Skip to content

Latest commit

 

History

History
272 lines (214 loc) · 15.1 KB

File metadata and controls

272 lines (214 loc) · 15.1 KB

CDDBS Execution Plan

Project: Cyber Disinformation Detection Briefing System (CDDBS) Start Date: February 3, 2026 Delivery Model: 2-week sprints Last Updated: 2026-03-28


Project Vision

CDDBS is a system for analyzing media outlets and social media accounts for potential disinformation activity. It uses LLM-based analysis (Gemini) to produce structured intelligence briefings assessing source credibility, narrative alignment, and behavioral indicators across multiple platforms (news outlets, Twitter/X, Telegram).


Sprint Roadmap

Sprint 1: Briefing Format Redesign (Feb 3-16, 2026) — COMPLETE

Target: v1.1.0 | Status: Done

  • Researched 10 professional intelligence briefing formats
  • Designed CDDBS briefing template with 7 mandatory sections
  • Created JSON schema (draft-07) for structured output
  • System prompt v1.1 with confidence framework and attribution standards
  • Frontend mockup with sample RT analysis
  • Compliance: BYOK architecture, confidence framework, AI labeling, .gitignore for secrets

Sprint 2: Quality & Reliability (Feb 17 - Mar 2, 2026) — COMPLETE

Target: v1.2.0 | Status: Done

  • Automated quality scorer (7 dimensions, 70 points)
  • Known narratives reference dataset (7 categories, 16 narratives)
  • Source verification framework for 5 evidence types
  • 41 tests (schema validation + quality scoring)
  • System prompt v1.2 with narrative detection + self-validation
  • Compliance: Deterministic quality rubric (independent of AI), automated testing

Sprint 3: Multi-Platform Support (Mar 3-16, 2026) — COMPLETE

Target: v1.3.0 | Status: Done

  • Telegram platform analysis and behavioral indicators
  • Cross-platform identity correlation framework
  • Network analysis enhancement (graph model, community detection)
  • Schema v1.2.0 with multi-platform fields and network graph
  • Platform adapters (Twitter + Telegram)
  • System prompt v1.3 (multi-platform aware)
  • API rate limiting design (Twitter v2 + Telegram MTProto)
  • 80 tests total (39 new)
  • Compliance: Data normalization via adapters, rate limiting respect

Sprint 4: Production Integration (Mar 1-3, 2026) — COMPLETE

Target: v1.4.0 | Status: Done

  • Integrated Sprints 1-3 research into live cddbs-prod application
  • Quality scorer wired into analysis pipeline (7 dimensions, 70 points)
  • Narrative matcher running against 18 known narratives post-analysis
  • 3 new API endpoints (quality, narratives, narratives DB)
  • 3 new database tables (briefings, narrative_matches, feedback)
  • Frontend: QualityBadge, QualityRadarChart, NarrativeTags components
  • Dashboard metrics: Avg Quality + Narratives Detected
  • Unplanned: Feedback system, keyboard shortcuts, cold start handling, skeleton loading
  • 56 new tests in production (quality: 23, adapters: 22, narratives: 11)
  • Compliance: Controlled research→prod transfer, analyst feedback loop

Sprint 5: Operational Maturity & Data Ingestion (Mar 3-16, 2026) — COMPLETE

Target: v0.5.0 | Status: Done

  • Twitter API v2 integration (direct account analysis via platform adapter)
  • Batch analysis support (multiple outlets in single request)
  • Export formats (PDF, JSON, CSV)
  • Operational metrics endpoint (GET /metrics)
  • Developer documentation (812-line DEVELOPER.md)
  • Platform routing in orchestrator (news/twitter with fallback)
  • 169 tests total (35 new)
  • Compliance: Export for auditing, operational metrics, comprehensive documentation
  • See docs/sprint_5_backlog.md for details

Sprint 6: Scale, Analytics & Event Intelligence (Mar 14-18, 2026) — COMPLETE

Target: v0.6.0 | Status: Done

  • Event Intelligence Pipeline: RSS (15 feeds) + GDELT Doc API v2 collectors
  • BaseCollector ABC + CollectorManager with async scheduling
  • URL deduplication (SHA-256) + Title deduplication (TF-IDF cosine similarity)
  • Telegram Bot API integration (wired into pipeline)
  • Quality and narrative trend endpoints
  • Webhook alerting (HMAC-SHA256 signing, auto-disable)
  • CI compliance pipeline: secret scanning, documentation drift detection, branch policy enforcement
  • Open-source hardening: CODEOWNERS, SECURITY.md, CONTRIBUTING.md, LICENSE, TROUBLESHOOTING.md
  • ~197 tests total (25 new)
  • Compliance: Major compliance sprint — secret scanning CI, docs drift detection, branch policy, SECURITY.md, CODEOWNERS
  • See docs/sprint_6_backlog.md for details

Sprint 7: Intelligence Layer & Compliance Hardening (Mar 14-18, 2026) — COMPLETE

Target: v0.7.0 | Status: Done

  • TF-IDF event clustering pipeline (agglomerative clustering, distance_threshold=0.6)
  • Z-score burst detection (24h baseline, 1h window, threshold=3.0)
  • Narrative risk scoring (4-signal composite: source concentration, burst magnitude, timing sync, narrative match)
  • /events API endpoints (list, detail, map, bursts)
  • Frontend: EventClusterPanel, BurstTimeline, EventDetailDialog, enhanced GlobalMap
  • Compliance practices documentation (7 documents: DSGVO, CRA, EU AI Act)
  • Recursive completeness audit PASSED — 204 tests, all CI green
  • Compliance: Full compliance documentation folder, recursive audit, vision alignment verification
  • See docs/sprint_7_backlog.md | retrospectives/sprint_7.md

Sprint 8: Topic Mode & Supply Chain Security (Mar 22-28, 2026) — COMPLETE

Target: v0.8.0 | Status: Done

  • Topic Mode: 5-step pipeline (baseline → discovery → per-outlet comparative analysis) with coordination signal detection, key claims/omissions extraction
  • OutletNetworkGraph.tsx: Force-directed outlet relationship graph in MonitoringDashboard
  • AIProvenanceCard.tsx: Tiered AI disclosure (EU AI Act Art. 50) — model ID, prompt version, quality score, legal text
  • SBOM generation in CI: CycloneDX sbom.yml on every push to main/development, 90-day artifact retention
  • Dependency vulnerability scanning: pip-audit in CI, fails on actionable HIGH/CRITICAL CVEs
  • GitHub Actions pinned to commit SHAs (GhostAction supply chain mitigation)
  • 10 new tests (coordination logic, key claims, API schema, ai_metadata)
  • Migration fixes: startup column migrations for Sprint 8 DB schema
  • Infrastructure: Cloudflare Workers (frontend + GDELT proxy), Fly.io/Koyeb exploration, keep-alive workflow
  • Compliance: SBOM artifact (CRA Art. 13(15)), pip-audit (CRA Art. 10(4)), AI provenance (EU AI Act Art. 50)
  • See docs/sprint_8_backlog.md | retrospectives/sprint_8.md

Sprint 9: AI Trust, Information Security & Compliance Automation (Mar 28, 2026) — COMPLETE

Target: v0.9.0 | Status: Done

  • AI Trust Framework: LLM output validation (output_validator.py), grounding score (TF-IDF claim verification), confidence calibration
  • Information Security Hardening: CORS fix, rate limiting (slowapi), prompt injection prevention (input_sanitizer.py), security headers, error sanitization, API key hygiene
  • Compliance Automation: Machine-readable /compliance/evidence endpoint, custom dependency scanner (replaces Dependabot)
  • OWASP LLM Top 10: LLM01, LLM02, LLM04, LLM06, LLM09 mitigated
  • 35 new tests, 249 total
  • Compliance: OWASP LLM Top 10 coverage, EU AI Act Art. 9/12/14, CRA security hardening
  • Versioning: Adopted semver 0.x.y — retagged v2026.03v0.5.0
  • See docs/sprint_9_backlog.md for details

Sprint 10: User Authentication & CDDBS-Edge (Apr-May 2026)

  • User authentication and authorization (JWT, role model, session management)
  • CDDBS-Edge Phase 0: Swap Gemini → Ollama, benchmark briefing quality
  • Analyst annotations and comments on briefings

Sprint 11: Collaboration & Advanced Features (May-Jun 2026)

  • Shared analysis workspaces (depends on Sprint 10 auth)
  • Automated monitoring schedules
  • API for third-party integration

Sprints 12+: Future (Jun-Aug 2026)

  • Machine learning model fine-tuning
  • Multi-language support
  • Currents API collector integration

CDDBS-Edge: Portable Offline Briefing System (Parallel Track)

Status: Concept — Experiment Phase 0 in planning Scope: Separate hardware prototype track, runs parallel to cloud sprints Design doc: research/cddbs_edge_concept.md

"What happens when the cloud goes down, the API gets blocked, or you're a journalist in a country that restricts internet access?"

A portable, offline-capable version of CDDBS that runs entirely on a Raspberry Pi 5 with a local quantized LLM (Phi-3 Mini 3.8B via Ollama), replacing all external API calls. Output delivered via MQTT broker to a connected display (e-ink HAT or external screen — approach TBD by experiment).

Experiment Phases:

  • Phase 0 (no hardware): Swap Gemini → Ollama on laptop, benchmark briefing quality vs cloud baseline
  • Phase 1: Deploy pipeline on Raspberry Pi 5 (8GB), benchmark speed/RAM/thermal
  • Phase 2: Wire MQTT output, prototype display options (e-ink HAT vs MQTT subscriber)
  • Phase 3: Design offline data ingestion (USB-based article import or minimal RSS fetch)

Why it matters for AI trust & governance: Demonstrates resilience, digital sovereignty, access equity, and privacy-preserving AI deployment — concrete artifacts for governance discussions that most researchers only address theoretically.


Architecture

Current Stack (as of v0.9.0)

  • Backend: FastAPI + uvicorn + slowapi on Render (Docker)
  • Frontend: React 18 + TypeScript + MUI 6 + Vite on Cloudflare Workers + Render
  • Database: PostgreSQL 15 (Neon managed, 12 tables)
  • LLM: Google Gemini 2.5 Flash via google-genai SDK
  • Data Sources: SerpAPI Google News, Twitter API v2, GDELT Doc API v2 (Cloudflare proxy), RSS (15 feeds)
  • CI: GitHub Actions (7 workflows)
  • Source Code: GitHub (cddbs-prod + cddbs-research)

Achieved Architecture (v0.6.0)

  • Structured briefing output validated against JSON Schema v1.2
  • 7-dimension quality scoring pipeline (70-point rubric)
  • Narrative detection against 50+ known disinformation narratives
  • Platform adapters for Twitter + Telegram (both wired into pipeline)
  • Multi-source event intelligence pipeline (RSS + GDELT)
  • URL + title deduplication (SHA-256 + TF-IDF cosine)
  • Webhook alerting with HMAC-SHA256 signing
  • CI compliance pipeline (secret scan, docs drift, branch policy)
  • Background task processing with auto-polling frontend
  • Batch analysis and export (JSON/CSV/PDF)
  • Operational metrics and trend endpoints

Achieved Architecture (v0.7.0)

  • Event clustering and burst detection (TF-IDF agglomerative + z-score)
  • Narrative risk scoring composite (4-signal: source_concentration, burst_magnitude, timing_sync, narrative_match)
  • Events API and frontend visualization (EventClusterPanel, BurstTimeline, GlobalMap overlay)
  • 204 tests, 3 CI workflows, 7 compliance documents

Achieved Architecture (v0.8.0)

  • Topic Mode: 5-step pipeline — baseline fetch, Gemini baseline, broad discovery, per-outlet comparative analysis, coordination signal detection
  • OutletNetworkGraph: force-directed outlet relationship visualization
  • AIProvenanceCard: tiered AI disclosure (model ID, prompt version, quality score, legal text)
  • SBOM generation (CycloneDX) and pip-audit vulnerability scanning in CI
  • GitHub Actions pinned to commit SHAs (supply chain hardening)
  • Infrastructure: Cloudflare Workers (frontend + GDELT proxy), keep-alive workflow

Achieved Architecture (v0.9.0)

  • AI trust framework: output validation, grounding score (TF-IDF claim verification), confidence calibration
  • Information security: CORS hardening, rate limiting (slowapi), input sanitization, security headers, error sanitization, API key hygiene
  • Compliance automation: /compliance/evidence endpoint, custom dependency scanner (replaces Dependabot)
  • OWASP LLM Top 10: LLM01, LLM02, LLM04, LLM06, LLM09 mitigated
  • 249 tests, 7 CI workflows

Target Architecture (v0.10.0+)

  • User authentication and authorization (JWT, RBAC)
  • CDDBS-Edge Phase 0 (Gemini → Ollama swap, benchmark)
  • Shared analysis workspaces

Key Principles

  1. Evidence over speed - Every claim must be traceable to evidence
  2. Confidence transparency - Always communicate uncertainty honestly
  3. Reproducibility - Analyses should be reproducible with the same inputs
  4. Professional standards - Output should meet intelligence community standards
  5. Cost discipline - Stay within free/low-cost tier limits
  6. Compliance by design - EU regulatory requirements (DSGVO, CRA, EU AI Act) addressed through engineering practices, not afterthought

Branching Policy

Repository Branch Policy
cddbs-prod Feature branches from development → merge to development → merge to main
cddbs-research Feature branches from main → merge to main

Production code flows through the development branch as a staging/integration area before reaching main. This is enforced by CI (branch-policy.yml).


Vision Alignment Check (as of Sprint 8 Planning)

Sprint Contribution to Vision On Track?
1 Briefing format — core intelligence output Yes
2 Quality scoring — reliability of AI analysis Yes
3 Multi-platform — broader disinformation coverage Yes
4 Production integration — making research usable Yes
5 Operational maturity — production-grade features Yes
6 Event intelligence — proactive monitoring capability Yes
7 Intelligence layer — automated event detection Yes ✓
8 Topic Mode, supply chain security, AI provenance — proactive discovery + compliance Yes ✓
9 AI trust, information security, compliance automation — output integrity + platform hardening Yes ✓

Drift assessment: No significant drift from project vision. All sprints serve the core mission of "analyzing media outlets and social media accounts for potential disinformation activity."

Sprint 9 reprioritization note: The original plan placed user authentication in Sprint 9. The Sprint 8 security audit revealed critical gaps (prompt injection, no rate limiting, CORS misconfiguration) that must be resolved before adding auth. Additionally, for a disinformation detection system, AI output trustworthiness (grounding scores, hallucination detection) is more mission-critical than access control. Auth is now Sprint 10 — this is a deliberate sequencing decision, not scope drift. The core features (auth, workspaces, annotations, CDDBS-Edge) remain on the roadmap with unchanged priority.

Potential drift risks:

  • CDDBS-Edge is a parallel track that could divert focus — mitigated by keeping it separate and experiment-phase only
  • Collaborative features (Sprint 9) could drift toward general-purpose workspace — must stay focused on analyst collaboration for disinformation analysis
  • Compliance documentation is valuable but must not become the primary focus — it supports engineering quality, not the other way around

Compliance Documentation

See compliance-practices/ for comprehensive documentation of all DSGVO, CRA, and EU AI Act measures implemented across Sprints 1-7.