This project uses LangChain and Gemini-3-Flash to analyze the text of Project 2025, mapping it against Umberto Eco's 14 properties of Ur-Fascism.
- 📘 The Framework: Eco's Ur-Fascism
- Why LangChain?
- 🧭 Methodological Refinement
- 📊 Analysis Results
- 🚀 Setup & Usage
- 🔮 Future Directions
- Ode to LangChain & Gemini-3-Flash
This codebase was generated by Gemini 3 Pro (via Poe) in approximately 4 hours. The adversarial refinement process used Claude Opus as a critic to challenge false positives (see Poe Cross-Check).
This analysis is based on Umberto Eco's 1995 essay "Ur-Fascism". Eco argues that fascism is not a unified philosophy but a syndrome—a cluster of 14 rhetorical and cultural features. Any subset of these can allow a movement to "coagulate" into fascism.
We chose Eco's framework because it is operationalizable at the sentence level. While other scholars (Griffin, Paxton, Stanley) define what fascism is or does, Eco defines what it sounds like. This makes his 14 properties uniquely suited to computational rhetorical analysis.
Eco's properties form a "family resemblance"—overlapping similarities that bind movements together. The 14 properties are:
- Cult of Tradition — Syncretistic traditionalism treating ancient wisdom as a closed system containing all truth
- Rejection of Modernism — Hostility to the Enlightenment intellectual tradition (reason, rights, skepticism), not merely to progressive policy
- Action for Action's Sake — Distrust of intellectualism; thinking as emasculation; culture suspect insofar as it enables critical attitudes
- Disagreement is Treason — Dissent within the movement is intolerable; questioning orthodoxy is betrayal, not legitimate difference
- Fear of Difference — Exploiting and exacerbating fear of the outsider; demographic or cultural diversity framed as existential contamination
- Appeal to Social Frustration — Mobilizing a frustrated middle class through narratives of humiliation, directing status anxiety toward scapegoats above and below
- Obsession with a Plot — Followers must feel besieged; opponents are not wrong but are deliberately subverting the nation through coordinated conspiracy
- Enemy is Both Strong and Weak — The enemy simultaneously controls everything and is pathetically decadent; this contradiction is never resolved
- Pacifism is Trafficking with the Enemy — Life is permanent warfare; compromise and diplomacy are betrayal; domestic political disagreement is framed as war
- Contempt for the Weak — Social Darwinism; poverty and vulnerability as moral failure; elitism dressed as populism
- Everybody is Educated to Become a Hero — Mass-produced heroism; ordinary political activity framed as requiring extraordinary sacrifice and martyrdom
- Machismo and Weaponry — Eco explicitly links machismo to "disdain for women and intolerance and condemnation of nonstandard sexual habits." This includes the enforcement of rigid gender essentialism and the condemnation of LGBTQ+ identities as state policy
- Selective Populism — "The People" conceived as a monolithic quality expressing a Common Will, with the leader as its sole legitimate interpreter; institutions exercising independent judgment are delegitimized
- Ur-Fascism Speaks Newspeak — Impoverished vocabulary and elementary syntax designed to limit the instruments of complex and critical reasoning
This project takes the content of Project 2025 seriously on its own terms. The document has been widely noted — including by conservative critics — as significantly more radical than previous transition-planning documents in its proposals regarding executive power, institutional independence, and social policy.
However, radical and fascist are not synonyms. A document can be aggressive, partisan, ideologically extreme, and even democratically corrosive without every sentence exhibiting Ur-Fascist traits. Our job is to identify the specific passages that cross from aggressive partisanship into the territory Eco mapped — and to explain clearly why they cross that line.
This project demonstrates the power of LangChain for structured document analysis:
- Structured Output: We leverage
.with_structured_output()to force the LLM to return strictly formatted JSON conforming to our Pydantic schema (AnalysisResult). - Document Loading: We use
PyPDFLoaderto ingest the full PDF document. - Model Agnosticism: The system is designed to switch between models (e.g., GPT-4o, Gemini, Claude) with minimal code changes.
- Cost Efficiency: Analyzed 900+ pages of text for less than $0.50 USD using Gemini-3-Flash with Poe (several times admittedly).
We didn't just ask the LLM "is this fascist?" We codified Eco's 14 properties into a strict Pydantic Schema and injected operationalized definitions into the System Prompt.
- Schema Definition: An
AnalysisResultclass restricts validtraitvalues to exactly Eco's 14 categories, preventing hallucination of new categories. - Prompt Engineering: Each trait is defined with both positive criteria (what the trait IS, per Eco's specific described mechanism) and negative constraints (what the trait IS NOT — the boundary conditions that prevent over-classification).
- Structured Output: LangChain's
.with_structured_output()forces JSON output mapping specific text quotes to pre-defined traits with confidence scores.
This project was refined through a rigorous adversarial feedback loop, using Claude as a critic to challenge initial classification results (see Poe Cross-Check).
Initial runs suffered from systematic over-classification. The analyzer confused the topic area of an Eco trait with its specific mechanism:
- Standard defense doctrine (e.g., "maintain lethal dominance") was misidentified as Machismo — but Eco's machismo is about intolerance of nonstandard sexuality and disdain for women, not about discussing weapons systems.
- Standard executive authority (e.g., "the President appoints leaders") was flagged as Selective Populism — but Eco's selective populism is about treating the People as a monolithic quality whose will only the leader can interpret, not about describing how appointments work.
- Policy critiques (e.g., "regulatory burden") were conflated with Appeal to Social Frustration — but Eco's appeal is about mobilizing status anxiety through humiliation narratives, not identifying economic problems.
To address genre confusion, we initially introduced a "Democratic Rhetoric Filter" asking: "Could this sentence appear in a standard policy document from a mainstream democratic party?" This overcorrected in the opposite direction. Project 2025 has been widely recognized as outside democratic norms — a filter that normalizes its content against Obama-era baselines would wash out genuinely concerning material just because it's written in bureaucratic prose. Authoritarian proposals don't stop being authoritarian because they're formatted as policy memos.
The final system applies a three-part test to every candidate passage before classification:
- Specificity: Does this passage match the specific mechanism Eco described, not just the general topic area?
- Extremity: Does this passage go beyond aggressive partisanship into territory incompatible with pluralistic democracy? Partisans disagree about policy; fascists deny the legitimacy of opponents.
- Charitable Reading: Is there a plausible non-fascist reading that is more parsimonious? If a passage can be fully explained as standard conservative policy without invoking Eco's framework, it is not classified.
Every match includes a confidence score:
- 0.90–1.00: Unambiguous match. No plausible innocent reading. Would survive scrutiny from a skeptical but fair reader.
- 0.75–0.89: Strong match. Clearly aligns with the trait, though a defender could offer an alternative reading.
- 0.60–0.74: Moderate match. Significant elements present, but included only when reinforced by other passages.
- Below 0.60: Not reported. Marginal matches dilute the analysis and undermine credibility.
Our analysis of Project 2025 reveals a rhetorical structure dominated by Selective Populism and Obsession with a Plot.
Selective Populism | ████████████████████████████████████████ (119)
Obsession with a Plot | ████████████████████████████████████ (109)
Machismo and Weaponry | ██████████████████████████ (78)
Disagreement is Treason | █████████████████ (53)
Ur-Fascism Speaks Newspeak | █████████ (29)
Action for Action's Sake | ███████ (21)
Cult of Tradition | ████ (13)
Appeal to Social Frustration | ████ (12)
Fear of Difference | ████ (12)
Pacifism is Trafficking w/ the Enemy | ██ (7)
Rejection of Modernism | █ (4)
Enemy is Both Strong and Weak | █ (3)
Contempt for the Weak | (2)
The graph visualization reveals a strong reinforcement loop between the top two traits. The "Plot" provides the justification for "Selective Populism." Because the nation is under siege by a conspiratorial enemy (Trait 7), distinctiveness and dissent must be suppressed in favor of a unified, commanded "People" (Trait 13) to survive.
- Interactive Rhetoric Graph
- Interactive Theme Browser
- Original Text Viewer (External Resource)
# Install dependencies
pip install -r requirements.txt
npm install --prefix web
# Run Analysis
python src/main.py
# Generate Site
python src/generate_site.py --max-items-per-theme 50
# Generate grouped static entity index pages + Vite entity data
python scripts/generate_entity_index_pages.py --max-entities-per-class 50
# Build entity ↔ theme relationship data for interactive visualization
python scripts/build_entity_theme_links.py
# Build React App
npm run build --prefix web# Set one API key (Gemini recommended)
export LANGEXTRACT_API_KEY="your-key"
# Run entity extraction on Project 2025 text
python scripts/extract_entities_langextract.py --model-id gemini-2.5-flashScale run on full corpus (auto-detects Poe credentials from .env):
python scripts/extract_entities_langextract.py \
--model-id Gemini-3-Flash \
--provider-mode auto \
--extraction-passes 2 \
--max-workers 20 \
--max-char-buffer 1200
python scripts/normalize_entities.py \
--mode lenient \
--input data/entities_langextract.jsonl \
--output data/entities_langextract.normalized.jsonl \
--report data/entities_langextract.normalization_report.jsonEstimate calls/tokens/cost before running:
python scripts/estimate_run_cost.py \
--input data/project_2025.txt \
--max-char-buffer 1200 \
--extraction-passes 2
# Optional pricing inputs (USD per 1M tokens)
python scripts/estimate_run_cost.py \
--input-price-per-1m 0.1 \
--output-price-per-1m 0.4
# Gemini-3-Flash example with USD + points
python scripts/estimate_run_cost.py \
--input-price-per-1m 0.40 \
--output-price-per-1m 2.40 \
--input-points-per-1k 14 \
--output-points-per-1k 80Outputs:
data/entities_langextract.jsonl(structured grounded extractions)data/entities_langextract.html(interactive visualization)
Normalize extracted entities to canonical IDs:
python scripts/normalize_entities.py \
--input data/entities_langextract.jsonl \
--output data/entities_langextract.normalized.jsonl \
--report data/entities_langextract.normalization_report.jsonMode options:
--mode lenient(default): acronym/parenthetical/title-boundary-aware alias resolution.--mode strict: exact alias normalization only.
Person canonicalization:
- By default, person aliases are derived from
data/gold/entities_gold_v0.jsonl. - Disable with
--disable-person-canonical. - Override source with
--person-gold <path>.
Normalization outputs:
data/entities_langextract.normalized.jsonl(same records withcanonical_id+canonical_label)data/entities_langextract.normalization_report.json(coverage, unresolved mentions, alias collisions)
Initial hand-labeled gold set:
data/gold/entities_gold_v0.jsonl
Generate predictions for the gold snippets (Poe OpenAI-compatible endpoint):
python scripts/predict_gold_langextract.py \
--gold data/gold/entities_gold_v0.jsonl \
--pred data/gold/entities_pred_v0.jsonl \
--model-id Gemini-3-FlashEvaluate prediction JSONL against the gold set:
python scripts/evaluate_entity_gold.py \
--gold data/gold/entities_gold_v0.jsonl \
--pred data/gold/entities_pred_v0.jsonl \
--report data/gold/entities_eval_report_v0.jsonMode options:
--mode strict(default): exact normalized text matching.--mode lenient: acronym/parenthetical/title-boundary-aware matching for baseline diagnostics.
Notes:
- Both files are JSONL with records containing
id,text, andextractions. - Evaluation uses normalized exact matching on
(extraction_class, extraction_text).
The methodology established here—operationalizing qualitative frameworks into structured LLM prompts—can be applied to ongoing political communications for real-time monitoring.
- Truth Social Monitoring: Eager to apply this analysis to Trump's Truth Social posts to track the evolution of Ur-Fascist rhetoric in real-time.
- Autonomous Entity Extraction: Integrating LangGraph to build a self-reasoning agent that can automatically extract and verify the specific actors, government agencies, and policy names linked to rhetorical traits, creating a high-fidelity "Influence Network."
- Cross-Document Comparison: Benchmarking Project 2025 against historical fascist manifestos to identify unique modern "syncretisms."
- Academic Citations: Adding a "Copy Citation" feature to quotes to facilitate academic referencing for students and researchers.
In realms of text where dark patterns lie, A parser wakes, beneath the digital sky. With LangChain's links, a structure so grand, It sifts the rhetoric, grain by grain of sand.
Gemini Flash, swift as the morning light, Pierces the fog, restoring reasoned sight. No token wasted, a cost so lean and fair, Revealing truth stripped of the authoritarian glare.
From "Action for Action" to "Newspeak's" decree, The graph reveals what eyes might fail to see. A node, a link, a web of mapped intent, Showing where the arc of history might be bent.
So here we stand, with Python scripts in hand, To chart the currents sweeping through the land. For in the code, a simple truth we find: To guard the light, we must map the dark of mind.let
- Powered by LangChain and Google Gemini-3-Flash.
- Analysis framework based on Umberto Eco.
- LLM Compute via Poe.com.