Methodology

1. Overview

RAQEB is an open platform for monitoring online electoral discourse. It analyzes that discourse at scale across major platforms and public broadcast feeds: content that is clean and consistent with the source of truth flows into aggregate dashboards, while cases that warrant scrutiny — the risky, unclear, serious, or contested — are routed to trained reviewers for human verification against electoral law and editorial guidelines. Verified findings are published with full evidence and an audit trail.

This page describes how that works in detail — what we collect, how we analyze it, how we verify it, what we publish, and the principles that govern every step.

2. Foundations — how this methodology was built

RAQEB's methodology was not improvised. It translates established election-observation practice into a system that operates at the scale and speed of online discourse — the digital transformation of a process traditionally done by monitors reading content one item at a time.

Grounded in evidence

• International election-observation standards and the published methodologies of established monitoring bodies
• A review of global reports and academic literature on electoral integrity, disinformation, and media oversight
• Contextual and comparative analysis of the deployment's electoral history and information environment
• The electoral law and regulatory framework of each deployment's jurisdiction

Authored with domain experts

The framework was developed with input from specialists — not written by engineers alone:

• Legal researchers
• Media-oversight researchers
• Election experts

Codified into an inspectable system

The output of that work is not a document on a shelf. Detection criteria, aggregation logic, and routing policies are encoded as explicit, versioned rules that the system executes consistently on every item — the same standard applied at scale.

We publish what the methodology covers and why. We do not publish the internal detection criteria in a form that would let bad actors reverse-engineer and evade monitoring — a limitation we state openly rather than hide.

3. Data sources

RAQEB monitors content that is either publicly accessible or licensed for analysis. We do not scrape, do not collect private content, and do not track individual users.

Licensed platform content

• Major social platforms (Meta, X) via licensed content API access. Content is pulled in structured batches under the terms of the platform's content licensing agreement.
• Public WhatsApp channels — only the channels that are publicly accessible by design.

Public broadcast and press

• TV and radio broadcasts via public feeds and transcripts
• Online press articles
• Published opinion polls and official electoral commission data

Official electoral content

• Public statements, speeches, and posts from official accounts of candidates, parties, and electoral bodies
• Meta Ad Library and equivalent public political advertising disclosures

Submissions

• Observer reports and citizen submissions submitted through structured intake forms (with safeguards described under Privacy)

All sources, per deployment, are documented and disclosed openly. Each deployment publishes its own source list as part of its public methodology.

4. Per-deployment data isolation

Each adopting organization operates an independent deployment of RAQEB. The data ingested for one deployment is isolated from every other deployment.

What that means in practice

• A deployment's data is stored in a dataset dedicated to that deployment
• That dataset is not shared, mixed, or cross-queried with any other deployment's data
• The adopting organization's team has access to their deployment; access controls are configured by them
• The shared methodology governs how data is analyzed and published; it does not pool the data itself

This isolation is structural, not just policy. It ensures each deployment's adopting organization controls its own monitoring environment under its own oversight.

5. AI detection — what the models do

RAQEB uses AI models to surface signals from the volume of monitored content. The role of the AI is detection and scoring, not judgment. Models do not publish findings.

What the models are trained on

• The electoral law and regulatory guidance of the deployment's context (silence-period rules, campaign-period rules, disclosure requirements, broadcaster obligations)
• Editorial framing and language guidelines used by RAQEB reviewers (how to describe potential issues without overclaiming)
• Known patterns of electoral issues: silence-period violations, coordinated narrative amplification, manipulated media, sponsored content lacking disclosure, false claims about voting procedures
• Linguistic context across Arabic (including its dialects), English, and French (with additional languages added per deployment as needed)

What the models do

• Classify content against the trained categories
• Assign a confidence score to each detection
• Cluster related content (e.g., coordinated posting patterns, narrative repetition)
• Surface candidates for human review

What the models do not do

• Decide whether a finding is published
• Assign legal labels (no "FALSE," no "BREACH," no fixed verdict labels)
• Replace human judgment at any stage

6. Actor registry and attribution

Every signal RAQEB surfaces is anchored to a specific, known actor. This is what separates a signal about a real candidate, party, or outlet from a loose keyword match.

The registry

• A curated registry of the candidates, parties, and media outlets relevant to each deployment
• Multiple name forms per actor — official names, common variants, transliterations, and aliases — so an actor is recognized across spellings and languages
• Arabic- and dialect-aware matching, so names resolve across dialectal and transliterated forms
• Verifiable relationships between actors (party affiliation, outlet ownership) where relevant

Attribution rigor

Anchoring to an actor is only half the discipline. The harder question is always: whose statement is this?

• Content is attributed to the speaker — the account or outlet that produced it — not to whoever is merely named in it.
• When a post criticizes or discusses a candidate, it is attributed to its author. A candidate is never associated with claims made against them.
• When an outlet quotes or reports a person's statement, the statement is attributed to the person who made it; the outlet is recorded as reporting it, not as making the claim.
• Mentions, quotes, and authorship are tracked as distinct relationships — never collapsed into one.

This distinction is enforced in how signals are aggregated, so that what is said about an actor is never miscounted as something the actor said.

7. Confidence scoring

Every AI detection is assigned a confidence score. The score reflects the model's certainty that the content matches the pattern it was trained to detect — nothing more.

The score is one input into a routing decision — not the decision itself. Issue type, severity, and a source-of-truth check also determine whether a case needs a human at all. Content that is clean and consistent with the source of truth is not sent to human review; it flows into aggregate analysis. What is risky, unclear, serious, or contested is routed to a reviewer, regardless of how confident the model was.

Confidence thresholds are configurable per deployment and per issue type. They are documented openly in each deployment's published methodology.

8. Escalation routing — multi-dimensional

Not every item needs a human. The first routing question is whether a case requires review at all:

• Content that is clean and consistent with the source of truth flows straight into aggregate analysis — no human review
• Content that is risky, unclear, serious, contested, or tied to results and oversight is routed to a reviewer

This is what keeps the system fast and focuses reviewer time where it matters. When a case does need review, it is routed to a verification track based on four dimensions, not a single severity score:

Routing tracks include

• Standard review — single reviewer applies framing guidelines, confirms or dismisses
• Dual review — two reviewers must independently confirm before publication
• Senior reviewer escalation — sensitive or high-severity findings go to a senior reviewer
• Editorial board escalation — findings with legal implications or potential public-interest weight are reviewed by the editorial board before publication

This multi-dimensional routing is the practical mechanism that ensures no flagged finding is published without proportionate review.

Aggregate analysis and the dashboards

Most of what RAQEB monitors is not a flagged finding — it is ordinary discourse. That content is aggregated into the public dashboards, which summarize patterns across the full monitored corpus: who is covered and how much, on what topics, in what framing and tone, and how narratives emerge and move over time.

These views are computed from the same per-item AI analysis, aggregated across the corpus — not from hand-picked examples. Most content flows into this aggregate layer directly, and only a minority is escalated for human review. (In the Lebanon pilot, roughly three-quarters of content routed straight to aggregation rather than the reviewer queue.)

9. AI governance

RAQEB treats its use of AI as something to be governed, not just deployed. The following controls apply across every deployment.

No automated decisions

• No flagged finding — a suspected breach, a misinformation incident, a serious allegation — is published on the strength of an AI output alone; a human reviewer confirms or dismisses it first. (Aggregate analytics summarize patterns at scale and are not individual findings.)
• The AI cannot assign a legal label, close a case, or publish. Its role ends at surfacing and scoring.

Human oversight and feedback

• Reviewers can dismiss any AI detection; dismissals are logged and feed back into refining the detection criteria over time.
• Detection quality is monitored per category; when quality on a category degrades, its criteria are revised.

Consistency and fairness

• Detection criteria are developed and validated against labeled, held-out, and adversarial test batches before and during deployment.
• The same criteria are applied regardless of the political affiliation of the actor involved.
• Coverage-balance metrics are descriptive — they surface how much, and how, actors are covered. They are signals for review, not a ruling on bias.

Stated limitations

AI detection is probabilistic: it produces false positives and can miss content. That is precisely why human verification — not the model — is the gate on every flagged finding that is published. Governance here is continuous, not a one-time certification.

10. Qualified language — by design

RAQEB does not adopt fixed verdict labels (FALSE, MISLEADING, BREACH, etc.). This is a deliberate editorial discipline that protects both accuracy and liability.

Instead, findings are described in qualified language that reflects what the evidence actually supports:

This is not hedging. It is editorial accuracy. RAQEB is not a court, an electoral commission, or a fact-checking authority that issues definitive verdicts. We document, verify, and publish — leaving determinations of legal breach to the bodies empowered to make them.

When a finding is later confirmed, updated, or refuted, the public record is updated and the change is logged.

11. Reviewer guidelines

Trained reviewers operate under documented guidelines covering:

Framing

• How to describe a finding without overstating its meaning
• How to attribute claims to their source rather than making them in RAQEB's voice
• How to handle content involving named individuals with appropriate care

Verification

• How to cross-check a claim against original sources
• How to confirm the authenticity of media (image metadata, broadcast logs, platform-level timestamps)
• How to identify coordinated behavior versus organic amplification
• When to escalate rather than confirm or dismiss

Language

• Use qualified language at all stages of documentation
• Use the source's own words when quoting; do not paraphrase in a way that changes meaning
• Apply the same standard to all sides, regardless of political affiliation

Dismissal

• When a detection does not meet the verification standard, it is dismissed and logged
• Dismissals are tracked to refine the AI's training over time

Senior reviewers handle escalations and apply heightened care to sensitive findings. An editorial board reviews findings with broader legal or public-interest implications.

12. What gets published vs. what is held

Not every verified finding is published immediately. Publication criteria:

Published

• Findings confirmed by the required number of reviewers
• Findings mapped to a documented issue type
• Findings supported by evidence accessible to readers

Held

• Findings where verification is incomplete
• Findings where evidence cannot be made public without compromising a source
• Findings under editorial review for sensitivity or legal implications

Never published

• Personally identifying information beyond what is already public
• Private content (private messages, non-public posts, restricted-access material)
• Submissions without sufficient evidence to support qualified language

Holds are logged internally. They are not concealed — adopters can review what was held and why.

13. Audit trail

Every published finding carries a public audit trail that includes:

• Source — the original content, with a link or archived reference
• Detection details — when it was surfaced, what AI track flagged it, the confidence score
• Review chain — which reviewers handled it, what verification steps were applied, escalation history if any
• Editorial decisions — any language changes between draft and publication, with timestamps
• Updates — any revisions made after publication, with the reason and date

The audit trail makes RAQEB's process inspectable. A reader, an adopter, a journalist, or an electoral body can trace any finding from public record back to its original signal.

14. Privacy principles

RAQEB monitors public discourse, not individuals. Our privacy principles:

What we do

• Aggregate analysis of public content
• Pattern and trend detection across volumes of posts
• Narrative clustering and amplification analysis
• Verification of public claims against evidence

What we do not do

• Track individual user behavior over time
• Collect personally identifying information beyond what is already public
• Publish private account details, personal communications, or restricted content
• Use monitoring data for any commercial purpose

Submissions and sources

• Citizen and observer submissions are handled with care for source protection
• A submitter's identity is not published unless they explicitly choose to be named
• Sensitive submissions are routed to senior reviewers; submitters are informed of how their report will be handled

Data retention

• Each deployment publishes its own data retention policy as part of its methodology

15. Non-partisanship

RAQEB is non-partisan. The methodology applies the same standard to every actor monitored — candidates, parties, broadcasters, official accounts, independent commentators — regardless of political position or affiliation.

The platform does not endorse, oppose, or recommend any candidate, party, or political position. Findings represent data-driven documentation of what occurred in the monitored discourse, mapped against the regulatory and editorial standards of the deployment.

Adopting organizations operate under the same non-partisanship principle. Eligibility criteria for adoption require organizations to operate independently of political parties and to commit to monitoring all actors under the same standard.

16. Limitations and caveats

RAQEB is rigorous, but it is not infallible. Honest limitations:

• AI models miss things. Detection coverage is high but not complete. Some issues will surface in the discourse without being flagged by the models.
• Human review can disagree. Two reviewers may reach different conclusions on the same finding. Dual-review and editorial escalation mitigate this but cannot eliminate it.
• Findings are not legal judgments. Whether something constitutes a breach of electoral law is a determination for the relevant electoral authority or court — not for RAQEB. We document; they decide.
• Findings may be updated. As evidence develops, a published finding may be revised, expanded, or — rarely — withdrawn. The change is logged publicly.
• Coverage is bounded by data access. RAQEB monitors what is publicly accessible or licensed. Content on closed platforms or in private channels is outside our scope by design.
• No real-time guarantee. AI detection and human verification take time. RAQEB is fast, but not instant. Some patterns become visible only with hours or days of accumulated signal.

Stating limitations openly is itself part of the methodology.

17. Continuous research and development

RAQEB runs on a shared methodology — one that is developed, maintained, and continuously improved by SmartGov, the team behind the platform. It is not a fixed specification frozen at launch; it is actively researched and refined.

What continuous improvement covers

• Detection logic and rules — improving accuracy, reducing false positives, and closing coverage gaps as they are identified
• Speed and efficiency — shortening the time from signal to review without lowering the verification standard
• Robustness and new capabilities — extending what the platform can monitor and how reliably it does so
• Emerging threats — adapting to new manipulation tactics as they appear, including synthetic and manipulated media (deepfakes), coordinated-behavior techniques, and evolving evasion methods

The information environment changes quickly and adversarial tactics evolve. Continuous research and development is how the shared methodology stays adaptable to what comes next — rather than locked to the tactics of today. Improvements are propagated across deployments, while each deployment's data remains isolated as described above.

18. Update policy

Published findings may be updated when new evidence becomes available.

• Minor corrections (typos, source link updates) are made without notice; the change date is logged.
• Substantive revisions (changes to the language of the finding, additional evidence, reassessment) are flagged with a visible "Updated" note and a brief description of what changed.
• Withdrawals (a finding that the editorial board determines should not have been published, or that has been refuted by subsequent evidence) are flagged with a visible "Withdrawn" note. The original finding remains visible with a strikethrough and a link to the explanation.

The update history of every finding is preserved. RAQEB does not silently edit the record.

19. Contact and corrections

Anyone affected by a published finding — a candidate, party, broadcaster, account holder, or third party — may submit a correction request.

• Correction requests are reviewed by a senior reviewer and, if substantive, by the editorial board.
• The submitter receives a response within a published timeframe (set per deployment).
• Substantive corrections trigger an update to the public record as described in the update policy.

Contact channels

• Corrections and disputes: corrections@raqebinitiative.org
• General inquiries: contact@raqebinitiative.org
• Press inquiries: press@raqebinitiative.org
• Methodology questions: methodology@raqebinitiative.org