Architecture

                ┌────────────────────────────┐
                │ Browser (Next.js / React)  │
                │  Mission Desk · Registry · │
                │  Dossier · Toolkit · Ledger│
                └──────────────┬─────────────┘
                               │ axios (REST/JSON)
                               ▼
   ┌──────────────────────────────────────────────────────┐
   │ FastAPI backend  (port 8000)                         │
   │  ┌─────────────┬────────────┬────────────────────┐  │
   │  │ analysis.py │ intelligence│ pdf_export.py      │  │
   │  │ (LLM +      │ (heuristics:│ (ReportLab + QR)   │  │
   │  │ heuristics) │ twins,      │                    │  │
   │  │             │ graph, geo, │ regulations.py     │  │
   │  │             │ lang, counter)                    │  │
   │  │             │             │ (citation library) │  │
   │  └─────────────┴────────────┴────────────────────┘  │
   │           │                                          │
   │           ▼ SQLAlchemy ORM                           │
   └──────────────────────────────────────────────────────┘
                               │
                ┌──────────────▼──────────────┐
                │ Postgres (pgvector image)   │
                │  items · audit_events ·     │
                │  toolkit_incidents ·        │
                │  app_settings               │
                └─────────────────────────────┘

items

Every dossier. Stores source text/URL, narrative, entities, claims, three core scores, confidence, headline, language/country, brief, legal flags, recommended actions, policy roadmap, status, and transparency state.

audit_events

Append-only trace of meaningful actions such as dossier creation, completed analysis, escalation, counter-narrative drafting, PDF export, and publication to the transparency ledger.

toolkit_incidents

Manual, non-LLM rapid-response incident logs from the toolkit. Captures the alert card, risk matrix inputs, chosen response, notes, and the saved risk band.

app_settings

Key/value store used for election-mode state, activation timestamp, and SLA duration. Keeps operational toggles separate from dossier data.

deepfake_risk

Returned by analysis.analyze_content() as ai_probability, clamped to [0, 1]. Fallback path returns 0.5 when the LLM is unavailable.

coordination_score

Returned by analysis.analyze_content() as coordination_score, clamped to [0, 1]. Fallback path returns 0.2.

disinformation_impact

Primary escalation score. The current backend blends 35% model output with 65% heuristic output, then caps low-signal content so harmless political reporting does not get over-tagged.

confidence

Computed from the spread between the three core signals: 1 − stdev(values), clamped to [0.45, 0.97]. The UI renders this as the +/- band around each meter.

• Baseline score starts at 0.02 instead of a heavier default.
• Benign context terms like report, committee, briefing, and interview activate a soft cap.
• If the text is only about elections/politics and lacks harm, amplification, or synthetic cues, impact is capped at 0.22.
• Benign political reporting is further capped at 0.15 where appropriate.
• Final impact is a weighted blend: 35% model + 65% heuristic, not a blind max().
• Low-signal contexts are capped at 0.25 even if the model score is aggressive.

Hero, KPI strip, live evidence ticker, intake form, escalation queue, EU heatmap, language pills, and the always-on policy framework block.

Filterable dossier table with search, status filters, risk-tier filters, and links into each dossier.

Hero controls, claims/entities, confidence meters, regulatory exposure, coordination graph, narrative twins, audit log, PDF export, counter-narrative drawer, and policy roadmap.

Manual alert card, two-minute triage checklist, risk calculator, decision menu, templates, saved incidents, escalation contacts, and AI literacy modules.

DSA Article 24-style public log of redacted interventions, legal basis chips, and action summaries.

This page: a technical blueprint for implementers, reviewers, and deployment owners.

• Queue threshold drops from 0.45 to 0.30 in dashboard_stats().
• Frontend risk tiering adds a +0.12 lift before threshold checks.
• Pending dossiers display a live SlaCountdown pill using created_at + slaMinutes.
• The topbar status pill switches to ELECTION MODE LIVE with an alert tone.
• Every toggle writes election_mode.toggled to the audit log with legal basis DSA Art. 35.

A curated in-repo dict keyed like 2024/1689:50 stores article title, paraphrased excerpt, and source PDF filename. lookup(citation) powers the citation drawers, and parse_flag_to_citation(flag) turns natural-language legal flags into lookup keys.

GET /api/items/{id}/pdf produces a one-page EP-style brief using ReportLab with a header, four metric cells, situation overview, regulatory exposure, recommended action, and a QR code back to the dossier endpoint. If ReportLab is unavailable, a minimal handwritten PDF fallback prevents runtime failure during demos.

Audit log

Every state-changing route calls record_event(item_id, action, detail, actor, legal_basis). The UI maps those keys to human-readable labels and action tones.

Transparency ledger

Publishing requires a completed or escalated dossier, flips transparency_published to true, and emits a DSA Art. 24 audit event. Narratives are redacted to avoid re-amplification.

Demo replay

POST /api/demo/replay seeds three prewritten dossiers: Marin synthetic audio, Marseille polling-station network, and a promoted opposition deepfake. They are intentionally scripted so every major UI feature has data.

Impact > origin

Impact > origin. Electoral harm and amplification matter more than synthetic provenance alone.

Conservative low-signal handling

Conservative low-signal handling. The backend now caps topic-only or benign reporting so ordinary political coverage does not drift into monitoring tiers.

Legal grounding is deterministic

Legal grounding is deterministic. The LLM writes prose, but augmenters guarantee AI Act, DSA, GDPR, and political-ads citations are present.

Auditability is a first-class feature

Auditability is a first-class feature. Every operator action and system action writes a reasoned event with legal basis where available.

Offline-friendly heuristics keep the demo explainable

Offline-friendly heuristics keep the demo explainable. Twins, graph, geo, confidence, and redaction all work without vector search or opaque external scoring services.