About this site
Disclaimer, methodology, limitations, objectives and reading guide.
Disclaimer
This site is NOT a journalistic news site.
It is an AI-assisted computational investigation project, intended for research and transparency purposes. The content presented includes hypotheses, speculations and unverified leads, systematically identified as such through explicit confidence levels.
What this site is
- ✓ An exercise in structuring and analyzing information
- ✓ A curated catalog of theories with credibility scoring
- ✓ An educational tool for critical thinking
- ✓ An open source project, transparent about its methods
- ✓ A database for exploring public documents (FOIA)
What this site is NOT
- ✗ A site that asserts anyone's guilt
- ✗A conspiracy theory site — it catalogs and evaluates them
- ✗ A substitute for judicial or journalistic investigation
- ✗ A reliable primary source for accusations
- ✗ A partisan or activist project
The mention of a person on this site in no way implies they are guilty of anything. The implication scores (IIE) measure the documentary density of links to the case, not a degree of guilt. Only the justice system is authorized to establish criminal responsibility.
The approach
The starting observation
The Epstein case has become the crossroads where all conspiracy theories converge. With over 1.3 million documents declassified by the DOJ under the Freedom of Information Act, it is a case where the following coexist:
Convictions, financial transactions, correspondence, sworn depositions, flight logs.
Documented connections whose meaning remains ambiguous. Correlations without established causation.
Theories without factual basis, conflations, magical thinking, sometimes vectors of antisemitism or violence.
The problem is not that people are interested in the Epstein case — it's that no tool helps them distinguish fact from speculation. Traditional fact-checking tools are binary (true/false) and fail to capture the gradient of credibility or the mechanisms by which legitimate curiosity transforms into conspiracy belief.
The approach: multi-agent computational investigation
This project takes an unprecedented approach: applying the methods of computational analysis and multi-agent artificial intelligence to a massive documentary corpus, with methodological rigor borrowed from both investigative journalism and academic research.
Massive corpus processing
OCR, named entity recognition (NER), indexing — transforming 1.3M raw documents into a structured and searchable database of over one million entities.
Systematic theory cataloging
200 theories identified, classified into 10 thematic categories (geopolitics, intelligence, elite networks, child exploitation...), each with evidence for, against, and a calculated credibility score.
Systematic fact-checking
135 fact-checks conducted on 61 theories, including 16 in Red Team mode (adversarial challenge) on confirmed theories. Result: 63% have a confirmed factual core, only 2% are formally disproven. The dominant pattern is the mixing of fact and interpretation.
Network analysis and cross-referencing
Construction of a graph with 2,710 nodes and 53,847 edges linking actors, documents and theories. Community detection, bridge actors, and connection patterns. Systematic cross-referencing between contact books, flight logs, call slips and depositions.
Detection and analysis of redacted zones
52,598 redaction zones identified textually + 3,684 detected visually by OpenCV. 65.8% of redacted zones mask people's names. The FOIA exemptions used (national security, personal protection) are analyzed for each document.
Transparent presentation
This site: presentation of all data, analyses and theories with all confidence levels displayed, explicit limitations, and educational modules to equip the reader.
Guiding principles
- ◆ Gradation, not binary — Instead of "true/false", a 6-level credibility spectrum with numerical scores. The world is more complex than a toggle.
- ◆ The process is the product — Transparency about the method is as important as the conclusions. Showing how we evaluate, not just what we conclude.
- ◆ Cataloging ≠ validating — Documenting a speculative theory does not mean endorsing it. Understanding why a theory circulates is itself valuable information.
- ◆ Built-in self-correction — Errors are tracked, agents recalibrated, scores adjusted. The system is designed to make mistakes, recognize them and improve.
What sets this approach apart
- ◆ Contradictory multi-agents — 8 AI agents with different roles that mutually challenge each other, reproducing the functioning of a diverse newsroom.
- ◆ Processing at scale — 1.3M documents, 1M+ entities, 200 theories, 684 actors — a volume no single human investigation could cover.
- ◆ Memeplexes and sociology — Beyond fact-checking, analysis of propagation mechanisms: why certain theories cluster together, how they mutually reinforce each other.
- ◆ Educational dimension — Not just "here are the facts" but "here is how to evaluate claims for yourself". Reader empowerment, not dependence on our authority.
Project objectives
1. Structure information
Given the immensity of the corpus (1.3 million FOIA documents), the primary objective is to catalog, classify and connect information systematically, rigorously separating documented facts from hypotheses and speculation.
2. Methodological transparency
Every claim is sourced, scored and challenged. Potential AI agent biases are tracked, errors cataloged, and confidence levels explicitly displayed. The process is as important as the result.
3. Critical thinking education
The site serves as a pedagogical laboratory for learning to distinguish facts, correlations, interpretations and speculations. Educational modules provide concrete critical thinking tools.
4. Computational investigation
Exploring the possibilities of AI-assisted analysis applied to a massive documentary corpus: NER, redaction detection, automatic cross-referencing, network analysis, and multi-criteria scoring.
Methodology
Processing pipeline
Rigor protocols
- 01 Dual evaluation — Each conclusion is evaluated by at least 2 independent AI agents with cross-perspectives.
- 02 Confidence decomposition — For any score > 0.60: factual basis, interpretation, causal inference and source diversity are evaluated separately.
- 03 Error tracking — Errors are cataloged by type (attribution, distortion, chronology...). >3 errors of the same type = flagged systematic pattern.
- 04 Anti-confirmation — Pre-session protocol: listing biases to monitor, falsification criteria, blind spots identified before analysis.
- 05 Calibration — Agents are regularly recalibrated on their prediction accuracy (Michel 0.67, Herodotus 0.79, Vidocq 0.64).
Capping rules
Weak causal inference
If causal inference is < 0.30, the overall score is capped at 0.70 maximum, regardless of the strength of other components.
Hypothesis chain
If more than 5 unverified hypotheses in the reasoning chain, the whole chain is flagged as "fragile chain".
Adversarial audit
Any theory > 0.60 must be challenged by a non-author agent in an independent epistemic audit.
Probability scales
All site content is framed by explicit scoring systems. A high score does not mean "true" — it indicates dense and coherent documentary support.
Theory credibility levels
Confirmed by official documents, trials, recognized investigations. Multiple independent primary sources.
Some elements verified by independent sources, others remain unconfirmed. Coherent but incomplete.
Consistent with known facts, but lacking direct evidence. Reasonable interpretation but unproven.
Based on interpretations, coincidences, or unverified testimony. To be treated with great caution.
Refuted by solid evidence, but persists in the collective imagination. Cataloged to understand its persistence.
No factual basis, magical or paranoid thinking. Included only for the study of conspiracy mechanisms.
Credibility score (0.0 — 1.0)
Each theory receives a numerical score between 0 and 1 that synthesizes all elements:
Current corpus average: 0.54 — the majority of theories fall in the grey zone.
Investigative confidence levels
Used in analyses and reports:
Theory statuses
Each theory follows a lifecycle:
Established documentary support
Investigation in progress, partial evidence
Factual elements + uncertain interpretation
Hypothesis without sufficient support
Refuted by contrary evidence
IIE — Epstein Implication Index (0–100)
The IIE measures the documentary density of links between an actor and the Epstein case. This is NOT a guilt score. A high IIE means the person appears frequently in the document corpus.
Private flights, routes to the Virgin Islands, co-flights with victims/inner circle
Contact book, MASSAGE section (critical signal), phone calls
Number of corpus documents where the actor appears (crossref)
6 levels: operator (5), facilitator (4), beneficiary (3), enabler (2), associate (1), contextual (0)
Depositions, testimonies, Maxwell proffer, fact-checks
Temporal span, contacts before/after pivot dates (2008, 2019)
Threshold: IIE ≥ 50 = significant documented implication. The score is not modified by redaction (censorship) data — these are complementary, not constitutive.
Multi-agent AI system
The investigation is conducted by 8 specialized AI agents, each with a role, identified biases and a calibration score. This system aims to reproduce the cross-checks of a journalistic newsroom.
Cataloging, classification and credibility scoring of theories. Identification of inter-theory connections and memeplexes.
Documentary corpus analysis, relational networks, pattern detection, anomalies and cross-referencing.
Fact-checking, debunking, source tracing. Michel's counterpart — systematic theory challenging. Red Team mode available.
Historical contextualization, precedents, chronologies, period geopolitics. Anachronism detection. Calibration: 0.79.
Psychological profiles, power dynamics, grooming, coercive control, victimology. Calibration: 0.64.
Popularization, interactive exercises, critical thinking, educational module design.
Social class analysis, belief systems, geography of power, group dynamics.
Inventory of belief systems (scientific, pseudo-scientific, religious, ideological), epistemic status, intellectual genealogies.
Documentary sources
Primary corpus
DOJ-OGR Archives (FOIA)
1,330,303 documents released under the Freedom of Information Act by the US Department of Justice.
EFTA DataSets 1–12
Structured datasets (DS1-DS12) including depositions, financial documents, correspondence.
MediaSet 9 (DS9)
Audio/video transcriptions: depositions, covert recordings, interviews.
Extracted data
Epstein's Black Book — including sensitive sections (MASSAGE)
Private jet records — passengers, destinations, frequencies
Message slips — caller, date, subject, urgency
Proffer and sworn depositions of Ghislaine Maxwell
52,598 redaction zones detected textually + 3,684 detected visually
249,587 entities (100,895 persons, 126,226 organizations)
Known limitations
AI limitations
- •AI agents can hallucinate — producing plausible-sounding but inaccurate connections or details.
- •NER (named entity recognition) can confuse homonyms or miss entities.
- •OCR on scanned documents produces artifacts that can distort the analysis.
- •Agent calibration scores (0.64 to 0.79) indicate a significant margin of error.
Methodological limitations
- •The Protection Index can generate false positives (generic entities detected in redacted zones).
- •Redacted content predictions are contextual estimates at [LEAD] or [HYPOTHESIS] level.
- •22 active theories have no documented key_actors, making the redaction signal impossible to calculate.
- •Documentary density (IIE) is not proof — an actor can be frequently mentioned for legitimate reasons.
Source limitations
- •FOIA documents contain massively redacted zones (28.7% exemption b7C, 25.3% b6).
- •The "Black Book" is a contested document whose complete authenticity is not established.
- •Sworn testimonies may contain deliberate lies or omissions.
- •The analysis does not cover the entire corpus — not all 1.3M documents have been processed.
Identified biases
- •Confirmation bias: tendency to seek connections that validate an existing theory.
- •Salience bias: well-known names attract more attention than obscure but potentially important actors.
- •Complexity bias: tendency to favor complex, interconnected theories over simple explanations.
- •Availability bias: widely covered theories are better documented, not necessarily more credible.
User guide
How to read this site
Always check the credibility level
Each theory displays a colored badge (DOCUMENTED → DELUSIONAL) and a numerical score. A high score does not mean "true" — it means "well-documented and consistent with known facts".
Distinguish correlation and causation
Two people who know each other does not imply complicity. The network graph shows documented links, not guilt. The presence of a flight proves nothing by itself.
Check the fact-checks
Each theory that has been fact-checked displays results with a confidence score. Fact-checks identify verified and refuted elements.
The IIE is not a verdict
An actor with an IIE of 90 is very present in the documents. This may reflect criminal involvement or simply a professional role (lawyer, manager, employee).
Explore with critical thinking
The educational modules are designed to develop critical thinking tools. Use them to question theories, including those presented here.
Site structure
Overview: statistics, top actors, theory distribution.
Interactive force-directed D3.js graph — navigate through the network of relationships.
601 actors scored by IIE, with profiles, connections and documentary appearances.
200 theories cataloged, scored and fact-checked across 10 categories.
Clusters of interconnected theories sharing common narratives.
Fact-checking results with confidence scoring.
Critical thinking education modules for conspiracy theories.
Upcoming tools
Two complementary modules are being designed to extend this archive site into interactive tools for investigation and critical thinking education.
Hellmate
In design"Just because it's partly true doesn't mean it's entirely true."
An interactive web application to test the credibility of a claim by confronting it with the database of 200 documented theories, 135 fact-checks and 15 mapped memeplexes. Instead of a binary "true/false" verdict, a multi-dimensional diagnosis.
Position on 6 levels (DOCUMENTED → DELUSIONAL) with numerical score and visual gauge.
Identification of documented real facts that make the claim credible — the base before the drift.
The exact moment where interpretation exceeds the evidence. The "logical leap" made visible.
Identification of exploited biases (confirmation, salience, precedent-as-proof...).
Which theory cluster the claim belongs to. How theories mutually reinforce each other.
How this belief can lead to more extreme positions. The funnel made visible.
Key concept: the Factual Trojan Horse
An original concept identified in the project. Most conspiracy theories are not entirely false — they start from a real fact that serves as a vehicle for an unfounded interpretation, itself a gateway to dangerous drifts.
Example:
Who is it for?
"I saw on Twitter that..." — Test a claim in 30 seconds.
Map the most dangerous theories, export sourced reports.
Show how a theory is built layer by layer.
Analyze the articulation between theories, access raw data via API.
Multi-Agent Investigation Module
Roadmap"Your investigation agents. Your inquiry. Our methodological rigor."
A platform that exposes the multi-agent system as a service, allowing external users to conduct their own investigations assisted by the 8 specialized agents, while benefiting from the existing knowledge base.
Capabilities
- ▸ Launch your own investigations using the 8 specialized agents
- ▸ Import your own corpus of documents for analysis
- ▸ Leverage conspiracy-db as a shared knowledge base
- ▸ Produce structured reports with scored conclusions
- ▸ Collaborate as a team on shared investigations
- ▸ Pedagogical feedback on reasoning quality (via Socrates)
Pre-configured workflows
Scan → Catalog → Historical context → Fact-check → Profiling → Sociology → Synthesis → Pedagogical evaluation
Michel explores the hypothesis → Factchecker counter-argues → Arbitration → Revised score
Steel Man → Best counter-argument → Defense → Pre-Mortem → Post-challenge score
Connection graph → Actor profiling → Sociological analysis → Structured report
Target audience
Three-track comparison
| Dimension | Archive (this site) | Hellmate | Investigation SaaS |
|---|---|---|---|
| Audience | Everyone | General public, teachers | Professionals |
| Mode | Browse | Claim testing | Directed investigation |
| AI Agents | Invisible | Transparent (outputs) | User-directed |
| Data | Read-only | Shared base | Base + importable corpus |
| Duration | Free | 5–30 min | Hours to days |
| Status | Active | In design | Roadmap |
This project is under continuous development. Scores, analyses and content are regularly updated as new sources are analyzed and errors are corrected.
Last update: February 2026