  3. 12

    Sign-off and publication

    The sign-off is recorded on the fingerprint of the text that was read. Every publication has a restore point, a change log and a test on the page readers receive.

    If the text changes after sign-off, the sign-off lapses.

The model

How a fact earns its place: the equations

These are the definitions the engine applies to every claim, written as they work. The thresholds and internal weights stay confidential; the logic does not, because a method that cannot be checked is not a method.

(1)

The sources that hold

R(a) = { p ∈ P(a) : p verified ∧ ¬ p contradicts }

For a claim a, only verified passages count — found word for word in the archived document — that do not contradict it.

(2)

Issuers and routes

O(a) = { issuer(p) : p ∈ R(a) }

V(a) = { route(p) : p ∈ R(a) }

O is the set of distinct issuers (seven rulings from the same chamber are one issuer); V is the set of independent routes the records arrived by.

(3)

Degree of corroboration

C(a) = |O(a)| if |V(a)| ≥ 2, otherwise min(|O(a)|, 1)

Several issuers count only if they arrive by different routes: through a single route, whatever it is, C is at most 1. In the case file C is published as a pair ⟨|O|, |V|⟩, so readers see where it comes from.

(4)

The state

contested ⇔ ∃ p : p verified ∧ p contradicts

confirmed ⇔ C(a) ≥ 2 ∨ ∃ p ∈ R(a) : primary(p)

attributed ⇔ C(a) = 1

unverified otherwise

A claim is confirmed if several issuers support it through different routes, or if the record itself establishes it in its operative part. Otherwise it is attributed, and the sentence says who states it.

(5)

The level shown to readers

corroborated across routes ⇔ |V(a)| ≥ 2 ∧ |O(a)| ≥ 2

confirmed by record otherwise

The two kinds of confirmation are never confused: a fact reported by several independent issuers is one thing, a fact established by a record is another.

(6)

Sentence coverage

κ(A) = |{ f ∈ F(A) : ∃ b ∈ B, b ⊆ f }| ⁄ |F(A)| = 1

Every factual sentence f in story A contains a passage b from the register. A piece with κ < 1 does not go out.

(7)

No copying of the record

f rejected ⇔ Gₙ(f ∖ «…») ∩ Gₙ(Π) ≠ ∅

If a sentence, leaving aside declared quotations, repeats n consecutive words from the record's passages Π, it is rejected and rewritten: records are quoted, not copied.

(8)

Fingerprint and sign-off

h(d) = SHA-256(bytes(d))

sign-off valid ⇔ h(T published) = h(T read)

Every document carries the fingerprint of its bytes; a sign-off is valid only for the text it was given on.

(9)

Cited records and channel yield

x accepted ⇔ number(x) ∈ text(x) ∧ year(x) ∈ text(x)

θc ∼ Beta(αc, βc)

A cited record is kept only if its text contains the reference. Each channel has a yield updated with every record accepted or rejected, and ARKA chooses where to search by sampling that yield.

What the registers say today

4987 claims in 50 registers, 5148 literal passages, 180 of them from the operative part of a record, no contradictions. Most claims are attributed: ARKA does not dress up what a single authority states as a confirmed fact.

  • attributed — the record states it, and the sentence names who4889
  • confirmed by the record's operative part93
  • corroborated by independent issuers across routes5

What users receive

  • The evidence register: claims, literal passages, issuers, routes, fingerprints
  • The investigative story in two sections, in Italian and English
  • The PDF case file generated from the register, with the register's fingerprint
  • The list of records to request when the archive is silent (freedom of information, official copies)
  • The history of every piece: checks passed, sign-off, restore point

What stays confidential

The logic is public, so it can be checked. These stay confidential:

  • the internal thresholds and weights of checks and yield
  • the engine's working instructions and code
  • the list of channels and how they are queried
  • the internal structure of the register and the case file

The principles

01

Every fact has an address

Every statement in a piece is tied to the page of the record it comes from. Whatever cannot be anchored to a record stays out of the text: at most, it becomes a question.

02

Sources are counted by issuer

Ten documents from the same office are one source. ARKA measures how many truly independent voices confirm a fact, and says so when there is only one.

03

Chain of custody

Every document acquired is logged with its provenance, date and an integrity fingerprint: the copy we read is the copy we cite, and anyone can check it.

04

Nodes, not summaries

ARKA does not summarise what is already known. It looks for what is missing between one record and the next: contradictions, gaps, links between proceedings that nobody has lined up.

05

Where the archive is silent, we ask

When a record is not public, ARKA prepares the requests to the institutions holding it — official copies, freedom-of-information requests — in the form the law provides.

06

The editors have the last word

ARKA works for a journalist, not in their place. Every piece is reviewed and signed by the Unclessify editorial team, edited by Graziano Costantino, who answers for it.

Where you see it at work

In every article the “How this was made” box states which record the piece starts from; below the text, the case file lists the evidence one by one, with its address. If you find an error, write to us: correction is part of the method.

Case studies

Investigations published with ARKA. In each, every factual sentence points to a piece of evidence in the case file, with the exact words of the record and its source: the simplest way to see what the method does.

Restricted access

ARKA under licence: product sheet, technical sheet and demo

ARKA is licensed to journalists, publishers and newsrooms, research bodies and universities, public administrations and investigative bodies: as a service run by Unclessify or installed on your premises, quoted separately on your volumes. Unclessify checks every request — identity, accreditation, purpose — and may decline it. It is not licensed to private individuals.

  1. 1Fill in the form: every field is required.
  2. 2We check the request and send you the product sheet, the technical sheet and the code for the restricted demo.
  3. 3If we go ahead, we agree terms, use and any quote for on-premises installation together.
Who you are
How you would like to use it

Installation on your premises is quoted separately, based on your volumes and your infrastructure.

Unclessify checks every request. If it is accepted, you receive the product sheet and the technical sheet as attachments at the address given, with a personal code for the restricted demo.

ARKA is a system designed by Graziano Costantino for Unclessify. All rights reserved. A request does not entitle you to a licence; the data in the form are processed as described in the privacy policy.